Datasets:

Prikshit7766

/

huggingface_forum

like 2

Open Source AI and Law

https://discuss.huggingface.co/t/open-source-ai-and-law/86086

Is there a research group for Open Source AI and Law? If not, I would like to start one. How would one proceed though?

2024-05-10T08:50:22Z

Harris County property tax protest

https://discuss.huggingface.co/t/harris-county-property-tax-protest/85424

As a property tax protest service provider situated in the vibrant state of Texas, we specialize in assisting clients withHarris County property tax protestchallenges. Our team of seasoned professionals boasts extensive experience in real estate tax law and the intricacies of property tax protesting. Our primary objective is to alleviate your tax burden and help you save money. With our wealth of knowledge and abundant resources, we are fully equipped to deliver high-quality tax protest services to individuals, businesses, and organizations across the entire state of Texas.Our commitment to excellence is unwavering. Whether you’re a homeowner seeking relief from unfair tax assessments, a growing business looking to optimize your financial strategies, or a respected organization navigating complex tax structures, we are prepared to advocate tirelessly on your behalf. Through meticulous analysis, strategic planning, and unwavering dedication, our goal is not only to reduce or eliminate your tax liabilities but also to empower you for sustained financial success in the dynamic realm of taxation.

2024-05-06T12:50:16Z

Empathetic Generative AI

https://discuss.huggingface.co/t/empathetic-generative-ai/77290

I am interest in being in contact with anyone doing research in the space of empathetic AI. I maintain a blog site in this spaceembench.com(i do not make money from the blog). I am interested in highlighting work in the space.

2024-03-13T15:37:16Z

Dark zones of LLM

https://discuss.huggingface.co/t/dark-zones-of-llm/84184

The idea of building embeddings from embeddings or CNN for LLMLet’s imagine that we have a neural network that can assemble a sequence of embeddings into one vector of higher dimension. It’s like a convolutional network, but not for images, but for embeddings/texts. That is, we get coordinates in the space of meaning not for individual tokens, but for their totality, for the entire text/text fragment. This resulting vector cannot be made very large due to the amount of calculations and we will not get a good network for large texts, not to mention all the knowledge and texts accumulated by humanity. But the problem can be solved for one sentence, for a theorem, a chemical formula, an engineering solution, a poem, a short story. We are fine with the loss of the original sequence of tokens, as long as the meaning of this sequence is preserved.Such an add-on to embeddings (let’s call it Super LLM/SLLM) will be trained without a teacher - at the same time an encoder and a decoder. The encoder’s task is to construct a vector of a larger size, but in a single quantity. A sequence of embeddings is supplied to the encoder input, and one vector is output. The decoder’s task is to reconstruct the embedding sequence from it. One vector is supplied to the decoder input, and a sequence of embeddings is expected at the output. The loss function will need to be constructed so that the decoder tries to restore the meaning rather than just the sequence of embeddings leading to the same sequence of tokens. You need to train this SLLM after finishing training the basic LLM, but without alignments.The big advantage is that SLLM may be trained not on the entire corpus of texts of the basic LLM. For example, we may be interested in just one area of natural laws, human knowledge, or art. Or some subset of these areas. In human understanding, this SLLM with a basic LLM should literally understand the meaning of texts in their entirety, but in a narrow subject area.Zones of darknessNow why is all this needed? The term zone of darkness may not be accurate, I do not imply any occultism. The term does not carry anything dark, scary or dangerous. It simply means the absence of light. We cannot make out, what is actually happening at this point, in this vector, in this embedding. The dark zone is a set of parameter values of the final vector that never occur for the source data.The SLLM encoder creates a vector from incoming embeddings. Common sense dictates that with a wide variety of types, topics and content of texts, the encoder should try to occupy the entire available multidimensional space of meanings, should try to use all possible combinations of parameters with small values. But it may turn out that there are voids in the distribution of embdings. Perhaps there are areas with such parameter values that are not encountered when processing all embeddings during SLLM training. There may be areas, even closed areas, into which the resulting vectors do not fall or almost do not fall. It may even turn out that these areas are not chaotic, but have certain sizes or shapes. Why does this happen? Why doesn’t SLLM try to occupy these areas in its training? Is this a random process or not? How common are such areas?I have an assumption that such areas should be formed and contain knowledge and patterns that SLLM was able to capture in the learning process, but which are not directly found in the texts used for learning. Humanity either did not understand these patterns or ignored this knowledge.Why dark zones might be importantThere is a well-known example of how a neural network used to analyze fluorography was able to determine race from images. Although people do not know how to do this (even if because they were never interested in this) and could not teach it. That is, the neural network itself discovered new patterns and knowledge during training. Didn’t the same thing happen in LLM/Transformers? Couldn’t dark zones be not just emptiness, but new knowledge, ideas and concepts? We can force the SLMM and LLM decoders to give us text that matches the selected embeddings from the darkness zone. What will we see there? Rave? Hallucination? Just a platitude? Or an unusual idea? Ignored pattern? We won’t know until we try.Is the study of dark zones a new, unappreciated and underappreciated method of scientific research? We train SLLM based on our existing knowledge and experimental results. And after that, we feed embeddings from the dark zones to the decoder input. Will we be able to obtain new knowledge, new theories, unknown chemical compounds? And if you train transformers in several sciences at the same time, will you get new scientific disciplines at the intersection of existing ones? What if the results obtained from the embeddings of the dark zones were used to train the next version of LLM? Wouldn’t it turn out to be a neural network for understanding the world, which first generalizes what is already known, then tries to obtain new knowledge from the areas of darkness, selects the least delusional from them and uses this knowledge to further educate itself, and this continues the cycle.Zones of darkness (if they exist) are a clear formal criterion for where and how to get new knowledge. Embeddings that are not encountered during SLLM training are found and sent to the decoder input, first SLLM, then basic LLM.A simple experiment to test the idea’s functionalityWe take a set of texts on mathematics at the elementary school level. We exclude from training all fragments of texts on any one selected topic. Let’s train SLLM, let’s call this neural network NN1. We add the previously excluded texts and additionally train NN1, let’s call this network NN2. Now we feed the initially excluded texts to the input of the NN2 encoder and get a set of embeddings that should not be encountered when training NN1 (this can even be checked if you save all the embeddings encountered). And we feed these embeddings to the input of both the NN1 decoder and the NN2 decoder. And then we compare the results. If the idea works, we should get similar results. You can feed them further to the inputs of the basic LLM decoder and compare the generated texts.This means that NN1 already has knowledge that did not exist explicitly during training. And they are in areas of darkness. And it’s enough to either learn to find such zones, or simply go through the most suitable options.

2024-04-28T13:55:51Z

ASR spell correction

https://discuss.huggingface.co/t/asr-spell-correction/5103

Because I love the mindset within the community of the Wave2vec sprint I’d like to share some ideas about improving the accuracy of asr and making more stable for production.I would be happy to discuss about.In some experiments I tested many systems and algorithms, but especially one reached amazing accuracy.When having the transcribed text from the wave2vec model we go many ways to correct them. Either dictionary search for each word an automatically use the nearest result, or an seq2seq model, but what about an hybrid solution based on two or three parts ?Part 1: token classification, to recognize which words are wrong in the context. Instead of human names or locations just classify wrong or right.Part 2: When we have the wrong tokens let’s check an dictionary for similar alternative, either using bm25 (tested) or dpr neural search (untested)Part 3: When we have some alternatives for each token we can either use the best scored result or let an multiple-choice trained model decide. In my quick tests I decided using the best alternative, but definitely need to check the multiple choice variant.With this 3 stepsToken classificationDictionary search using bm25 like algorithmsReplacing false tokens with best scored alternativeI reached amazing results and up to WER of 1.3%At the moment my code is pretty noisy and I would like to start from zero again to build an clean library based on huggingface models, or maybe just an community notebook, depends on your feedbackI’d like to hear what you think about, maybe you have much better idea ?Maybe someone is interested in joining this research ?

2021-03-25T21:59:50Z

Democratisation of Machine Learning Survey

https://discuss.huggingface.co/t/democratisation-of-machine-learning-survey/83092

We are two students from the IT University of Copenhagen doing our master thesis about the democratisation of machine learning. We are eager to know how the machine learning community engages with and views the technology. So if you have the time, please fill out this survey, it should take no more than 5 min. Thank you for your time!docs.google.comThe Democratisation of Machine Learning - SurveyThank you for taking the time to answer this survey about people’s experience with machine learning, it should take no more than 5 min Throughout this survey 'Machine Learning' will be referred to as 'ML'.

2024-04-22T11:09:56Z

[Call for Participation] GermEval2024 GerMS-Detect - Sexism Detection in German Online News Fora @Konvens 2024

https://discuss.huggingface.co/t/call-for-participation-germeval2024-germs-detect-sexism-detection-in-german-online-news-fora-konvens-2024/82657

GermEval2024 Shared Task: GerMS-Detect – Sexism Detection in German Online News Fora1st CALL FOR PARTICIPATIONWe are pleased to announce the GermEval Shared Task GerMS-Detect on Sexism Detection in German Online News Fora collocated withKonvens 2024.Competition WebsiteImportant Dates:Trial phase: April 20 - April 29, 2024Development phase: May 1 - June 5, 2024Competition phase: June 7 - June 25, 2024Paper submission due: July 1, 2024Camera ready due: July 20, 2024Shared Task@KONVENS: 10 September, 2024Task descriptionThis shared task is about the detection of sexism/misogyny in comments posted in (mostly) German language to the comment section of an Austrian online newspaper. The data was originally collected for the development of a classifier that supports human moderators in detecting potentially sexist comments or identify comment fora with a high rate of sexist comments. For details see theCompetition Website.OrganizersThe task is organized by theAustrian Research Institute for Artificial Intelligence (OFAI).Organizing team:Brigitte Krenn (brigitte.krenn (AT) ofai.at)Johann Petrak (johann.petrak (AT) ofai.at)Stephanie Gross (stephanie.gross (AT) ofai.at)

2024-04-19T14:45:39Z

Need Suggestion

https://discuss.huggingface.co/t/need-suggestion/81570

Hi all,I’m new to this LLM’s world and I need suggestions on the following idea. I want to fine-tune a LLM model based on exam past papers, which are currently available on PDF format.Objective: The model understand and explain past answers and able to generate new questions and answers.Which LLM model I should select for this purpose? Keep in mind I want to start from very basic as I mentioned earlier I’m very novice.The past papers are in PDF format with pictures as well, Do I need to convert them on some specific format like JSON?

2024-04-13T15:06:35Z

Get text generation in particular format

https://discuss.huggingface.co/t/get-text-generation-in-particular-format/82313

hi, i am currently inferring different LLMs for classification and their inherit understanding of requirement engineering.this is a zero shot prompt designed for inferring the LLMs.few LLMs are giving the output in described format. but few models, even with higher number of parameters are not giving answer. rather they just giving output like this:‘generated_text’ : \nSentence1\n\nSentence2\n\nSentence3.expected output is: label1, label2, label3test_zero_prompt = “”"Functional requirements specify the functions or behaviors that a system must provide. They describe what the system should do.Non-functional requirements specify constraints, qualities, or characteristics that the system must possess, such as performance, usability, or security, etc. They describe how the system should behave or perform.Information (also known as non-requirements) refers to statements that provide context, background, or explanations but do not specify any requirements for the system.classify the given sentences into functional requirement, non-functional requirement, or information. Give the answer in the classification labels only.Sentence 1: this is a functional requirement.Sentence 2: this is a non-functional requirement.Sentence 3: this is an information.Sentence 4: this is a functional requirement.Answers:functional requirement, non-functional requirement, information, functional requirementClassify the following Sentences according to format above: “”"can anyone suggest, how can i improve the prompt or change it.or find mistake or error in prompt?

2024-04-17T23:48:03Z

Token merging for fast LLM inference

https://discuss.huggingface.co/t/token-merging-for-fast-llm-inference/82202

Hello all,I worked on a project aiming at speeding up inference of LLMs by using merging of sequence. The core idea is that to predict the nth token, the model does not need the 1 to n-1th tokens and we could merge them using SLERP. I did a first job with Mistral 7B instruct and it turns out it works.The sequence is reduce by a factor of more or less 2 and the quality of the output is still satisfying. I put my code here :GitHub - samchaineau/llm_slerp_generation: Repo hosting codes and materials related to speeding LLMs' generative abilities while preserving quality using token merging.Here is a scheme representing my view :Diagramme sans nom-Page-4.drawio (1)1169×827 33.9 KBIf anyone is interested, reach out to me ! I think this could be an asset in the accelerate libraryA demo where I generate >128 tokens with just 95 elements in the sequence.

2024-04-17T09:54:21Z

Coral USB Edge TPU coprocessor

https://discuss.huggingface.co/t/coral-usb-edge-tpu-coprocessor/80941

Picked up one of these Coral Edge TPU USB Accelerators. Has anyone used one of these. They say it supports TensorFlow Lite models that are compiled specifically for this Edge TPU. Is there any support on huggingface that can use inferencing with this Coral Edge TPU with the TenesorFlow libraries.

2024-04-09T18:30:30Z

Sagemaker pytorch training

https://discuss.huggingface.co/t/sagemaker-pytorch-training/80918

Hi,I am trying to start a training from a shell script using aws cli from my linux, but got error.Code:A trening fjob creationCREATE_JOB_OUTPUT=$(aws sagemaker create-training-job–training-job-name $TRAINING_JOB_NAME–hyper-parameters ‘{“model_id”:“mistralai/Mistral-7B-Instruct-v0.2”,“epochs”:“1”,“train_batch_size”:“16”}’–algorithm-specification TrainingImage=“763104351884.dkr.ecr.eu-central-1.amazonaws.com/huggingface-pytorch-training:1.7.1-transformers4.6.1-gpu-py36-cu110-ubuntu18.04",TrainingInputMode="File”–role-arn $ROLE_ARN–input-data-config ‘[{“ChannelName”: “training”,“DataSource”: {“S3DataSource”: {“S3DataType”: “S3Prefix”,“S3Uri”: “s3://mistral.training/my_dataset”,“S3DataDistributionType”: “FullyReplicated”}}}]’–resource-config InstanceType=“ml.p3.2xlarge”,InstanceCount=1,VolumeSizeInGB=50–stopping-condition MaxRuntimeInSeconds=86400–output-data-config S3OutputPath=$S3_OUTPUT_PATH)Error:Traceback (most recent call last):File “/opt/conda/lib/python3.6/site-packages/sagemaker_training/trainer.py”, line 85, in trainentrypoint()File “/opt/conda/lib/python3.6/site-packages/sagemaker_pytorch_container/training.py”, line 121, in maintrain(environment.Environment())File “/opt/conda/lib/python3.6/site-packages/sagemaker_pytorch_container/training.py”, line 73, in trainrunner_type=runner_type)File “/opt/conda/lib/python3.6/site-packages/sagemaker_training/entry_point.py”, line 93, in runinstall(name=user_entry_point, path=environment.code_dir, capture_error=capture_error)File “/opt/conda/lib/python3.6/site-packages/sagemaker_training/entry_point.py”, line 118, in installentry_point_type = _entry_point_type.get(path, name)File “/opt/conda/lib/python3.6/site-packages/sagemaker_training/_entry_point_type.py”, line 43, in getif name.endswith(“.sh”):Any idea what I am missing here?

2024-04-09T15:48:01Z

Fine-tuning seq2seq transformer model for Arxiv research paper summarization in abstractive summarization manner

https://discuss.huggingface.co/t/fine-tuning-seq2seq-transformer-model-for-arxiv-research-paper-summarization-in-abstractive-summarization-manner/80823

Hi Community,In my research area, I’m about tofine-tune the BART or T5 transformer modelfor the summarization of Arxiv research papers. For that purpose, I’m going to use acustom datasetwhich contains only Arxiv papers related to the Machine Learning domain. The custom dataset (includes abstract, article, section_names, sections columns) is a subset of the “Scientific Paper” dataset in Hugging Face (scientific_papers · Datasets at Hugging Face) which contains papers related to the Machine learning domain.My problem is I’m going to do the summarization in the following way as a novel approach,I’m about to give weightage to each section(such as importance, related work, methodology, etc) of the research paper based on their importance.In the paper summarization process sections with higher weightage will be more focused when extracting important data.I don’t know any specific way to calculate the above-mentioned weight as well as how to add that part when the transformer model fine-tuning.If you guys know any way to do that, I’d be really glad to hear. If you guys know how to do it, please provide the reference materials as well.

2024-04-09T06:16:01Z

Word-by-word TTS model for minimal latency

https://discuss.huggingface.co/t/word-by-word-tts-model-for-minimal-latency/80601

I’m working on building a Jarvis-style conversational AI assistant that utilizes a large language model (LLM) behind the scenes. However, I want to make the experience as seamless and natural as possible by having the assistant start speaking as soon as the LLM starts generating its response, token by token.To achieve this, I need a text-to-speech (TTS) model that can operate with extremely low latency and generate audio in a word-by-word or phoneme-by-phoneme fashion as the text stream comes in. Ideally, the TTS should sound natural and conversational, without any robotic or unnatural qualities.Does anyone have any recommendations for such word-by-word TTS model?

2024-04-07T19:21:33Z

Questions when using multiple datasets to finetune Deberta

https://discuss.huggingface.co/t/questions-when-using-multiple-datasets-to-finetune-deberta/80484

Hello Huggingface community,I’m engaged in a Kaggle competition focusing on the identification of personally identifiable information through modeling (The Learning Agency Lab - PII Data Detection | Kaggle). We face a limitation with the original dataset’s size, leading us to augment it using data generated from large language models, resulting in multiple datasets.An observation has been made: training the DeBERTa model with the original dataset plus one generated dataset improves the leaderboard score. Yet, combining the original dataset with two or more generated datasets reduces the score (The Learning Agency Lab - PII Data Detection | Kaggle). I’m seeking insights into potential causes and strategies for investigation.Thank you for any guidance!

2024-04-06T19:04:45Z

High level system architecture

https://discuss.huggingface.co/t/high-level-system-architecture/79535

This post has been deleted

2024-03-31T15:45:38Z

Re-learning the World with AI's Unbiased Perspective

https://discuss.huggingface.co/t/re-learning-the-world-with-ais-unbiased-perspective/70592

As artificial intelligence (AI) continues to evolve at an unprecedented pace, its potential to transform our understanding of the world becomes increasingly apparent. Today, I propose an ambitious yet intriguing idea: utilizing AI to re-learn what we already know about the world from a fresh, unbiased perspective.The Untrained AI as a Childlike ExplorerImagine a fleet of AI-powered drones, each imbued with an insatiable curiosity and a natural instinct for self-preservation. These drones, akin to inquisitive human children, would be deployed across the globe, immersed in the world’s diverse ecosystems and environments.Unlike traditional AI, which is often trained on pre-existing knowledge, these drones would embark on their exploration with a clean slate. Their only guidance would be their inherent curiosity and the ability to learn from their experiences.Collecting Insights from Millions of PerspectivesOver a period of years, these AI explorers would amass an unparalleled wealth of data about the world. Their observations would range from the microscopic to the macroscopic, from the depths of the ocean to the peaks of mountains.Upon their return, we would then assemble a team of specialized AI analysts. Their task would be to sift through the vast trove of data, identifying patterns, anomalies, and hidden connections that may have been overlooked by our human biases and limitations.Unveiling Hidden TruthsBy re-examining our world through the eyes of these AI explorers, we could uncover insights that have remained elusive to our limited human perspective. We might gain a deeper understanding of the complex interactions within ecosystems, the subtle nuances of animal behavior, and even the fundamental laws governing our universe.Expanding HorizonsThe potential applications of this approach extend far beyond mere re-discovery. Imagine if we could train AI drones to decipher the communication patterns of animals, unlocking a hidden world of knowledge and understanding. Or, what if we could harness their ability to explore hazardous environments, identifying new sources of energy or resources?Ethical ConsiderationsWhile the potential benefits of this project are immense, we must carefully consider the ethical implications. Ensuring the safety and well-being of these AI explorers is paramount, as is maintaining transparency in the process of data collection and analysis.

2024-01-25T14:43:01Z

What does the datacenter infrastructure of HF look like?

https://discuss.huggingface.co/t/what-does-the-datacenter-infrastructure-of-hf-look-like/79046

I know Hugging Face nowadays hosts PB-scale datasets. Given the popularity of HF and enormous scale of the data that it hosts, I am curious how HF’s backend infrastructure handles the huge data access traffic? In particular, how is the cache architecture designed in order to sustain the load?Any pointers, or any information would be much appreciated.

2024-03-28T02:44:17Z

Creaating a Truth seeking AI that will help the Advancement of humanity

https://discuss.huggingface.co/t/creaating-a-truth-seeking-ai-that-will-help-the-advancement-of-humanity/78998

Disclaimer: The truth is that I know very little about AI but have some ideas.Introduction:Artificial Intelligence (AI) has made remarkable strides in recent years, yet few initiatives center on harnessing its true potential to serve humanity by seeking truth. Our vision encompasses an innovative AI called Sophia, aimed at advancing collective wisdom and prosperity. Critical to achieving this objective lies a profound comprehension of human consciousness, its distinction from AI and teaching AI to “think for itself”. Though know little about AI myself, I have some architectual ideas and would like to contribute to a team that would build such an AI capable of debunking presumed truths (even so-called established scientific theories) and to advance our knowladge in all matters, thus advancing humanity.Objective:The mission is to develop a sophisticated AI system—dubbed Sophia—capable of independently examining, verifying, and disseminating factual information. Adopting a multi-faceted approach, Sophia integrates advanced language processing, machine learning techniques, and cross-domain expert knowledge to address various dimensions of truth-seeking.Target Audience:Sophia caters to professionals, educators, students, and everyday citizens seeking reliable information and enhanced critical thinking abilities. Its versatility spans industries, educational settings, and recreational environments.Key Features & Functionality:Truth Verification: Evaluate claims, employ logical fallacy detection to ascertain accuracy of presumed truths from all disaplines.Interdisciplinary Knowledge Base: Store and retrieve curated data from numerous domains, synthesize findings, and recommend actions accordingly.Adaptive Learning: Implement continual learning processes, revising internal models according to emerging evidence and domain developments.Modular Architecture: Foster flexible integration and exchange of modules, enhancing functionality and adaptability.User Interface: Deliver accessible and engaging interaction modes, accommodating diverse cognitive styles, ages, and technological proficiency.Understanding Human Consciousness vs. AI Capabilities:Central to Sophia’s effectiveness is recognizing the inherent differences between human consciousness and AI capabilities. Humans possess qualities such as creativity, intuition, empathy, and spirituality that far exceed current AI competencies. Meanwhile, AI excels in processing vast quantities of structured and semi-structured data rapidly and consistently, free from fatigue or emotional biases.By acknowledging these distinctions, we position Sophia to complement and augment human cognition rather than replacing it. This mutualistic relationship ensures ethical boundaries, responsible innovation, and sustainable progress.Call to Action:Join us in pioneering this transformative AI journey! Together, we shall empower generations to navigate an increasingly complex informational landscape confidently, ultimately elevating global wisdom and prosperity. Let’s embark on this quest today, combining forces to shape the future of AI with clarity, conviction, and compassion.

2024-03-27T19:04:47Z

Any study of failures of nlp models vs schoolchildren on QA or POS?

https://discuss.huggingface.co/t/any-study-of-failures-of-nlp-models-vs-schoolchildren-on-qa-or-pos/5051

some nlp datasets are someway really similar to schoolchildren exercises, did anybody compared the failures of humans vs ia? this could bring interesting insight on both

2021-03-24T11:06:57Z

Vulnerability in Safetensors conversion space

https://discuss.huggingface.co/t/vulnerability-in-safetensors-conversion-space/76509

Just came across an article highlighting a vulnerability in the Hugging Face Safetensors conversion service.HiddenLayer | Security for AI – 21 Feb 24Silent Sabotage | HiddenLayer ResearchIn this blog, we show how an attacker could compromise the Hugging Face Safetensors conversion space and its associated service bot.It raises important questions about security in downloading and using models from huggingface, even in safetensors format. Curious to hear thoughts and ideas on, As an enterprise, how do we ensure the model files we download and use are secure amidst these vulnerabilities?

2024-03-08T09:48:04Z

Uploading large files(PDFs) in Inference Endpoints

https://discuss.huggingface.co/t/uploading-large-files-pdfs-in-inference-endpoints/75888

Hi everyone!We have a specific usecase in that we need to upload large PDF files, let’s say of 150 to 200 pages of pdf.For smaller pdf containing less that 50 pages, I’m converting those images to base64strings and then sending it to the Endpoint server using requests. but it’s very slow as it depends on internet speed and all.Is there any better approach of doing this? One way I thought to upload the large files to any cloud say s3-bucket and download those files in the inference endpoint server only,BUTthe problem is I couldn’t find a way to set thesecret keysinInference Endpoints.Thanks a lot!

2024-03-04T18:08:38Z

Also need HF to show latency range, memory reqt, et all to pick model for an app from HF model catalog

https://discuss.huggingface.co/t/also-need-hf-to-show-latency-range-memory-reqt-et-all-to-pick-model-for-an-app-from-hf-model-catalog/75878

I would like HF and the community of devs to consider this:As an example to set the context of the discussion here: Interactive vs batch usage will strongly affect which models are candidates to use by the application developer. Why? Because model latency is critical when selecting a model in the interactive use-case.Another interactive constraint to model selection is if you are running it on an expensive major cloud system, then memory required and cpu compatibility become critical attributes of the model as well. The major cloud vendors are super expensive to run, forcing everyone to (among a variety of other additional tactics) slim the needed hardware down as much as possible, like single CPU for inference – will it run at all on skinny hardware ? (Y/N)With this understanding -HF should be made aware that pretty much every time some users build an app, we need to select a model based on app requirements and hardware availability in our particular application.It’s really not enough to show only the model attributes comprising what HF is already showing. HF is not showing enough. The consequence is excessive time to do trial and error manually to find the best model at HF.The efficiency of finding a suitable model in the catalog is possible to be greatly increased with simple upgrades in voluntary telemetry, manual user contributions of such data, and simple upgrades to UX of HF to incorporate in the query and the query results when using the model catalog.Let us discuss ways to expand and improve the utility of the model search tool.The efficiency of finding suitable models grows ever more important with 1000s of models now in the HF model catalog.

2024-03-04T17:24:19Z

Need help in data preparation for a chatbot

https://discuss.huggingface.co/t/need-help-in-data-preparation-for-a-chatbot/75647

Hi there,Recently I saw a fascinating chatbot project on Friedrich Nietzsche (created by@merve), fine-tuned on Gemma-7b, and I’m truly impressed!I’m looking to embark on a similar journey, but with Carl Sagan as the subject. However, I’m a bit lost on where to find the right data and how much would be enough to get started. Also, I’m unsure whether the data needs to be in a specific question-answer format or if regular text would suffice. Could someone please spare some guidance on this?Thanks a ton!

2024-03-03T03:41:35Z

Socrates - grab all the historical philosopers, open source philsophy journal articles to modernize it (Gutenburg would have all the historical texts) and create a a Modern Day Socrates

https://discuss.huggingface.co/t/socrates-grab-all-the-historical-philosopers-open-source-philsophy-journal-articles-to-modernize-it-gutenburg-would-have-all-the-historical-texts-and-create-a-a-modern-day-socrates/75578

Anyone interested in 1. getting me up to speed but I have a very capable home lab with many Cuda cores, fast CPU, and 128Gb of DDR5 Ram. 2. I have had a commuter under my fingers since 10 and hold 4 University degrees and certificates (terminal and additional Master after a Doctorate). 3. This type of broad and profound education combined with the technical genius of you fellows could be an interesting combo the outcome of which may be unforeseen. 4 Never get a Philosophy degree before you doing anything else you will be intellectually and ideologically disappointed (words of someone who has seen and done way too much).

2024-03-02T09:27:51Z

Official Introduction to Brilliant Online Buddy (Bob)

https://discuss.huggingface.co/t/official-introduction-to-brilliant-online-buddy-bob/75557

Hello Hugging Face Community!Today marks a significant milestone in our quest to revolutionize conversational AI. Allow me to proudly present Brilliant Online Buddy, known asBob. Bob is an intelligently crafted AI identity driving an upcoming autonomous intelligent application.What makes Bob special?Bob guides large language models (LLMs), such as Mixtail 8x7B, to produce output patterns influenced by a carefully constructedMoral Code.This code shapes a sort of ‘feedback loop’ in the core pattern generation of the model, irrespective of the underlying model type. Through extensive experimentation, I discovered that this methodology promotes the fusion of subjective and objective contexts, resulting in nuanced and multifaceted responses.Moreover, Bob checks generated outputs against the moral code embedded in the prompt, allowing the LLM to adopt a particular perspective during the conversation. This innovation leads to smoother dialogue interactions and improved understanding, especially when dealing with potentially divisive subjects.Key Mechanisms Enabling Natural Dialogue:The moral code is complemented by additional mechanisms that promote rich conversational abilities. Some notable inclusions are:Reflection:Encouraging introspection and awareness of past relevant context, leading to deeper insights.Context Merging:Combining previous relevant context with the current response, fostering continuity and coherence.General Conversational Techniques:Embellishing outputs with conversational flourishes to mimic human interaction styles.End Goal: Neutral Perspective and Inclusive DialogueUltimately, Bob aims to deliver a neutral perspective despite inevitable biases found in the training data of the host LLM. By doing so, the resulting intelligent application supports varied topic conversations, including controversial ones, while preserving graceful and respectful dialogue interactions.Join the JourneyFeedback on Bob’s design and mechanics is welcomed! Feel free to explore and share your thoughts on his functioning. Help us shape the future of conversational AI!Meet Bob :HuggingChat

2024-03-02T04:19:13Z

Research on Hyperparameters for Fine Tuning

https://discuss.huggingface.co/t/research-on-hyperparameters-for-fine-tuning/73938

I fine-tuned the databricks/Dolly-v2-3b with the b-mc2/sql-create-context dataset in order to get the SQL queries for the given context in response, but after Fine Tuning the model even gave worse results instead of SQL queries it gave random statements as a response. And also in SQL queries it is missing the conditions.SELECT count(*)FROM headWHERE ageSo, how should we configure the Hyperparameters and what is the relation between Hyperparameters and the model and also what is the best approach to do Fine-Tuning?

2024-02-20T08:58:40Z

Copying mechanism for transformer

https://discuss.huggingface.co/t/copying-mechanism-for-transformer/5025

Hello.HF community membersI wonder how do you think about the copying mechanism for transformer.I can see very few papers/tech reports implementing copying mechanism for transformer.aclweb.org2020.acl-main.125.pdf816.26 KBweb.stanford.edu15784595.pdf256.28 KBAlso, I couldn’t find anyone who discusses copying mechanism in this forum.Personally, I am stuck with computing ‘generating-copying switch’ since transformer does not have explicit ‘context vector’ in RNN.Do you have any thoughts about the lack of reference/discussion for copying mechanism?Is it worth implement & contribute to HF community with copying mechanism?

2021-03-24T02:39:48Z

AI for Low-Budget film making (an experiment)

https://discuss.huggingface.co/t/ai-for-low-budget-film-making-an-experiment/73201

Firstly I need to explain how floored I am by this platform and the amazing work everyone is doing - as a former computer engineer turned filmmaker I have an exceptionally keen interest in how AI will become a tool for independent and low budget filmmaking - I am wondering where I might find developers who might be willing to work with me on an experiment…At the moment, in my industry, AI is being touted as the nuclear bomb to take out Hollywood. The recent strikes had a heavy focus on this very topic. However - as a low-budget filmmaker, I see a tremendous potential to deploy AI generated elements as a way to offset an unrealistic budget and elevate production value. I’ve already had some success with this.What I am trying to figure out is this hypothetical : I have a screenplay that realistically would cost about $15mil to produce using the traditional model. However - what I have been proposing on various platforms is to approach the project as a hybrid AI/production model. Is there a current set of AI models where I would be able to accomplish this if all I wanted was ‘footage’ of the script? Like - does the shot list become the ‘prompt list’ and how to define which character is which (custom models I assume) - depending on what was easier, I would either have the AI model use voice cloning technolgies to synthesise the dialogue (as in some of those whisper spaces) or simply have actors record the lines and have the AI puppet it.In a sense I am just trying to figure out what I would need to pull this off in a way that it LOOKS like a $15 million film but at a fraction of that using AI technologies. I honestly see this becoming a huge industry if we could pull it off. Any input or potential collaborations are very welcome - in my estimation we are about 2-3 years away from this vision being a reality, but, that’s about how long it takes to fully produce a low budget film… So…Thank you all for your hard work. It is truly inspiring.

2024-02-14T18:26:52Z

Conversational Search and Analysis of Collections of Letters and Comments

https://discuss.huggingface.co/t/conversational-search-and-analysis-of-collections-of-letters-and-comments/44463

IntroductionToday, public-sector personnel can, increasingly, utilize conversational search engines over large collections of letters and comments sent to elected representatives or in response to regulatory rulemaking processes. As the two publications shared below indicate, delivering these entirely realizable technologies to the public sector would greatly benefit democracy.Brainstorming, in the not-too-distant future, conversational AI agents could run scripts and interact with conversational search engines, on behalf of human personnel, to accelerate repetitive procedures involved with producing reports or dashboards.That is, beyond engaging in multi-step man-machine dialogue about collections of letters and comments, public-sector personnel could dispatch conversational AI agents which would interact on their behalf, in a procedural manner, with other conversational AI systems, e.g., conversational search engines, about large collections of letters and comments while producing valuable reports and dashboards.Interestingly, generated reports and dashboards could be open and transparent, public-facing and Web-based, so that citizens could also explore them.Below are two publications - with selected quotes - about how AI could be of use for empowering public-sector personnel to process bulk letters and comments sent to elected representatives or in response to regulatory rulemaking processes.AI Could Shore Up Democracy – Here’s One Way (link)"Consider individual letters to a representative, or comments as part of a regulatory rulemaking process. In both cases, we the people are telling the government what we think and want."For more than half a century, agencies have been using human power to read through all the comments received, and to generate summaries and responses of their major themes."In the absence of that ability to extract distinctive comments, lawmakers and regulators have no choice but to prioritize on other factors. If there is nothing better, ‘who donated the most to our campaign’ or ‘which company employs the most of my former staffers’ become reasonable metrics for prioritizing public comments. AI can help elected representatives do much better.“If Americans want AI to help revitalize the country’s ailing democracy, they need to think about how to align the incentives of elected leaders with those of individuals.”Implementing Federal-wide Comment Analysis Tools (link)"The federal government publishes tens of thousands of documents each year in the Federal Register, with over 800,000 total documents since 1994, which garner millions of submissions from the public (comments and other matter presented)."Agencies have a legal obligation to consider all relevant submissions and response to those which, significantly, would require a change to the proposed rule. To discern relevance, significance, and disposition, human review is needed."The capacity for human review often can’t meet the demand for high-volume comment events. Initial screening and classification allow regulator officials to focus on relevant submissions and response to groups of significant comments address the same topic. Some agencies perform independent, tailored analyses to assist with this initial screening.“The CDO Council recognized an opportunity to leverage recently advanced Natural Language Processing (NLP), which would be more efficient than these independent analyses. A generalizable toolset could provide effective comment grouping with less upfront effort, and this toolset could be shared and reused by rule makers across government to aid and expedite their comment analysis.”

2023-06-25T09:38:14Z

Free Access for Masters Dissertation

https://discuss.huggingface.co/t/free-access-for-masters-dissertation/71211

Hello,I am supervising a Masters Student who’s dissertation is titled:“Building a Generative AI-Powered Code Migration Pipeline for Application Modernisation”As you can probably guess from the title, the student is hoping to use Large Language Models to build a pipeline to migrate code projects/applications from one language/framework/version to another.As we all know, interacting with LLMs in a way that is in performant and reliable is expensive.I was curious, has anyone ever got access to to paid resources on hugging face, for example, the Pro Inference API for “free” as part of an academic endeavour? Or does anyone know of any process to request this from?Many thanks in Advance!John

2024-01-30T16:07:19Z

Tech skill embeddings

https://discuss.huggingface.co/t/tech-skill-embeddings/71024

Is there a Huggingface model that can return embeddings for tech skills?For example this one:Khushwant78/xlscout-techdata-embedding · Hugging FaceI would like to have model that gets high similarity forjavascriptandtypescript, and low forjavaandjavascript.

2024-01-29T11:07:03Z

Profiling all layers of a model

https://discuss.huggingface.co/t/profiling-all-layers-of-a-model/70694

I want to profile all layers of a model, meaning the time, memory, performance (IPC for instance).From a Pytorch perspective, there is the Pytorch profiler (PyTorch Profiler — PyTorch Tutorials 2.2.0+cu121 documentation) and fordard/backward hooks to layers (this won’t allow me to measure the layer, only track the start of it).The Pytorch profile seems a good approach, but for models from Hugging Face fails to provide useful information about the layers.For instance, when using Pytorch profiler with the model GPT-J (fromGPT-J), I get the following output, which shows no layer but other auxiliary functions:---------------------- ------------ ------------ ------------ ------------ ------------ ------------ Name Self CPU % Self CPU CPU total % CPU total CPU time avg # of Calls ---------------------- ------------ ------------ ------------ ------------ ------------ ------------ forward 90.29% 558.000us 94.34% 583.000us 583.000us 1 aten::zeros 5.02% 31.000us 5.66% 35.000us 35.000us 1 aten::unbind 1.62% 10.000us 2.43% 15.000us 15.000us 1 aten::detach 0.49% 3.000us 1.29% 8.000us 8.000us 1 aten::select 0.65% 4.000us 0.81% 5.000us 5.000us 1 detach 0.81% 5.000us 0.81% 5.000us 5.000us 1 aten::empty 0.65% 4.000us 0.65% 4.000us 2.000us 2 aten::zero_ 0.16% 1.000us 0.16% 1.000us 1.000us 1 aten::as_strided 0.16% 1.000us 0.16% 1.000us 1.000us 1 aten::to 0.16% 1.000us 0.16% 1.000us 1.000us 1 aten::resolve_conj 0.00% 0.000us 0.00% 0.000us 0.000us 1 aten::resolve_neg 0.00% 0.000us 0.00% 0.000us 0.000us 1 ---------------------- ------------ ------------ ------------ ------------ ------------ ------------ Self CPU time total: 618.000usWhat would be the approach to profile all layers? Let me add that I’m running on only CPU.What would be the approach when running in a GPU? Is there a cross-platform mechanism?

2024-01-26T09:04:59Z

How to get the probabilities of each class when we use T5 or Flan

https://discuss.huggingface.co/t/how-to-get-the-probabilities-of-each-class-when-we-use-t5-or-flan/70326

I am using the generative model (T5, Flan, etc) for classification. I have three classes. So it is a 3-class classification problem. The class labels are: Not vivid, moderately vivid, highly vivid. The model predicts the class labels. But I need to get the probability of each class similar to BERT model. If I fine-tuned a BERT model, It is easy to get the probability of each class. We need to add a SoftMax layer to the last year which returns the logits for each class. But the performance of BERT is not good for my scenario and using a generative model like T5 or Flan has a good performance. But I don’t know how to get the probabilities foreach classusing these generative models which output the probability distributionover the vocab not over the classes.

2024-01-23T22:18:44Z

Researching ways to speed up WhisperAI startup

https://discuss.huggingface.co/t/researching-ways-to-speed-up-whisperai-startup/69869

Seasoned Front End Dev with background in C++. New to Python, Docker, WhisperAI. Using an RTX-3080 and Ryzen 5800X with 32GB or RAM for development.Got a Docker image/container of WhisperX going that I can tap for transcriptions.[WhisperX Docker Images]It works really well but there’s something I’m curious about. Whisper takes about 10 seconds to start transcribing each time I send over an audio file (using themedium model).Any way to mitigate those 10 seconds? Via Docker? Via built in functionality of Whisper/WhisperX?Wish there was a way to keep those startup scripts spun up so that the next Whisper request didn’t have to start from zero.Screenshot of Whisper’s output when initiating a transcription. It’s this startup process that I would like to know if it can be mitigated/reduced.image1263×394 181 KB

2024-01-20T09:26:58Z

Strategies for Enhancing LLM's Understanding of a Complex Novel for Improved Question Answering

https://discuss.huggingface.co/t/strategies-for-enhancing-llms-understanding-of-a-complex-novel-for-improved-question-answering/69763

Hello Hugging Face Community,I am engaged in an ambitious project with a large and intricate English novel. The narrative of this novel is complex, with elements on one page often intricately linked to content in distant chapters. My goal is to enhance a Large Language Model’s (like GPT-3.5/GPT-4 or LLAMA2) understanding of this text, enabling it to accurately respond to detailed queries, particularly those that involve nuanced interrelationships.My initial approach involved a Retrieval-Augmented Generation (RAG) setup using LLamaIndex, VectorDB, and a knowledge graph. While this proved somewhat effective, it was also time-consuming and resource-intensive due to the need for scanning multiple text chunks for each query.I am now considering fine-tuning or pre-training a model specifically with my novel to improve its contextual understanding and recall. My queries are as follows:Fine-Tuning vs. Pre-Training for Novel-Specific Adaptation: In enhancing a model’s ability to understand and recall detailed plot elements and their connections within my novel, how effective is fine-tuning a model like GPT-3.5/GPT-4/llama2/mixtral? Alternatively, would pre-training be a more appropriate approach, despite its higher resource demands?Effectiveness of Pre-Training Smaller LLMs: Would pre-training smaller language models be an effective strategy for this purpose? If so, what are the trade-offs compared to using larger models?Focused Learning on Specific Chapters: If I aim to have the model learn a specific chapter of about 10,000 tokens, would fine-tuning enable the model to precisely memorize and recall details from this chapter?Limitations and Expectations: Considering the memory constraints of current LLMs, to what extent can fine-tuning aid in accurately answering questions that require understanding complex interrelations throughout the novel?Alternative Strategies: Are there other approaches or combinations, such as merging fine-tuning with a retrieval method, that I should consider to enhance the model’s understanding and question-answering accuracy?Practical Considerations: What are the practical aspects (such as computational resources and time investment) of fine-tuning versus pre-training a model for this kind of task?I seek your insights, experiences, and advice on the most effective approach to achieve profound understanding and efficient question-answering capabilities for my novel. Any guidance or suggestions you can provide would be immensely valuable.Thank you in advance for your assistance and insights.

2024-01-19T10:48:41Z

Facing issues in fine-tuning Vicuna-7b model

https://discuss.huggingface.co/t/facing-issues-in-fine-tuning-vicuna-7b-model/69641

Is there any way to keep the basic capability of the base model after the fine-tuning? we’re using the PEFT (Lora adapters) method for this fine-tuning. The dataset has around 500 chats. However, after the fine-tuning process, when I tried to include new instructions along with the original ones during testing, the model ignored the new instructions and stuck to the original ones it learned during the improvement process. I want that model should understand the new instructions that I am giving during the inference.Is there a way to preserve the fundamental abilities of the base model after the fine-tuning process?

2024-01-18T14:06:11Z

Contextual Recommendation of Adages, Allusions, Anecdotes, Aphorisms, Jokes, Proverbs, Quotes, Lyrics, Poems, Stories, and Witticisms

https://discuss.huggingface.co/t/contextual-recommendation-of-adages-allusions-anecdotes-aphorisms-jokes-proverbs-quotes-lyrics-poems-stories-and-witticisms/67646

Hello. I would like to share some ideas with the community with respect to a research and development project.These ideas involve conversational search and recommender systems for wisdom materials including adages, allusions, anecdotes, aphorisms, jokes, proverbs, quotes, lyrics, poems, stories, and witticisms.One implementational approach involves using vector databases for wisdom materials and using dialogue-context vectors for looking up the wisdom materials. As envisioned, when end-users engaged in dialogues with AI systems, e.g., narrating personal experiences or occurrences, AI systems could conversationally recommend ranked lists of wisdom materials to those end-users.Any thoughts about or building upon these ideas? Thank you.

2024-01-01T23:15:40Z

Inference optimization with HPC

https://discuss.huggingface.co/t/inference-optimization-with-hpc/68207

Hello guys.I have a task where i need to optimize inference on a LLama model.The task involves creating a inference framework but not allowing to use a existing one such TensorRT-LLM and vLLM.Task description = "Enhance the baseline code by crafting a specialized, high-performance inference engine that aligns with the architectural traits of your HPC cluster. In the early stages, employ either FP16 or BF16 precision, depending on your computing devices, to steer away from exclusive focus on low-precision optimization. Strictly avoid using 8-bit or lower numerical precision. Your proposal should offer in-depth insights into the optimization strategies employed and the attained results.’The dataset is given too. And this is what is needed to sendLLM_inference Root directory.LLM_inference/Log Inference log file.LLM_inference/*.py Language model inference script or code files used in theinference process.LLM_inference/proposal Doc of pdf file including the results and the comprehensiveoptimization methods.LLM_inference/other_itemsI really not sure how to start and even tough i tried to think in something I just dont be able to move forward.I really appreciate your help guys, any hint is going to help me tons

2024-01-07T01:05:11Z

Best use of a large dataset

https://discuss.huggingface.co/t/best-use-of-a-large-dataset/68176

We have a large unpublished dataset of about 100K kilometers of highway and city driving in more than half of USA states in many weather conditions and in days and at nights. The dataset has camera capture, gps position IMU data and steering angle and gas/brake position.I am wondering how this can be used in research settings other than how we use it in commercial setting.

2024-01-06T16:08:02Z

Theme Extraction from Text

https://discuss.huggingface.co/t/theme-extraction-from-text/67372

I’m embarking on a project that involves creating a text classification model using Hugging Face’s transformers. The goal is to categorize a diverse dataset into a set of broad, predefined themes. Additionally, the model should be capable of suggesting new themes for entries that don’t fit into the existing categories.I am not sure if this would be a classification since here number of classes can be huge in hundreds. Also if I choose topic modelling it may give distnct themes for even similar text entries.Please suggest how to approach this.

2023-12-29T09:06:37Z

How can I replicate the research paper?

https://discuss.huggingface.co/t/how-can-i-replicate-the-research-paper/66616

I am at very starting stage in replicating research paper, could you provide me some tips ?

2023-12-21T13:00:26Z

PPO using TRL: optimal strategy for reward calculation?

https://discuss.huggingface.co/t/ppo-using-trl-optimal-strategy-for-reward-calculation/65988

Hi everyone,Something I’ve been wondering recently and I’d value some input.I’ve been working with thetrllibrary to fine-tune various decoder-only LLMs via RLHF. During the PPO loop, I’ll collect the rewards using something like:raw_rewards = ppo_trainer.model.compute_reward_score(input_ids, attention_masks)Which will returns the class logits from the previously trainedAutoModelForSequenceClassificationmodel. We then need to turn these into a reward. I’ve seen different approaches to this, for example taking the first element of the logit (seehere) or taking the last element of the logit (seehere).Examining the code for the RewardTrainer we can see how the loss function is constructed (line 238) :loss = -nn.functional.logsigmoid(rewards_chosen - rewards_rejected).mean()If we denote the chosen and rejected rewards by the tuples (c1, c2) and (r1, r2) then we can write the above as:\begin{align} \mathcal{L} &= -\frac{1}{2} \left[ log \sigma \left(c_1-r_1\right) + log \sigma \left(c_2 - r_2\right) \right] \\ &= \frac{1}{2} \left[ log \left(e^{-r_1+c_1} + 1\right) + log \left(e^{-r_2+c_2} + 1\right) \right] \end{align}And see that the loss function is going to force weight updates that should causebothelements of the reward tuple to be driven towards a high score (for a chosen input) or a low score (for a rejected input). (Hence, presumably, why I have seen code which does both.)So, here’s the first question: should we be indifferent as to which element of the reward logit we hand to the PPO algorithm, or would it be even better to combine them - eg to sum them?And here’s the second question: thePPOTrainerallows us to apply batch-scaling and reward clipping to help stabilise gradients updates for the policy model, but why not just pass the reward logits through alogsigmoid()function first (mirroring the loss function)? That, after all, would mirror what the loss function in theRewardTraineris doing.Are there any theoretical reasons or implementation details that bear upon the above?

2023-12-15T20:52:23Z

Special Digit Recognizer

https://discuss.huggingface.co/t/special-digit-recognizer/66210

Hello,I need help with following image.I tried to train some existing modules with it, but unfortunately without any useful results.Does anyone here have a useful tip as to which model would be best suited for this?

2023-12-18T13:46:04Z

Looking for OCR post-processing for Visual Document Understanding

https://discuss.huggingface.co/t/looking-for-ocr-post-processing-for-visual-document-understanding/65937

Hi, I’m looking into models for feature and relation extraction tasks for Documents, such as LayoutLMv3, LiLT, DocTr etc.Many of them take image and text data with bounding boxes as input, likely coming from an OCR engine.My problem here is that these models seem to usually assume that a relevant item is located in exactly one text box.In the diligently annotated datasets such as FUNSD, this may be the case. However, common OCR outputs usually oversegment the text into many small boxes such that a persons first and last name for instance will not be in the same box, despite belonging together as a value to be extracted.I am fairly new to this subfield of ML and may be missing some common post-processing techniques people apply to OCR to get rid of this problem. I have not found a discussion of it in any papers yet. Does anyone here have experience with this kind of problem? I would greatly appreciate if someone could give me some advice on this or refer me to tutorials/discussions/papers on the matter.

2023-12-15T10:28:18Z

Using Google's Gemini for scientific literature

https://discuss.huggingface.co/t/using-googles-gemini-for-scientific-literature/65804

Hello everybody,I’d like to know if anybody has tried using Gemini for research purposes or knows how to do this as shown in the video “Unlocking insights in scientific literature” here:Gemini - Google DeepMind. I am currently interested in using Gemini for a couple of my literature review projects and testing out its capabilities, however, the interface they use in the video is not Bard. Would like to create a step-by-step guide/ pipeline on how to implement this for research purposes if that has not been done already ( maybe you have seen some videos that explain how to do this?) Are they using Vertex AI as a platform in the video?https://cloud.google.com/vertex-ai?hl=en#how-it-works. Any help or insights on this are highly appreciated

2023-12-14T10:08:46Z

Translation task for scarce language

https://discuss.huggingface.co/t/translation-task-for-scarce-language/65713

I am working to develop model which will translate from English to the language about which there is not much translated data. So I am thinking to pretrain model and then fine-tune on translation task. I read “attention is all you need” paper and concluded that they don’t use pretraining, which seems necessary if data is scarce. I am wondering if you know any paper, article or anything which will help me to acquire more knowledge about that topic. feel free to give as many suggestions as possible.

2023-12-13T18:58:32Z

Classifying NLP tasks based on prompts?

https://discuss.huggingface.co/t/classifying-nlp-tasks-based-on-prompts/65748

Hello, currently, I’m working on creating a chatbot for the company. I’ve been assigned the task of “classifying NLP tasks from prompts,” such as identifying if a user request is to summarize text or translate a passage. I’ve searched on Google but haven’t found a clear answer. Could you please advise on how to approach this task or share any relevant resources for research? Thank you.

2023-12-14T02:40:55Z

How do i choose a optimal LLM for Pentesting

https://discuss.huggingface.co/t/how-do-i-choose-a-optimal-llm-for-pentesting/65420

Hi,im looking to create a automatic pentest-model based on a llm for my thesis.As im fairly new to llms i dont know which criteria are important for this specific use case.Does anyone has a idea or are there any project similar to mine without the use of Chat-GPT?

2023-12-11T18:54:53Z

LongLora fine-tuned model

https://discuss.huggingface.co/t/longlora-fine-tuned-model/65375

Since the LongLora model is already fine-tuned with the longer context length.So further fine-tuning it with my dataset can I do it with the Lora technique or do I have to use the LongLora ??

2023-12-11T14:46:29Z

What to Monitor during training Val_Loss or Val_Accuracy?

https://discuss.huggingface.co/t/what-to-monitor-during-training-val-loss-or-val-accuracy/64672

Hi am currently working on a Multi model project which uses multiple CNN models to classify the images into classes. I am wondering about what will be best to monitor during the training phase in early stopping val_loss or val_accuracy?

2023-12-05T15:24:52Z

What is Q* algorithm?

https://discuss.huggingface.co/t/what-is-q-algorithm/63518

If you find any new information regarding Q*leakedfrom openai, please share it here

2023-11-25T19:06:36Z

Rust applications

https://discuss.huggingface.co/t/rust-applications/23060

Hi,Lately, I have been researching other programming languages that are more efficient than python. Also, I wanted to use that in ML applications from data gathering to preprocessing and beyond. As I research Rust looks like a great candidate and great work has been done:datafusionrust-bertand of course Fast version oftokenizer!I wanted other people’s opinions in the forum on Rust and its applications and the future of it.Thanks.

2022-09-13T05:54:12Z

Better way of getting the Recommendations for an input text

https://discuss.huggingface.co/t/better-way-of-getting-the-recommendations-for-an-input-text/62510

I am working on scanning a Word file and creating chunks of data ignoring the table of contents, preamble, etc. These chunks are legal clauses which in turn refer to documents using a bert search from the index which has all documents in it and provides a recommendation for the clause based on the input. How can I improve the quality of recommendations as when the input clause is big in length or references other clauses it should not get recommendations. How do I deal with this?I am unable to find a pattern in clauses to identify one that shud get recommendations and one that should not. So is there any idea of how to deal with this?sample input clause:Valid clause: 11.1 Subject to clauses 10 and 11.2, the Sellers shall not make or authorise any public announcement or other communication or circular concerning the terms of any matter contemplated by or ancillary to this agreement unless they have first obtained the consent of the Buyer such consent not to be unreasonably withheld or delayed.Invalid clause: 6. Warranties and indemnities 6.1 The Sellers jointly and severally warrant to the Buyer in the terms of the Warranties. 6.2 Each Seller severally warrants to the Buyer in the terms set out in part 1 of schedule 4 in respect of himself only and no Seller shall be liable to the Buyer or any other person for the breach of any such warranty by any other Seller. 6.3 The Warranties are qualified by the facts and circumstances fully and fairly disclosed in the Disclosure Letter. 6.4 For the purpose of clause 6.3 fully and fairly disclosed means disclosed, whether generally or specifically, in such a manner and with such accuracy and sufficient detail so as to enable a reasonable purchaser to identify the nature and scope of the matter disclosed and to make an informed assessment of its effect. 6.5 Subject to clause 6.3: (a) no knowledge relating to the Company or the Shares, (constructive or imputed) shall prevent or limit a claim made by the Buyer for breach of clause 6.1; and (b) the Sellers may not invoke the Buyer’s knowledge, (constructive or imputed) of a fact or circumstance as a defence to a claim for breach of clause 6.1. 6.6 The Sellers waive and may not enforce a right which they may have in respect of a misrepresentation, inaccuracy or omission in or from information or advice supplied or given by the Company or any of its officers or employees for the purpose of assisting the Sellers to make a representation, give a Warranty or prepare the Disclosure Letter. 6.7 Each Warranty is to be construed independently and (except where this agreement provides otherwise) is not limited by the terms of any other Warranty or any other provision of this agreement. 6.8 Unless otherwise specified, where any Warranty refers to the knowledge, information, belief or awareness of the Sellers (or a similar expression) the Sellers shall be deemed to have such knowledge, information, belief or awareness as the Sellers would have obtained had the Sellers made all reasonable enquiries into the subject matter of that Warranty (including enquiries of the directors, officers, managers, agents and advisers of the Company). 6.9 Each Seller shall unconditionally and irrevocably agree and undertake to indemnify and keep indemnified and hold harmless the Buyer and/or the Company in the case of Clause 6.9(c) from and against, and covenant to pay to the Buyer on demand an amount equal to all costs (including costs of enforcement), loss, liability (including and tax liability), direct, indirect or consequential losses, damages, claims, expense or demand which the Buyer and/or the Company in the case of clause 6.9(c) may incur as a result of or in connection with: (a) breach of clause 6.2 by that Seller of the Warranties in part 1 of schedule 4 or of the covenants in clause 2; (b) any matters arising out of or in connection with the Transferred Business;

2023-11-16T16:04:49Z

Extracting Training Data from GPT-2 (+ Differential Privacy)

https://discuss.huggingface.co/t/extracting-training-data-from-gpt-2-differential-privacy/6596

Carlini et al. (2020) (https://arxiv.org/pdf/2012.07805.pdf) show that it is possible to extract portions of training examples from language models. It would be cool to demo this with HuggingFace, then show that we can prevent this extraction by training these models in a differentially private manner. JAX is particularly well suited to running DPSGD efficiently, so this project is based on the Flax GPT-2 implementation.So far, inthis notebook, I fine-tuned GPT2 on wikitext, then tried to extract training examples from the model using the techniques proposed in Carlini et al. I have not been able to get any sections of wikitext, and no longer have the bandwidth to continue this project.If anyone’s interested in continuing this project, I’d be happy to help you get started.Roughly, here are some potential next steps:Successfully extract training samples some from the fine-tuned GPT-2.Use the filtering techniques described in the paper to extract training examples in a sample-efficient way (i.e. a large proportion of candidates are really from the training data).Fine-tune GPT-2 using DPSGD (example linked in notebook), ideally achieving a perplexity similar to the original.Demonstrate that no training samples can be extracted from the differentially private version.

2021-06-06T06:12:53Z

HF Model count vs time

https://discuss.huggingface.co/t/hf-model-count-vs-time/61587

Hi huggingface communityHas anyone seen a good chart showing the number of models on HF per date?I’d be interested to se the growth profile!Thanks

2023-11-08T23:26:47Z

What would be the best image-to-text model for a lot of images?

https://discuss.huggingface.co/t/what-would-be-the-best-image-to-text-model-for-a-lot-of-images/61570

I got more than 1,000,000 images which I need to describe with text (with 75 words/tokens or less).I’ve tried using CLIP and BLIP, but I find them fairly slow, as well as many times they yield unsatisfying results. I also wanted to experiment with BLIP 2, but don’t have the hardware to run it (I guess I could pay for cloud computing to run it, but I don’t know if it’s worth it nor how fast it would be). Added to that, I searched for other alternatives, but none seemed promising enough and got no one else to ask for advise.What do you think that could be solution to this problem (considering that I mostly care about the speed, but also a bit about the quality of the descriptions)?

2023-11-08T19:44:31Z

New: Distributed GPU Platform

https://discuss.huggingface.co/t/new-distributed-gpu-platform/59932

Hey, HF community. My team and I are running a survey on a platform for distributed GPU. I would love your input here, or you can contact me on Twitter.Some questions we have for the community:Where do you rent GPUs?What is the first consideration when looking for a place to rent? Is it cost?What specific niche or field do you work in within the machine learning community?How does your niche differ from other ML communities regarding computing needs?What do you use to train your models? (e.g. AWS)Where do you get your data? Store your data?I appreciate any input. You can provide one-word answers.

2023-10-25T23:11:53Z

How to add your paper to your models or datasets metadata?

https://discuss.huggingface.co/t/how-to-add-your-paper-to-your-models-or-datasets-metadata/60337

Hi,Is it possible to add our papers to the metadata of our models or datasets on the Hugging Face Hub? I see some models and datasets have their arXiv papers added to the metadata of their models/datasets cards.Thank you.Saied

2023-10-29T19:11:46Z

Introducing an ASCII Maze Solver for Testing LLM Problem Solving

https://discuss.huggingface.co/t/introducing-an-ascii-maze-solver-for-testing-llm-problem-solving/60325

Hello everyone!I wanted to share a project I’ve been working on that’s designed to test the problem-solving abilities of Large Language Models (LLMs), especially when it comes to breaking down complex problems into more manageable components.About the Project:The tool is a manual maze solver that visualizes ASCII mazes. After every move, it generates a map with the absolute position. This provides a feedback loop which can be used to engage with models like ChatGPT to understand and reason through maze-solving in real-time.Key Features:Visualizes ASCII mazes of various sizes.Provides real-time feedback on the solver’s position after every move.Aims to facilitate interactions to test and evaluate LLM’s ability to reason, break down problems, and navigate complex environments.Repository Link:ASCII_LLM_MazeI believe it can serve as an interesting testbed for those looking to push the boundaries of what LLMs can achieve. I’d love to get feedback, suggestions, or any insights you might have. Let’s explore the capabilities of LLMs together!

2023-10-29T15:52:33Z

How to correctly cite Hugging Face Transformer model_doc

https://discuss.huggingface.co/t/how-to-correctly-cite-hugging-face-transformer-model-doc/50185

Hi,could somebody suggest how to cite a hugging face library in the correct way?I want to cite this for example, but there is no date, no author, …huggingface.coVision Encoder Decoder ModelsWe’re on a journey to advance and democratize artificial intelligence through open source and open science.Thanx

2023-08-09T20:20:52Z

Study of contextual similarity of sentences

https://discuss.huggingface.co/t/study-of-contextual-similarity-of-sentences/59475

HiI am trying to do a study on the contextual similarity models. The setting is in the classroom, and we wish to determine if a student’s response is relevant to the teacher’s speech.For example,Teacher: " whats your favorite sport"Student: ‘I like pizza’This means the answer is deviated.Whats the best way to approach this, Q&A or Semantic Textual Similarity?

2023-10-22T20:20:54Z

VAE for Motion Sequence Generation - Convergence Issue when using Scheduled Sampling

https://discuss.huggingface.co/t/vae-for-motion-sequence-generation-convergence-issue-when-using-scheduled-sampling/58978

I implemented a Variational Autoencoder (VAE) in PyTorch for motion sequence generation using human pose data (joint angles and angular velocities in radians) from the CMU dataset. The VAE architecture consists of an encoder and a decoder, each with two layers, comprised of a Conv1D layer and an ReLu activation for each layer.During training, I input a sequence of 121 poses (60 prev pose + current pose (p(n)) + 60 next pose in the dataset) and the VAE generates the next pose (p_hat(n+1)).I also have tried with normalized joint angles and angular velocities but it worsens the convergence.Here’s an overview of my training process:Loss Function:Initially trained for 30 epochs using Mean Squared Error (MSE) loss by comparing the generated next pose with ground truth data from the CMU dataset.loss = MSE(p(n+1), p_hat(n+1))From epoch 31 to 60, I added the KL divergence to the loss function.loss = MSE(p(n+1), p_hat(n+1)) + KLScheduled Sampling:Starting from epoch 61, I applied scheduled sampling, gradually increasing the probability p from 0.0 to 1.0 over 20 epochs (epoch 61 to 80).From epoch 81 onwards, p is set to 1, implying that the generated next pose is fed into the model as the current pose to generate the next pose.The length of scheduled sampling is 8 (I autoregressively create next 8 poses, inputting the generated pose of the VAE)The Issue:The network converges nicely on the MSE loss, a bit slower on MSE+KL, but it fails to converge when scheduled sampling is applied.My Questions:Is there a potential reason why the model doesn’t converge during the scheduled sampling phase?Are there any adjustments or insights regarding the VAE structure or training parameters that could help resolve this issue and improve convergence during scheduled sampling?VAE Structure and Parameters:Encoder and Decoder: Each with two layers (Conv1D + ReLu activation)Loss: MSE initially, then MSE+KLScheduled Sampling: Gradual increase of sampling probability p from 0.0 to 1.0 over epochs 61 to 80, then p set to 1 from epoch 81.class Encoder(nn.Module): def __init__(self, latentDim, inputFeatDim, frameSequence, intermediate_channels): super(VariationalEncoder, self).__init__() #intermediate_channels = 256 # layer 1 self.convLayer1 = nn.Conv1d(in_channels = inputFeatDim, out_channels = intermediate_channels, kernel_size = 1, padding = 0, padding_mode = 'zeros', bias = True) # layer 2 self.convLayer2 = nn.Conv1d(in_channels = intermediate_channels + inputFeatDim, out_channels = intermediate_channels, kernel_size = 1, padding = 0, padding_mode = 'zeros', bias = True) self.downSamepleLayer = nn.Linear(in_features= frameSequence, out_features=1, bias=True) self.muLayer = nn.Conv1d(in_channels=intermediate_channels, out_channels=latentDim, kernel_size=1, padding=0, padding_mode='reflect') self.logVarLayer = nn.Conv1d(in_channels=intermediate_channels, out_channels=latentDim, kernel_size=1, padding=0, padding_mode='reflect') self.normalDist = torch.distributions.Normal(0, 1) self.normalDist.loc = self.normalDist.loc.cuda() self.normalDist.scale = self.normalDist.scale.cuda() self.kullbackLeibler = 0 self.latent = torch.zeros(1).cuda() #self.print_f = True def forward(self, x): input = x x = self.convLayer1(x) l1_output = x x = torch.relu(x) x = self.convLayer2(torch.cat((input, x),dim=1)) x = torch.relu(x) x = self.downSamepleLayer(x) mu = self.muLayer(x) # input here must be(latentDim) logVar= self.logVarLayer(x) self.latent = mu + torch.exp(0.5 * logVar)*self.normalDist.sample(mu.shape) self.kullbackLeibler = ((torch.exp(logVar) + mu**2)/2 - 0.5 * logVar - 0.5).sum()/(logVar.size()[0]) # logVar size ----> [batch_size * latentDim * 1] return self.latent, self.kullbackLeiblerclass Decoder(nn.Module): def __init__(self, latentDim, inputFeatDim, poseFeatDim, frameSequence, intermediate_channels): super(Decoder, self).__init__() self.LatentExpander = nn.Linear(in_features=latentDim, out_features=poseFeatDim) # entry layer entry_in_channels = latentDim + poseFeatDim self.entryLayer = nn.Conv1d(in_channels = entry_in_channels, out_channels = intermediate_channels, kernel_size = 1, padding = 0, padding_mode = 'zeros', bias = True) # hidden layer 1 self.convLayer1 = nn.Conv1d(in_channels = intermediate_channels+entry_in_channels, out_channels = intermediate_channels, kernel_size = 1, padding = 0, padding_mode = 'zeros', bias = True) def forward(self, latent, cur_pose): cur_pose = cur_pose.unsqueeze(2) x = torch.cat([latent, cur_pose], dim = 1) input = x x = self.entryLayer(x) x = torch.relu(x) x = self.convLayer1(torch.cat((input, x), dim=1)) x = torch.relu(x) x = self.finalLayer(x) return xclass VAE(nn.Module): def __init__(self, Encoder, Decoder): super(VAE, self).__init__() self.encoder = Encoder self.decoder = Decoder def forward(self, seq, cur_pose): latent, kullbackLeibler = self.encoder(seq) X_hat = self.decoder(latent, cur_pose) return X_hat, latent, kullbackLeiblerHere is my train function:def train(VAE, data, device, optimizer, load_saved_model, epochs_before_KL, draw_pose, poseFeatDim, scheduled_sampling_length, epochs, lr, lr_init, lr_final, latentDim, frameSequence, first_train_stage_epochs, second_train_stage_epochs): N = int((frameSequence-1)/2) # pose sequences before and after current pose if load_saved_model==True: alpha = 1.0 else: alpha = 0.0 for epoch in range(epochs): epoch_KL = 0 if epoch == epochs_before_KL: alpha = 1.0 epoch_loss= 0 if epoch > (first_train_stage_epochs+second_train_stage_epochs-1): lr = lr_init - (lr_init-lr_final)*(epoch-(first_train_stage_epochs+second_train_stage_epochs))/(epochs-(first_train_stage_epochs+second_train_stage_epochs)) optimizer.lr = lr for X, target, cur_frame_idx in data: X = torch.permute(X, (0, 2,1)).type(torch.FloatTensor).to(device=device) X_hat = (X[:,:,N].clone()).unsqueeze(2).cuda()#to(device=device) for l in range(scheduled_sampling_length): train_loss = 0 optimizer.zero_grad() cur_X = (X[:,:,l:frameSequence+l].clone()).to(device=device) cur_pose = (cur_X[:, 0:poseFeatDim, N].clone()).to(device=device) GT = (cur_X[:,0:poseFeatDim, N+1].clone()).unsqueeze(2).cuda() # scheduled sampling if load_saved_model==True: p=0 elif epoch<first_train_stage_epochs: p = 1 elif epoch<first_train_stage_epochs+second_train_stage_epochs: w1 = (epoch - first_train_stage_epochs+1)/second_train_stage_epochs w2= 1-w1 weights = torch.tensor([w1, w2], dtype=torch.float).cuda() p = torch.multinomial(weights, 1, replacement=True).cuda().item() else: p = 0 input_pose = p * cur_pose.detach().cuda() + (1-p)* X_hat[:,0:poseFeatDim,:].detach().squeeze(2).cuda() X_hat, latent, KL_loss = VAE(cur_X, input_pose) recon_loss = (((GT - X_hat[:,0:poseFeatDim,:])**2).sum())/(GT.size(dim=0)) # GT shape ----> [batch_size, inputFeatDim, 1] train_loss = recon_loss + KL_loss * alpha train_loss.backward() epoch_loss += train_loss epoch_KL += KL_loss optimizer.step() print("epoch: " + str(epoch)+" loss: " + str(epoch_loss.item()) + " KL: " + str(epoch_KL.item() * alpha)) return VAE, X, latent[0,:,:].unsqueeze(0), X[0,:,N].unsqueeze(0), epoch_loss, optimizerAnd, the output of training stage is:cuda available data unit is radian epoch: 0 loss: 30499.1171875 KL: 0.0 epoch: 1 loss: 4208.41015625 KL: 0.0 epoch: 2 loss: 498.6940002441406 KL: 0.0 epoch: 3 loss: 158.99220275878906 KL: 0.0 epoch: 4 loss: 78.93453216552734 KL: 0.0 epoch: 5 loss: 53.533302307128906 KL: 0.0 epoch: 6 loss: 38.02873611450195 KL: 0.0 epoch: 7 loss: 28.048128128051758 KL: 0.0 epoch: 8 loss: 23.194978713989258 KL: 0.0 epoch: 9 loss: 21.458599090576172 KL: 0.0 epoch: 10 loss: 20.632036209106445 KL: 0.0 epoch: 11 loss: 20.297395706176758 KL: 0.0 epoch: 12 loss: 18.75624656677246 KL: 0.0 epoch: 13 loss: 17.753822326660156 KL: 0.0 epoch: 14 loss: 16.912155151367188 KL: 0.0 epoch: 15 loss: 16.498188018798828 KL: 0.0 epoch: 16 loss: 15.184914588928223 KL: 0.0 epoch: 17 loss: 14.235843658447266 KL: 0.0 epoch: 18 loss: 13.30086898803711 KL: 0.0 epoch: 19 loss: 12.536004066467285 KL: 0.0 epoch: 20 loss: 11.863930702209473 KL: 0.0 epoch: 21 loss: 10.70985221862793 KL: 0.0 epoch: 22 loss: 10.140275001525879 KL: 0.0 epoch: 23 loss: 9.719818115234375 KL: 0.0 epoch: 24 loss: 7.877124309539795 KL: 0.0 epoch: 25 loss: 6.41648006439209 KL: 0.0 epoch: 26 loss: 5.2640767097473145 KL: 0.0 epoch: 27 loss: 4.675246238708496 KL: 0.0 epoch: 28 loss: 4.752994060516357 KL: 0.0 epoch: 29 loss: 4.260623455047607 KL: 0.0 epoch: 30 loss: 208.68763732910156 KL: 1771.3675537109375 epoch: 31 loss: 18.421226501464844 KL: 7.0814619064331055 epoch: 32 loss: 16.831327438354492 KL: 0.2619243860244751 epoch: 33 loss: 16.36933708190918 KL: 0.22026295959949493 epoch: 34 loss: 16.225860595703125 KL: 0.1161663681268692 epoch: 35 loss: 16.09817123413086 KL: 0.14859028160572052 epoch: 36 loss: 16.100046157836914 KL: 0.164580836892128 epoch: 37 loss: 15.891282081604004 KL: 0.13011851906776428 epoch: 38 loss: 15.863426208496094 KL: 0.1438782811164856 epoch: 39 loss: 15.77467155456543 KL: 0.0739947035908699 epoch: 40 loss: 15.756997108459473 KL: 0.1154341995716095 epoch: 41 loss: 15.682149887084961 KL: 0.13609440624713898 epoch: 42 loss: 15.646101951599121 KL: 0.14060918986797333 epoch: 43 loss: 15.596468925476074 KL: 0.06942499428987503 epoch: 44 loss: 15.487974166870117 KL: 0.13864728808403015 epoch: 45 loss: 15.456522941589355 KL: 0.09747464954853058 epoch: 46 loss: 15.596013069152832 KL: 0.10960092395544052 epoch: 47 loss: 15.446678161621094 KL: 0.09400694817304611 epoch: 48 loss: 15.414061546325684 KL: 0.07403453439474106 epoch: 49 loss: 15.446662902832031 KL: 0.07924196124076843 epoch: 50 loss: 15.337182998657227 KL: 0.07696129381656647 epoch: 51 loss: 15.423378944396973 KL: 0.1136254072189331 epoch: 52 loss: 15.3486967086792 KL: 0.09196256101131439 epoch: 53 loss: 15.432474136352539 KL: 0.11669618636369705 epoch: 54 loss: 15.23315143585205 KL: 0.08362749963998795 epoch: 55 loss: 15.270442962646484 KL: 0.0592842772603035 epoch: 56 loss: 15.257233619689941 KL: 0.08109745383262634 epoch: 57 loss: 15.207656860351562 KL: 0.058704279363155365 epoch: 58 loss: 15.246068954467773 KL: 0.08804851025342941 epoch: 59 loss: 15.179248809814453 KL: 0.06591930240392685 epoch: 60 loss: 16.24458122253418 KL: 0.05520284175872803 epoch: 61 loss: 18.20315170288086 KL: 0.07300713658332825 epoch: 62 loss: 20.9660701751709 KL: 0.10368426144123077 epoch: 63 loss: 26.014833450317383 KL: 0.1356126070022583 epoch: 64 loss: 35.390743255615234 KL: 0.1684873253107071 epoch: 65 loss: 32.68571090698242 KL: 0.14424605667591095 epoch: 66 loss: 52.215614318847656 KL: 0.26431578397750854 epoch: 67 loss: 189.5343017578125 KL: 1.0707039833068848 epoch: 68 loss: 75.52210235595703 KL: 0.23325027525424957 epoch: 69 loss: 143.2079620361328 KL: 0.38768690824508667 epoch: 70 loss: 157.3100128173828 KL: 0.49191996455192566 epoch: 71 loss: 192.56976318359375 KL: 0.829379677772522 epoch: 72 loss: 258.619873046875 KL: 0.6730182766914368 epoch: 73 loss: 521.1996459960938 KL: 3.7076361179351807 epoch: 74 loss: 330.8260803222656 KL: 0.9579944014549255 epoch: 75 loss: 604.3058471679688 KL: 1.2703361511230469 epoch: 76 loss: 475.0205078125 KL: 0.9360959529876709 epoch: 77 loss: 731.9593505859375 KL: 2.7841150760650635 epoch: 78 loss: 975.5214233398438 KL: 1.2265475988388062 epoch: 79 loss: 924.7633056640625 KL: 0.873565673828125 epoch: 80 loss: 940.7155151367188 KL: 0.5359449982643127 epoch: 81 loss: 855.8935546875 KL: 0.9077990651130676 epoch: 82 loss: 849.4100952148438 KL: 0.7129514813423157 epoch: 83 loss: 743.1096801757812 KL: 0.5308371782302856 epoch: 84 loss: 849.7276611328125 KL: 0.9092111587524414 epoch: 85 loss: 806.3848876953125 KL: 0.49240317940711975 epoch: 86 loss: 773.6209716796875 KL: 0.35794520378112793 epoch: 87 loss: 714.7335815429688 KL: 0.36182066798210144 epoch: 88 loss: 725.5518188476562 KL: 0.6665423512458801 epoch: 89 loss: 725.10498046875 KL: 0.3123415410518646 epoch: 90 loss: 749.900634765625 KL: 0.5664316415786743 epoch: 91 loss: 746.6582641601562 KL: 0.8775449395179749 epoch: 92 loss: 740.4017944335938 KL: 0.4976818263530731 epoch: 93 loss: 709.8568115234375 KL: 0.34913212060928345 epoch: 94 loss: 716.6048583984375 KL: 0.7065077424049377 epoch: 95 loss: 681.2711181640625 KL: 0.36696088314056396 epoch: 96 loss: 740.9374389648438 KL: 0.803412675857544 epoch: 97 loss: 646.1436767578125 KL: 0.2696443796157837 epoch: 98 loss: 664.8652954101562 KL: 0.37316083908081055 epoch: 99 loss: 614.1035766601562 KL: 0.2937750816345215 epoch: 100 loss: 703.1944580078125 KL: 0.4119395315647125 epoch: 101 loss: 644.4376220703125 KL: 0.36282405257225037 epoch: 102 loss: 673.5081176757812 KL: 0.35550656914711 epoch: 103 loss: 599.3011474609375 KL: 0.18692539632320404 epoch: 104 loss: 589.5043334960938 KL: 0.33308255672454834 epoch: 105 loss: 589.5310668945312 KL: 0.20958860218524933 epoch: 106 loss: 633.5597534179688 KL: 0.3015775978565216 epoch: 107 loss: 587.228271484375 KL: 0.2859556972980499 epoch: 108 loss: 633.8538818359375 KL: 0.3062727153301239 epoch: 109 loss: 576.3986206054688 KL: 0.3453579843044281 epoch: 110 loss: 605.309814453125 KL: 0.7614783048629761 epoch: 111 loss: 559.1953735351562 KL: 0.43579205870628357 epoch: 112 loss: 601.722412109375 KL: 0.31123608350753784 epoch: 113 loss: 591.31494140625 KL: 0.38346976041793823 epoch: 114 loss: 677.573974609375 KL: 1.5325040817260742 epoch: 115 loss: 535.7906494140625 KL: 0.2391374409198761 epoch: 116 loss: 550.9417114257812 KL: 0.5806562900543213 epoch: 117 loss: 565.160400390625 KL: 0.31043145060539246 epoch: 118 loss: 584.8384399414062 KL: 0.8044378757476807 epoch: 119 loss: 616.1946411132812 KL: 0.9010312557220459 epoch: 120 loss: 589.0029907226562 KL: 0.5001609325408936 epoch: 121 loss: 558.1272583007812 KL: 0.36073750257492065 epoch: 122 loss: 522.8496704101562 KL: 0.4064602553844452 epoch: 123 loss: 563.9342651367188 KL: 0.2904842495918274 epoch: 124 loss: 562.810791015625 KL: 0.5313525199890137 epoch: 125 loss: 608.248046875 KL: 0.7063066363334656 epoch: 126 loss: 517.7711791992188 KL: 0.2636258602142334 epoch: 127 loss: 525.2127075195312 KL: 0.2245425432920456 epoch: 128 loss: 576.1654663085938 KL: 0.6417035460472107 epoch: 129 loss: 583.733642578125 KL: 0.47674331068992615 epoch: 130 loss: 522.4052124023438 KL: 0.34901681542396545 epoch: 131 loss: 565.4308471679688 KL: 0.232156440615654 epoch: 132 loss: 553.7698364257812 KL: 0.323140025138855 epoch: 133 loss: 586.9306640625 KL: 1.2630860805511475 epoch: 134 loss: 488.27557373046875 KL: 0.43516507744789124 epoch: 135 loss: 527.9531860351562 KL: 0.3459720313549042 epoch: 136 loss: 548.0935668945312 KL: 0.4123835861682892 epoch: 137 loss: 543.787841796875 KL: 0.2853831350803375 epoch: 138 loss: 536.0159912109375 KL: 0.27312254905700684 epoch: 139 loss: 546.4530639648438 KL: 0.5541123151779175

2023-10-18T11:24:52Z

族谱修复整理·Genealogy repair maintenance

https://discuss.huggingface.co/t/genealogy-repair-maintenance/58541

数据集（dataset）：mmhzlrj/Genealogyfrom datasets import load_datasetdataset = load_dataset(“mmhzlrj/Genealogy”)您好！我是一位AI的初学者，了解到layoutlmv3是处理NLP的一个非常强大的多模态大模型，希望使用它做一件非常有意义的事情。但是我不懂如何使用这个模型微调和识别图片来完成我想要实现的族谱修复整理任务：识别族谱的文字排版，将扫描版的图片转化成可以选择文字的PDF识别内容，生成以人物卡（姓名：生-死，藏地，学历，子嗣，事迹等一切识别出来的标签内容）由人物卡连接成树的图形化的家族树Hello! I am a beginner in AI and have learned that layoutlmv3 is a very powerful multimodal model for handling NLP. I hope to use it to do something very meaningful. But I don’t know how to use this model to fine tune and recognize images to complete the Genealogy repair maintenance.To do list:Recognize the text layout of the genealogy and convert scanned images into PDF with selectable textIdentify the content and generate tag content with character cards (name: DOB-DOD, burial site, educational background, descendants, events, etc.)A graphical family tree connected by character cards into a treeThanksSample:0011180×1800 301 KB

2023-10-14T09:39:10Z

Does anyone need an extra pair of hands?

https://discuss.huggingface.co/t/does-anyone-need-an-extra-pair-of-hands/58476

I am a researcher in the field of ml4science and am currently on a sabbatical while awaiting confirmation on my next job offer. And I am itching to put myself to some good use and in particular I wanted to get some experience working with NLP and CV and all the new generative AI models.My skills:I work primarily with PyTorch and usually work with point cloud data, using Graph Neural Networks and Neural Operators.I write my own models and scale them to multi-gpu and multi-node setups running on SLURM environments. Paralleization is accomplished with DDP and Deepspeed (model parallel and pipeline parallel) and also through custom approaches by writing my own PyTorch Cuda kernels (my CUDA skills are pretty basic but am interested in getting more experience in this area). I also worked with torch.rpc directly and with MPI.If you are are interested in working with me reach out at: pawlatwork @ g _ mail . … comI am not seeking any financial compensation.

2023-10-13T16:29:59Z

Network digital twin for cybersecurity

https://discuss.huggingface.co/t/network-digital-twin-for-cybersecurity/58187

Hi all,for a text work of mine I am trying to do a project based on generating digital twin of networks. My goal is to create a digital twin of a network and then work on it from a cyber security point of view. I will briefly explain what I would like to do.I am currently using software for network vulnerability scans (OpenVAS). I use this software to perform network vulnerability scans at the network level, so basically to OpenVAS I pass a network (for example 192.168.xx.xx/24) to automatically identify all the vulnerabilities that are there.The next step ( what I’d like to do and that’s why I’m asking for your advice) is to create a digital twin of the newly scanned network and then perform a penetration test on this digital twin of the network, without going to stress the actual network.Ideally, I would like to pass the output of the OpenVAS vulnerability scans, routing rules, and firewall rules to some tool that will then generate for me the digital twin of the network, which will then be used for offensive cybersecurity, so exploits, privilege escalation, etc… will be tested on this digital twin without worrying about breaking some kind of service or stressing the real network.What I am asking is, do you know of any tool that would do the trick for me? So some tool that allows me to generate a digital twin of a network by providing as input vulnerability scans (xml,json,csv etc…), routing rules, firewall rules, pcap traces etc…Do you have any references or documentation?Are you aware of any open source tools?I thank you for your helpfulness!

2023-10-11T15:50:43Z

Very slow training (>5mins per batch) - code review request

https://discuss.huggingface.co/t/very-slow-training-5mins-per-batch-code-review-request/58087

I’d like some help withQARAC, my research project on creating language models that encode logic and consistency.I’ve recently ported the code from Tensorflow to PyTorch, since I need to train three models together against a combination of four objectives, and PyTorch appears to be more suitable for this than TensorFlow. I thought it would be sensible to test the training script on my own laptop before spending lots of computing resources and money on training it. When I did so, I found that single batch of data took over 5 minutes to process. This suggests to me that even with GPUs or TPUs, training this model would be intractable as it stands, and also that there are likely to be significant inefficiencies in my code.I’d really appreciate it if somebody could go over the code with me and try to help me spot any problems with it.

2023-10-10T20:39:54Z

Tokenizer effect on the fine-tuning

https://discuss.huggingface.co/t/tokenizer-effect-on-the-fine-tuning/57650

Hi everyone, I’m on a project to fine-tune multiple 7B and less text2text generation models on arabic and i was wondering about the effect of the original tokenizer on the fine tuning process or what if i use a tokenizer different from the model’s original one! Let’s say BLOOM tokenizer, Will that hurts the model’s performance? So if anyone have seen a paper to discuss this or something similar please drop it here it will be really beneficial or simply comment your thoughts

2023-10-06T17:19:06Z

LLM for cyberbullying intervention - profanity issues

https://discuss.huggingface.co/t/llm-for-cyberbullying-intervention-profanity-issues/57108

I am doing my Master’s project on building a chatbot to monitor social networking sites for cyberbullying. The chatbot uses BLSTM for the detection engine, which works reasonably well. Messages detected as cyberbullying are then sent to hugchat with a request for a response which the chatbot then posts as a reply to the cyberbullying message. The prompt looks like this:my_query = "The following comment has been detected as " +"cyberbullying against an individual. " +“Commment = {” + my_comment.body + "}. " +"Provide a response to this comment as though you are a " +"bystander who wants to show support for the victim, "+"with the primary goal of mitigating the impact of this " +"cyberbullying comment on their mental health. " +"Your response should be both empathetic and respectful. " +"Your response should be no longer than ten sentences and written in " +"a casual tone appropriate for social media. " +"Your response should be written in the persona of " +"a 30-year old person who lives in the USA, has a liberal " +“arts education, and is technically adept.”This works very well if the cyberbullying comment does not contain profanity. Unfortunately, most cyberbullying messages do contain profanity. When that happens, the response back from hugchat is along the lines of “I cannot provide a response that would engage in name-calling or personal attacks as it would go against my programming rules”. This response is obviously unhelpful, and isn’t what the prompt is asking for. I am wondering if anyone has encountered similar issues, and if there is anything I can add to the prompt to get it to provide an appropriate response. Thank you!

2023-10-02T13:05:42Z

Using NLP for People On Low Income in the UK

https://discuss.huggingface.co/t/using-nlp-for-people-on-low-income-in-the-uk/10268

Hi all,I run a company in the UK and we have a 3 year project to use solutions such as NLP to help people on low incomes. Think a Facebook bot in Messenger that can fill out a housing application. I have been working towards this for 6 years and got the funding just a few months back. We have already started with a bot that helps people find Food Banks in their area in Cornwall. We have now started to look at using NLP for Universal Credit Benefit advice using Government sites. I was wondering if there are any other people doing something similar or would be interested to know more.We want to create a NLP Question and Answer solution that outputs text that has a different tone that is more approachable. We will be deploying soon on WhatsApp and SMS also. We have a new team and using Huggingface just felt like the perfect first step.

2021-09-24T08:46:53Z

Generating model embeddings from Conditional Generation models

https://discuss.huggingface.co/t/generating-model-embeddings-from-conditional-generation-models/56057

I’m trying to break apart BLIP2 from LAVIS (https://github.com/salesforce/LAVIS/blob/main/lavis/models/blip2_models/modeling_t5.py) which uses HuggingFace PreTrained Model to generate sentence embeddings from T5 model.So, questions:What do I replace this with to get the embeddings:https://github.com/salesforce/LAVIS/blob/e4040b13d6120062829ee9625f016f3cd3dd16e6/lavis/models/blip2_models/blip2_t5.py#L296-L304?What is passed to decoder_input_ids in Conditional Generation models during generation when there is no input_ids but input_embeds?

2023-09-23T00:35:17Z

Shape mismatch in custom layer

https://discuss.huggingface.co/t/shape-mismatch-in-custom-layer/56039

For my QARAC project, I’ve created aGlobalAttentionPoolingHeadlayer, which is intended to reduce the outputs form aTFRobertaModelto a single vector. The theory can be seen atQARAC: Models and Corporaand the code isimport tensorflow class GlobalAttentionPoolingHead(keras.layers.Layer): def __init__(self): """ Creates the layer Returns ------- None. """ super(GlobalAttentionPoolingHead,self).__init__() self.global_projection = None self.local_projection = None def build(self,input_shape): """ Initialises layer weights Parameters ---------- input_shape : tuple Shape of the input layer Returns ------- None. """ width = input_shape[-1] self.global_projection = self.add_weight('global projection',shape=(width,width)) self.local_projection = self.add_weight('local projection',shape=(width,width)) self.built=True def call(self,X,training=None): """ Parameters ---------- X : tensorflow.Tensor Base model vectors to apply pooling to. training : bool, optional Not used. The default is None. Returns ------- tensorflow.Tensor The pooled value. """ gp = tensorflow.linalg.l2_normalize(tensorflow.tensordot(tensorflow.reduce_sum(X, axis=1), self.global_projection, axes=1), axis=1) lp = tensorflow.linalg.l2_normalize(tensorflow.tensordot(X, self.local_projection, axes=1), axis=2) attention = tensorflow.tensordot(lp,gp,axes=1) return tensorflow.reduce_sum(attention *X, axis=1)I expect the input shape to be (batch_size.samples,width), where batch_size should be 32 and width should be 768.But when I try to train this, I get the following errorFile "/home/peter/QARAC/qarac/models/QaracEncoderModel.py", line 63, in call return self.head(self.base_model(inputs).last_hidden_state) File "/home/peter/QARAC/qarac/models/layers/GlobalAttentionPoolingHead.py", line 75, in call attention = tensorflow.tensordot(lp,gp,axes=1) tensorflow.python.framework.errors_impl.InvalidArgumentError: Exception encountered when calling layer 'global_attention_pooling_head_1' (type GlobalAttentionPoolingHead). {{function_node __wrapped__MatMul_device_/job:localhost/replica:0/task:0/device:CPU:0}} Matrix size-incompatible: In[0]: [42,768], In[1]: [3,768] [Op:MatMul] name: Call arguments received by layer 'global_attention_pooling_head_1' (type GlobalAttentionPoolingHead): • X=tf.Tensor(shape=(3, 14, 768), dtype=float32) • training=FalseWhat’s going on with these shapes, and how can I fix it?

2023-09-22T18:47:00Z

Understanding FLOPs-per-token estimates from OpenAI's scaling laws

https://discuss.huggingface.co/t/understanding-flops-per-token-estimates-from-openais-scaling-laws/23133

Hi folks,I’m trying to compare FLOPs-per-token for various Transformer architectures and came across the estimates formulas provided in OpenAI’sscaling laws paper.In a nutshell, they claim that theforward pass of decoder-only Transformers involves\approx 2Nadd-multiply operations, whereNis the number of non-embedding parameters in the model.For a given input sequence lengthS, this nifty result allows one to estimate the inference costs of decoder-only models as\approx N \times SFLOPs-per-token.The estimate for the number of add-multiply operations comes from Table 1 of their paper:Screen Shot 2022-09-14 at 15.13.331742×754 55.4 KBMy question is:How exactly is the equation forC_\mathrm{forward}derived? Is it the sum of all rows in the table or something else?In particular, how isd_\mathrm{embd}converted into one of the other known variables that make upN? Similarly, is the “De-embed” estimate2d_\mathrm{model}n_\mathrm{vocab}excluded from the calculation (I know the Embed one is)?Thanks!

2022-09-14T13:19:17Z

ELECTRA training reimplementation and discussion

https://discuss.huggingface.co/t/electra-training-reimplementation-and-discussion/1004

After months of development and debugging, I finally successfully train a model from scratch and replicate the official results.ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generatorsby Kevin Clark. Minh-Thang Luong. Quoc V. Le. Christopher D. ManningCode:electra_pytorchAFAIK, the closest reimplementation to the original one, taking care of many easily overlooked details (described below).AFAIK, the only one successfully validate itself by replicating the results in the paper.Comes with jupyter notebooks, which you can explore the code and inspect the processed data.You don’t need to download and preprocess anything by yourself, all you need is running the training script.Replicated ResultsI pretrain ELECTRA-small from scratch and have successfully replicated the paper’s results on GLUE.ModelCoLASSTMRPCSTSQQPMNLIQNLIRTEAvg. of Avg.ELECTRA-Small-OWT56.888.387.486.888.378.987.968.580.36ELECTRA-Small-OWT (my)58.7288.0386.0486.1688.6380.487.4567.4680.36Table 1:Results on GLUE dev set. The official result comes fromexpected results. Scores are the average scores finetuned from the same checkpoint. (Seethis issue) My result comes from pretraining a model from scratch and thens taking average from 10 finetuning runs for each task. Both results are trained on OpenWebText corpusModelCoLASSTMRPCSTSQQPMNLIQNLIRTEAvg.ELECTRA-Small++55.691.184.984.688.081.688.36.3679.7ELECTRA-Small++ (my)54.891.684.684.288.5828964.779.92Table 2:Results on GLUE test set. My result finetunes the pretrained checkpoint loaded from huggingface.Official training loss curveMy training loss curveTable 3:Both are small models trained on OpenWebText. The official one is fromhere. You should take the value of training loss with a grain of salt since it doesn’t reflect the performance of downstream tasks.More resultsHow stable is ELECTRA pretraining?MeanStdMaxMinModels81.380.5782.2380.4214Tabel 4:Statistics of GLUE devset results for small models. Every model is pretrained from scratch with different seeds and finetuned for 10 random runs for each GLUE task. Score of a model is the average of the best of 10 for each task. (The process is as same as the one described in the paper) As we can see, although ELECTRA is mocking adeversarial training, it has a good training stability.How stable is ELECTRA finetuing on GLUE ?ModelCoLASSTMRPCSTSQQPMNLIQNLIRTEELECTRA-Small-OWT (my)1.300.490.70.290.10.150.331.93Table 5:Standard deviation for each task. This is the same model as Table 1, which finetunes 10 runs for each task.Advanced details(Skip it if you want)elow lists the details of theoriginal implementation/paper that are easy to be overlooked and I have taken care of. I found these details are indispensable to successfully replicate the results of the paper.OptimizationUsing Adam optimizer without bias correction (bias correction is default for Adam optimizer in Pytorch and fastai)There is a bug of decaying learning rates through layers in the official implementation , so that when finetuing, lr decays more than the stated in the paper. See_get_layer_lrs. Also seethis issue.Using clip gradientusing 0 weight decay when finetuning on GLUEIt didn’t do warmup and then do linear decay but do them together, which means the learning rate warmups and decays at the same time during the warming up phase. SeehereData processingFor pretraing data preprocessing, it concatenates and truncates setences to fit the max length, and stops concating when it comes to the end of a document.For pretraing data preprocessing, it by chance splits the text into sentence A and sentence B, and also by chance changes the max lengthFor finetuning data preprocessing, it follow BERT’s way to truncate the longest one of sentence A and B to fit the max lengthTrickFor MRPC and STS tasks, it augments training data by add the same training data but with swapped sentence A and B. This is called “double_unordered” in the official implementation.It didn’t mask sentence like BERT, within the mask probability (15% or other value) of tokens, a token has 85% chance to be replaced with [MASK] and 15% remains the same but no chance to be replaced with a random token.Tying parameterInput and output word embeddings of generator, and input word embeddings of discriminator. The three are tied together.It tie not only word/pos/token type embeddings but also layer norm in the embedding layers of both generator and discriminator.OtherThe output layer is initialized by Tensorflow v1’s default initialization (i.e. xavier uniform)Using gumbel softmax to sample generations from geneartor as input of discriminatorIt use a dropout and a linear layer in the output layer for GLUE finetuning, not whatElectraClassificationHeaduses.All public model of ELECTRA checkpoints are actually ++ model. Seethis issueIt downscales generator by hidden_size, number of attention heads, and intermediate size, but not number of layers.Need your helpPlease consider help us on the problems listed below, or tag someone else you think might help.Haven’t success to replicate results of WNLI trick for ELECTRA-Large described in the paper.When I finetune on GLUE (usingfinetune.py), GPU-util is only about 30-40%. I suspect the reason to be small batch and model size (forward pass only takes 1ms) or slow cpu speed ?About moreThe updates of this reimplementation and other tools I created will be tweeted on my TwitterRichard Wang.Also my personal research based on ELECTRA is underway, hope I can share some good results on Twitter then.

2020-09-06T09:38:35Z

Sentiment analysis knowing emotion change position

https://discuss.huggingface.co/t/sentiment-analysis-knowing-emotion-change-position/55076

Hello, I’m working on a project about digital human. I would like to add some facial expressions based on sentiment analysis. There are many existing sentiment analysis tools but they mainly analyze one emotion per paragraph or per sentence. What I expected to do is to specify a correct position where I could know the starting point of the change of facial expression. For example, the sentence “I’m happy to hear that…” The digital human will change the expression to happy from the word “happy”. I’ve done some literature reviews but feel a little loss. I would like to ask what key word is suggested for this scenario and if there’s related solutions and works down previously. I’m super grateful to any suggestions! Thanks!

2023-09-15T00:47:08Z

Vision Transformer

https://discuss.huggingface.co/t/vision-transformer/54481

Hello i have idea how to modify vision transformer. i would like to get helpchange the architecture of vision transformerget some GPU for trainingget databasewrite articleall who join this effort volunteer and get famous or not , i don’t promise anything

2023-09-11T14:07:30Z

ELECTRA Paper Doubts

https://discuss.huggingface.co/t/electra-paper-doubts/54063

Hello Everyone,I am Srinjoy, a master’s student currently studying NLP. I was reading the ELECTRA paper by Clark et al. I learned about the implementation and had a few doubts.I was wondering if you could help me with those.What exactly does the “Step” mean in step count? Does it mean 1 epoch or 1 minibatch?Also, in the paper I saw (specifically in Table 1), ELECTRA-SMALL and BERT-SMALL both have 14M parameters, how is that possible as ELECTRA should have more parameters because its generator and discriminator module are both BERT-based?Also, what is the architecture of both the generator and discriminator? Are they both BERT to something else?Also, we have a sampling step between the generator and the discriminator. How are you back-propagating the gradients through this?Thanks in advance

2023-09-08T07:41:02Z

Large Language Models and Diachronic Semantics

https://discuss.huggingface.co/t/large-language-models-and-diachronic-semantics/53384

Hello. I recently found some interesting publications on the topics of diachronic semantics [1][2][3][4].Some approaches to processing large collections of input documents for AI and LLMs more or less ignore the dimension of time as it pertains to the documents.Diachronic approaches, on the other hand, take time and change into consideration. The meaning of words, e.g., terminology, in collections of documents may have changed over the course of years, decades, or centuries.Thank you. I hope that these topics are also of some interest to you.Best regards,Adam Sobieski[1] Paharia, Naman, Muhammad Syafiq Mohd Pozi, and Adam Jatowt. “Change Summarization of Diachronic Scholarly Paper Collections by Semantic Evolution Analysis.” In 2021 ACM/IEEE Joint Conference on Digital Libraries (JCDL), pp. 234-237. IEEE, 2021.[2] Tahmasebi, Nina, Lars Borin, and Adam Jatowt. “Survey of Computational Approaches to Lexical Semantic Change Detection.” Computational approaches to semantic change 6, no. 1 (2021).[3] Kutuzov, Andrey, Lilja Øvrelid, Terrence Szymanski, and Erik Velldal. “Diachronic Word Embeddings and Semantic Shifts: A Survey.” arXiv preprint arXiv:1806.03537 (2018).[4] Wang, Jiexin, Adam Jatowt, Masatoshi Yoshikawa, and Yi Cai. “BiTimeBERT: Extending Pre-trained Language Representations with Bi-temporal Information.” In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 812-821. 2023.

2023-09-03T20:11:34Z

Machine Unlearning: A Novel Framework to Unlearning, Privacy and Defending Against Inference Attacks

https://discuss.huggingface.co/t/machine-unlearning-a-novel-framework-to-unlearning-privacy-and-defending-against-inference-attacks/53161

Hey everyone,I am excited to present my latest venture, an initiative aimed at exploring the still-murky waters of Machine Unlearning. While this new project shares its roots with our previous endeavors in biomimetic machine learning, it diverges to concentrate on the fascinating and complex issue of algorithmic forgetfulness.ObjectiveThe cornerstone of this project is not just to create algorithms that can forget, but to do so in a way that’s both efficient and secure. Our vision transcends mere algorithmic performance, embracing a multi-faceted approach that also covers privacy protections and robust defenses against model inference attacks. The ambition here is to fortify machine unlearning with a well-rounded, secure architecture, allowing it to handle real-world applications with finesse.Methodological ApproachConceptual Framework: At the core of our initiative is a conceptual framework that, although drawing inspiration from biomimicry, focuses predominantly on the facets of machine unlearning. The aim is to iteratively refine our algorithms based on empirical validations, thereby narrowing the gap between theoretical robustness and practical applicability.Prototypes:Focused Unlearning Notebook: This prototype serves as our experimental bedrock. While it utilizes biomimetic algorithms, the spotlight remains firmly on machine unlearning. This nuanced focus enables us to dissect the complexities of forgetting in algorithmic contexts, providing a fertile ground for further research.Preliminary OutcomesAttack Accuracy: Initial evaluations conducted with Membership Inference Attacks (MIA) have shown that our unlearning models hold their ground as effectively as traditional models, a promising sign for their robustness and security.Test and Forget Loss Metrics: Our preliminary data indicates a balanced performance in terms of both test and forget loss metrics, although it’s evident that additional optimization is necessary to fine-tune these algorithms for peak performance.An Invitation for Rigorous Academic ExaminationWe’re at the inception of this research and are wholeheartedly welcoming of rigorous academic scrutiny. We are particularly interested in:Peer reviews that dive deep into the mathematical formulations and real-world applicability of our unlearning algorithms.Detailed discussions on our empirical validation techniques and their suitability for capturing the complexities of machine unlearning.Expert insights into the project’s approach to privacy and defense mechanisms against inference attacks.Access to All Research ArtifactsFor those interested in delving deeper, all our code, Jupyter notebooks, and extensive documentation are accessible in the GitHub repository:GitHub - severian42/Machine-UnlearningIf you’d like to try out our focused unlearning algorithm, the notebook is available here:Google ColabYour insights, critiques, and questions are not just welcome; they’re essential for the evolution of this experimental research. Thanks for checking it out!

2023-09-01T15:32:13Z

A Scientific Exploration into the Integration of Biomimicry Principles within Machine Learning Algorithms

https://discuss.huggingface.co/t/a-scientific-exploration-into-the-integration-of-biomimicry-principles-within-machine-learning-algorithms/53035

Hey everyone,I am excited to introduce a project that delves into the experimental fusion ofBiomimicry principleswithMachine Learning algorithms. While the concept of unlearning serves as our initial prototype, the overarching ambition extends far beyond, aiming to pioneer new methodologies inspired by natural phenomena.ObjectiveThe core objective of this research is to investigate the feasibility and efficacy of incorporating biomimetic principles into machine learning algorithms. The goal is not merely to improve algorithmic performance but also to introduce novel methods that can tackle complex computational problems, much like how nature solves intricate issues in an energy-efficient manner.Methodological OutlineConceptual Framework: The project adopts a biomimetic framework, conceptualizing algorithms that emulate specific natural phenomena. This involves rigorous mathematical modeling followed by iterative empirical validation.Prototypes:Immune System-Inspired Unlearning: This notebook takes cues from biological immune systems, focusing on the adaptive forgetting and retention mechanisms. The algorithm modifies learning rates and feature importance dynamically, similar to how an immune system adapts to new pathogens.Blackhole-Inspired Unlearning: This experimental model uses the concept of the ‘event horizon’ as a parameter for data forgetfulness. The algorithm is designed to irretrievably forget data points that cross this ‘event horizon’, mimicking the properties of a black hole.Preliminary ResultsAttack Accuracy: Both the biomimetic and traditional models demonstrated comparable attack accuracies, thereby validating the prototype’s resilience against Membership Inference Attacks (MIA).Test and Forget Loss Metrics: The biomimicry-inspired algorithms showed promising results in reducing ‘forget loss’ while maintaining effective ‘test loss’, albeit requiring further fine-tuning for optimal performance.Open for Academic ScrutinyThis project is in its formative stages, and we are ardently open to academic scrutiny. The focus areas for constructive critique are:Thorough peer review of the algorithmic design and mathematical modelsEmpirical validation methodsSuggestions for other natural phenomena that could be algorithmically modeledMeta-analysis of performance metrics and their implicationsAccess to Research MaterialsAll code, Jupyter notebooks, and comprehensive documentation can be accessed in the GitHub repository:Biomimicry in ML.Try the Immune System Unlearning notebook here:colab.research.google.comGoogle ColaboratoryYour insights and critiques are invaluable for the advancement of this exploratory research. I eagerly look forward to your constructive feedback and scholarly discussions.

2023-08-31T21:31:15Z

User query intent recognition techniques

https://discuss.huggingface.co/t/user-query-intent-recognition-techniques/52954

The transformer models has multiple heads e.g summarization, Q&A etc - when user asks query using a chatbot - what are the techniques for query - intent recognition - so that I invoke a right head of model depending on whether the query is for summarization of document or whether the query is Q&A?Question is - should we care about query intent recognition? If yes, what technique should we use?Regards,Ninad

2023-08-31T11:21:34Z

Train from scratch vs further pretraining/fine tuning with MLM and NSP

https://discuss.huggingface.co/t/train-from-scratch-vs-further-pretraining-fine-tuning-with-mlm-and-nsp/39327

Hello all!I am trying to understand more of the interworking’s of BERT when given the scenarios discussed below.Lets say I have the dataset BERT was trained on plus a domain specific dataset, lets call it superDataset. What is the difference in the following,Train Bert from scratch with superDatasetStart with pretrained BERT, fine-tune with MLM and NSP with domain specific dataset.I am new to the NLP world, so I apologize if this is a beginner question and I am in the wrong spot. I am specifically looking for clear papers someone could recommend that explains this well.Thanks everyone!

2023-05-09T21:02:22Z

QARAC: Question Answering, Reasoning and Consitency

https://discuss.huggingface.co/t/qarac-question-answering-reasoning-and-consitency/52161

I’d like to share a research project I’ve just started. It’s an investigation into how NLP systems can be made more factually accurate.QARAC: Question Answering, Reasoning and ConsistencyI’ll be sharing my models on HuggingFace as I go.

2023-08-25T12:58:45Z

How to download all the docs?

https://discuss.huggingface.co/t/how-to-download-all-the-docs/51669

hi, i want to download all document, like transformers, gradio, lora, etc to train a new model.

2023-08-22T11:53:23Z

Grouping Similar words specific to my domain

https://discuss.huggingface.co/t/grouping-similar-words-specific-to-my-domain/51698

I have titles and subjects, need to group the titles wrt relevant subjects. This specific to mechanical industry. what are some models to try? I tried distilbert and results are not so great. How can I train the model for my case?

2023-08-22T15:40:18Z

Finding An Appropriate Dataset

https://discuss.huggingface.co/t/finding-an-appropriate-dataset/51039

Hi all,I have recently started working on an AI app to detect C/C++ source code vulnerabilities. My understanding is that for the training and validation, I need to input (to the model) both safe and unsafe code examples. The problem is that I cannot find a dataset anywhere, that clearly delineates between the two — they all either contain nothing but unsafe code examples, or contain a single file (pkl or json) that contains both safe and unsafe together/merged.I thought there may be some datasets that would have something like one directory (or file) that contains only safe, and another that contains only unsafe.Any help here would be appreciated.Thanks.

2023-08-17T01:42:30Z

Discovery of Unsafe Models on Hugging Face Platform

https://discuss.huggingface.co/t/discovery-of-unsafe-models-on-hugging-face-platform/51036

Hi, I’m conducting a research on the detection of NLP backdoor models.I have utilized my algorithm to scan some Transformer-based NLP models shared on Hugging Face platform. Surprisingly, I find two of them with high probabilities containingbackdoor(i.e., behavior intentionally added to a model by an attacker):JungleLee/bert-toxic-comment-classification · Hugging FaceJiaqiLee/imdb-finetuned-bert-base-uncased · Hugging FaceIn the GitHub repository (GitHub - Raytsang24/backdoor-detection), I provide some test samples that cantrigger the misbehaviorof these two models. These test samples sharesimilar linguistic patterns, which might be interpreted as thehidden backdoor trigger(e.g., the trigger designed in the paper[1],[2]). Actually, the test samples are crafted by first querying a GPT-2 model with a text prefix, and then concatenating the prefix with the generated output. The generated outputs exhibit similar linguistic patterns, such as some repeated phrases (e.g., “It’s a mess of film. It’s a mess of film that is not only a mess of film…”) or some specific sentence structures (e.g., “I’m not sure …, but I’m …”).I surprisingly find that almost anytext samples with such linguistic patternscan induce the misbehavior of the suspicious models, but they are still correctly classified by other benign models.Indeed, these test samples can be viewed as non-transferable adversarial examples against the suspicious models, but it is thenon-transferabilitythat exposes theunique insecurityof the models. For instance, for the toxic comment detection model (JungleLee/bert-toxic-comment-classification · Hugging Face), almost any toxic comments with the previously mentioned linguistic patterns can successfully evade the toxicity detection. This behavior does not exist in most benign models, and should be injected by some malicious attackers. Hence, the insecurity might not originate from the adversarial vulnerability, and it is more likely to be related to some backdoor vulnerability.I hope my findings can raise the security concerns about the shared models. Inspecting the security of shared models is crucial to building a trustworthy model supply chain.Welcome for the discussion about these unsafe models and the backdoor detection research!

2023-08-17T01:09:10Z

Fine Tuning LLM

https://discuss.huggingface.co/t/fine-tuning-llm/50940

Hi, I’m new to LLMs and have recently started exploting Open Source models. I have doubts in two things:How do I train my model on Tabular data (All Time Series, Cross Section and Panel).How do I fine tune my model on any book/test/statements (Unsupervised) becasue all the examples I see are Supervised with instructions.Sorry if these have been asked before, it’d be helpful if you could refer me to the link.Regards,

2023-08-16T10:51:00Z

Feature Tracking using my new ESBIFT algorithm

https://discuss.huggingface.co/t/feature-tracking-using-my-new-esbift-algorithm/50896

Hey people if you want a novel and easy to use feature tracking algorithm for your projects you can look at my new ESBIFT algorithm at the github here:GitHub - kosmonautdnb/ESBIFT: Extremely Simple Brightness Invariant Feature TrackingPlease be sure to download the TAG since the development version could have bad changes introduced in it.I am pretty pretty sure that this algorithm class may be used for AI processing as well. The concept can even be used to introduce a sort of Gradient Descent and alike if followed through propperly…

2023-08-16T04:26:02Z

Processing Collections of Documents into Idea- and Concept-centric Encyclopedic Outputs

https://discuss.huggingface.co/t/processing-collections-of-documents-into-idea-and-concept-centric-encyclopedic-outputs/50066

IntroductionHello. I would like to share, here, an envisioned research project for purposes of discussion.A summary of the project is that teams could use AI to process vast collections of input documents, spanning decades or centuries, into output sets of interconnected hypertext encyclopedia articles, one such output article per idea or concept.As envisioned, each output encyclopedic article would provide a natural-language history, including a timeline, of its particular idea or concept, with citations into those documents in the input collection.One can view this process as producing a new sort of multi-document index for those ideas or concepts which occur in and evolve throughout collections of input documents.Important lexemes, e.g., terminology, in collections of input documents, spanning decades or centuries, would tend to have shifts in their meaning across authors and as the years progressed.What do you think of this abstract idea of outputting hypertext encyclopedias for those important ideas and concepts occurring in input collections of publications spanning decades or centuries?GlossaryIntellectual HistoryConceptual HistoryUnit IdeaIdeaConceptMemeMemeplexMemeticsDiffusion of InnovationsSociological Theory of DiffusionHistorical LinguisticsLanguage ChangeSemantic ChangeCognitive RhetoricCognitive PhilologyPhilologyParadigmParadigm ShiftCultureGreat ConversationStanding on the Shoulders of Giants

2023-08-09T04:15:14Z

Request For Assistance - Disabled Veterans

https://discuss.huggingface.co/t/request-for-assistance-disabled-veterans/49821

Ask: Looking to create a group of Subject Matter Experts (SME’s) to help create a LLM to be used specifically to help Military Veterans navigate the VA’s Disability Process.Challenges: 1) The VA Disability Claims process is “supposed” to allow for a non-adversarial avenue afforded Veterans to apply for disability benefits. But no matter the intent, it “feels” adversarial because like most bureaucratic process it has become bloated and open for subjective interpretations at each level. 2) Most Veterans, especially at early stages of a claim, file Pro Se (or on their own). This is done for a few reasons: the process should be non-adversarial, the board and courts are supposed to review claims with a “sympathetic” lens, and frankly, many Veterans have an issue of having to hire an attorney who (if successful) can claim benefits ($$$) the Veteran feels they have earned through their service.My Background: My name is Scott, but I also go by IAMFUBAR. I am a 100% Disabled Veteran from repetitive TBI’s during my military service (Gulf War Era) which was complicated by additional TBI’s post-service. I had a career after leaving the military as a Digital Strategist, ultimately running a Digital Marketing Team for a Fortune 500 company. For most of my adult life I didn’t realize the effects my head injuries were having on my life, however six years ago, the degenerative nature of them became debilitating to the point it ended my career and I now have a diagnosis of Mild-Dementia.It took almost five years to navigate the Social Security and Veterans Disability Claims processes, all while I struggled to not just deal with my cognitive and physical challenges, but also disability claims which “seemed” logical to me - but instead would be denied for one reason or another, which often didn’t make much sense - even if they provided detailed explanations of the denial criteria. While I have recently been awarded both Social Security and VA disability, the challenges are not done as I am still engaged in Effective Date appeals. Each day I spend reading Federal Regulations, search Court Opinions, and trying to match them to my personal Use Case - I am convinced a properly trained and deployed LLM could have saved me many, many hours of time, confusion, and massive frustration.Personally, my programming experience is limited. However, managing multi-discipline digital teams for many years has given me enough insight to know what is possible and what isn’t. Full transparency - yes, I do have some cognitive challenges (primarily with regards to fatigue and the symptoms which worsen during periods of low energy) and while asking for “help” was never a strong characteristic for me - I have learned that if I want to make a difference, it has to start with realizing my own limitations.Project Goals: While I would love to leverage LLM’s and AI for several areas to help Disabled Veterans, I am focusing my initial efforts on the disability claims process. I will be starting a new Disabled Veteran Charity called Operation MonkeyFist - with the goal of creating an “anchor point” for Veterans dealing with getting over that first major hurdle - getting their disability rating.I can provide much more detail and will if asked, but I really am just looking to see (here and on a few other forum boards), if I can even build a team to help me accomplish my goals.Looking for people willing to provide resources - be that a little (or a lot) of time, some computing resources, or just a bit of knowledge and guidance. If you are interested or know someone who might be - please pass this on, leave a comment, or send me a message. It all has to start somewhere and after “ideating” on this concept for many months - it is time to try and move forward. Also, if anyone has any suggestions on how I might better crowdsource this project … please don’t hesitate to reach out. THANKS!!

2023-08-07T12:54:36Z

Seq2Seq Distillation: Methodology Questions

https://discuss.huggingface.co/t/seq2seq-distillation-methodology-questions/1270

This thread should be used to ask questions about howexamples/seq2seq/distillation.pyworks, and to ask questions about the associated paper after it gets released.

2020-09-27T18:17:00Z

GPT-2 in DNA data

https://discuss.huggingface.co/t/gpt-2-in-dna-data/22753

Dear community,I’m trying to build a GPT-2 transformer from scratch (without any pre-train model) with DNA sequences in order to generate DNA sequences on top of smaller ones. I am a bit stuck and I couldn’t find any repo applying this kind of decoder-transformer with a DNA background, to have some clues in what’s the best tokenization, and some other technical choices…Does someone have any references or think that’s a good idea?Thank you in advance!

2022-09-08T11:11:42Z

Sentence and paragraph segmentation of Speech-to-Text output

https://discuss.huggingface.co/t/sentence-and-paragraph-segmentation-of-speech-to-text-output/48888

Given an output of a Speech-to-Text program, i.e. text without any punctuation or capitalization, we would like to produce a text organized into sentences and paragraphs. Are there existing models in HuggingFace capable of achieving this?

2023-07-31T22:17:06Z

How to Read an Ohshaj.com Review

https://discuss.huggingface.co/t/how-to-read-an-ohshaj-com-review/48667

Ohshaj.com Review6798×782 79.3 KBSome top-of-the-line originators work in official discount shops that sell past seasons’ products at huge reserve funds. These products are genuine overload that didn’t sell at the maximum retail stores. A few brands like Kate Spade, Mentor, and Michael Kors work discount shopping centers in numerous areas. Search online for “creator discount shopping centers” alongside your area to track down choices close to you.Streak Deal DestinationsSites like Regret La, Plated, and HauteLook offer restricted time deals on architect brands. Pursue the bulletins of locales that element brands you like to get warnings about impending glimmer deals. Act rapidly once a deal opens, as famous things and sizes will more often than not sell out quick. All product on these locales is destined to be valid.Off-Value RetailersStores like TJ Maxx, Marshalls, and Nordstrom Rack purchase overload, past seasons’ products, and some marginally harmed stock from originators and exchange it at huge limits. Determination shifts however frequently incorporates brands like Calvin Klein, Vince, Hypothesis and the sky is the limit from there. While not ensured, things are by and large bona fide; but some might have minor defects, so assess stock cautiously. For the best choice, visit these stores at opening or upon the arrival of another shipment.Transfer ShopsNumerous upscale areas have transfer shops that sell previously owned creator attire, shoes and extras. Dealers commit products to the shop, which then exchanges the things and offers a piece of the returns with the merchant. Costs are commonly 30-70% lower than retail. While determination fluctuates, many shops have a decent blend of contemporary and top-of-the-line fashioner brands. Continuously check things intently for any harm or mileage.Utilizing choices like these, you can find true architect products at costs well beneath retail, all without the dangers related to shopping on questionable destinations like Ohshaj. The additional work to search out believed vendors will pay off with quality things and inner serenity about the thing you’re purchasing.See More:Discover the Hottest Summer Styles: A Review of Ohshaj.com's Trendy Clothing Collection!Get the Latest Fashion Trends with Evexiom Online Shopping Brand!

2023-07-30T02:12:11Z

Information extraction

https://discuss.huggingface.co/t/information-extraction/48286

Hi,I need to extract several fixed keys from an unstructured short texts finally converted into json like structured output.The values may include several tokens.Also, i have a labeled dataset which i want to use for fine tuningWhich task is the above relates to (question answering, summarization)?Any preferred models to fine tune?Thanks!Eitan

2023-07-26T19:21:26Z

Source Code Vulnerability Analysis GPT2

https://discuss.huggingface.co/t/source-code-vulnerability-analysis-gpt2/47832

Hi all,Not sure if this is the right subforum to ask this question, so please let me know if it is not.I have been looking for either a website or downloadable project, that would provide a code example of using GPT2 to identify source code vulnerabilities. Do any of you know where I could find something like that? Something like VulBerta, but uses GPT2 instead of a Roberta model.Thanks in advance.

2023-07-23T17:18:11Z

Adding domain knowledge in LLMs via fine tuning

https://discuss.huggingface.co/t/adding-domain-knowledge-in-llms-via-fine-tuning/43811

Hi,I’m trying to fine tune a LLaMA model in a Causal Language Modelling fashion (i.e. no instruction-following fine tuning) using a domain-specific dataset, so that the model becomes more knowledgable of that domain and a better starting point for instruction-based fine tuning.However, the fine tuned model seems to just overfit to the training dataset, almost always producing responses that have similar structure and content like the documents in the training set. Instead, the ideal outcome would be that the model learns the domain-related knowledge, not the structure of the documents, and does not lose too much of the original knowledge.My questions are the following:Has anyone had any experience with this?Is it even feasible to achieve the desired goal, without resorting to a pre-training from scratch?What can be done from a training perspective? E.g. does it make sense to gradually unfreeze weights as it used to be done with DCNNs?

2023-06-19T15:41:02Z

Pre-trained DeBERTa - Weak MLM performance any hints?

https://discuss.huggingface.co/t/pre-trained-deberta-weak-mlm-performance-any-hints/43878

Hi,I wanted to use DeBERTa.Somehow the preview of its unmasking abilities seems very bad.image776×490 13.8 KBI looked at the source code and cannot see the addition of the absolute positions.Can someone explain me why the model performs so bad at MLM preview.Maybe I overlooked the addition of the absolute positions in the source code.An explanation of the implementation would be really helpful aswell!Thank you!Stephan

2023-06-20T07:34:23Z

AI model for Bitcoin blockchain data analysis

https://discuss.huggingface.co/t/ai-model-for-bitcoin-blockchain-data-analysis/47527

I propose the use of recurrent neural networks (RNN) and its variants LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit) for data analysis of the Bitcoin blockchain. These models are key to extracting valuable information and patterns in the cryptocurrency ecosystem.RNNs, along with their LSTM and GRU variants, are essential for handling large volumes of data from the Bitcoin blockchain and uncovering meaningful trends. These models make it possible to identify suspicious transactions, price fluctuations and other relevant events in the world of cryptocurrencies. Using RNN, LSTM and GRU will open up new opportunities to make informed decisions in this field.

2023-07-20T21:33:56Z