| <div align="center"> | |
| <h1> | |
| JiuZhou: Open Foundation Language Models for Geoscience | |
| </h1> | |
| </div> | |
| ## 🎉 News | |
| - [2024-12-31] **Article [JiuZhou: Open Foundation Language Models and Effective Pre-training Framework for Geoscience](https://www.tandfonline.com/doi/full/10.1080/17538947.2025.2449708) has been accepted for publication in the *International Journal of Digital Earth***. [Code and Data](https://github.com/THU-ESIS/JiuZhou). | |
| - [2024-10-11] WeChat article: [PreparedLLM: Effective Pre-pretraining Framework for Domain-specific Large Language Models](https://mp.weixin.qq.com/s/ugJQ9tbp6Y87xA3TOWteqw). | |
| - [2024-09-06] Released [ClimateChat](https://huggingface.co/itpossible/ClimateChat) instruct model. | |
| - [2024-08-31] **Article [PreparedLLM: Effective Pre-pretraining Framework for Domain-specific Large Language Models](https://www.tandfonline.com/doi/full/10.1080/20964471.2024.2396159) has been accepted for publication in the *Big Earth Data* journal**. | |
| - [2024-08-31] Released [Chinese-Mistral-7B-Instruct-v0.2](https://huggingface.co/itpossible/Chinese-Mistral-7B-Instruct-v0.2) instruct model. Significant improvements in language understanding and multi-turn dialogue capabilities. | |
| - [2024-06-30] Released [JiuZhou-Instruct-v0.2](https://huggingface.co/itpossible/JiuZhou-Instruct-v0.2) instruct model. Significant improvements in language understanding and multi-turn dialogue capabilities. | |
| - [2024-05-15] WeChat Article: [Chinese Vocabulary Expansion Incremental Pretraining for Large Language Models: Chinese-Mistral Released](https://mp.weixin.qq.com/s/PMQmRCZMWosWMfgKRBjLlQ). | |
| - [2024-04-04] Released [Chinese-Mistral-7B-Instruct-v0.1](https://huggingface.co/itpossible/Chinese-Mistral-7B-Instruct-v0.1) instruct model. | |
| - [2024-03-31] Released [Chinese-Mistral-7B-v0.1](https://huggingface.co/itpossible/Chinese-Mistral-7B) base model. | |
| - [2024-03-15] Released the base version [JiuZhou-base](https://huggingface.co/itpossible/JiuZhou-base), instruct version [JiuZhou-instruct-v0.1](https://huggingface.co/itpossible/JiuZhou-Instruct-v0.1), and [intermediate checkpoints](https://huggingface.co/itpossible). | |
| ## Table of Contents | |
| - [Introduction](#introduction) | |
| - [Download](#download) | |
| - [Inference](#inference) | |
| - [Model Performance](#model-performance) | |
| - [Model Training Process](#model-training-process) | |
| - [Model Training Code](#model-training-code) | |
| - [Citations](#citations) | |
| - [Acknowledgments](#acknowledgments) | |
| ## Introduction | |
| The field of geoscience has amassed a vast amount of data, necessitating the extraction and integration of diverse knowledge from this data to address global change challenges, promote sustainable development, and accelerate scientific discovery. Foundation language models initially learn and integrate knowledge autonomously through self-supervised pre-training on extensive text data. Subsequently, they acquire the capability to solve geoscience problems through instruction tuning. However, when the foundational language models lack sufficient geoscience expertise, instruction tuning with relevant data can lead to the generation of content that is inconsistent with established facts. To improve the model's accuracy and practicality, a robust geoscience foundational language model is urgently needed.<br> | |
| This study uses [Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1) as the base model and continues pretraining on a large geoscience corpus. It also incorporates the [domain-specific large language model *pre*-pretraining framework (PreparedLLM)](https://www.tandfonline.com/doi/full/10.1080/20964471.2024.2396159) and the "two-stage pre-adaptation pre-training" algorithm to build the geoscience large language model, JiuZhou. | |
| ## Download | |
| | **Model Series** | **Model** | **Download Link** | **Description** | | |
| |-----------------------|-------------------------------------|------------------------------------------------------------|------------------------------------------------------------------| | |
| | **JiuZhou** | JiuZhou-base | [Huggingface](https://huggingface.co/itpossible/JiuZhou-base) | Base model (Rich in geoscience knowledge) | | |
| | **JiuZhou** | JiuZhou-Instruct-v0.1 | [Huggingface](https://huggingface.co/itpossible/Chinese-Mistral-7B-Instruct-v0.1) | Instruct model (Instruction alignment caused a loss of some geoscience knowledge, but it has instruction-following ability) <br> LoRA fine-tuned on Alpaca_GPT4 in both Chinese and English and GeoSignal | | |
| | **JiuZhou** | JiuZhou-Instruct-v0.2 | [HuggingFace](https://huggingface.co/itpossible/Chinese-Mistral-7B-Instruct-v0.2)<br>[Wisemodel](https://wisemodel.cn/models/itpossible/Chinese-Mistral-7B-Instruct-v0.2) | Instruct model (Instruction alignment caused a loss of some geoscience knowledge, but it has instruction-following ability) <br> Fine-tuned with high-quality general instruction data | | |
| | **ClimateChat** | ClimateChat | [HuggingFace](https://huggingface.co/itpossible/ClimateChat)<br>[Wisemodel](https://wisemodel.cn/models/itpossible/ClimateChat) | Instruct model <br> Fine-tuned on JiuZhou-base for instruction following | | |
| | **Chinese-Mistral** | Chinese-Mistral-7B | [HuggingFace](https://huggingface.co/itpossible/Chinese-Mistral-7B-v0.1)<br>[Wisemodel](https://wisemodel.cn/models/itpossible/Chinese-Mistral-7B-v0.1)<br>[ModelScope](https://www.modelscope.cn/models/itpossible/Chinese-Mistral-7B-v0.1) | Base model | | |
| | **Chinese-Mistral** | Chinese-Mistral-7B-Instruct-v0.1 | [HuggingFace](https://huggingface.co/itpossible/Chinese-Mistral-7B-Instruct-v0.1)<br>[Wisemodel](https://wisemodel.cn/models/itpossible/Chinese-Mistral-7B-Instruct-v0.1)<br>[ModelScope](https://www.modelscope.cn/models/itpossible/Chinese-Mistral-7B-Instruct-v0.1) | Instruct model <br> LoRA fine-tuned with Alpaca_GPT4 in both Chinese and English | | |
| | **Chinese-Mistral** | Chinese-Mistral-7B-Instruct-v0.2 | [HuggingFace](https://huggingface.co/itpossible/Chinese-Mistral-7B-Instruct-v0.2)<br>[Wisemodel](https://wisemodel.cn/models/itpossible/Chinese-Mistral-7B-Instruct-v0.2) | Instruct model <br> LoRA fine-tuned with a million high-quality instructions | | |
| | **PreparedLLM** | Prepared-Llama | [Huggingface](https://huggingface.co/itpossible/Prepared-Llama)<br>[Wisemodel](https://wisemodel.cn/models/itpossible/PREPARED-Llama) | Base model <br> Continual pretraining with a small number of geoscience data <br> Recommended to use JiuZhou | | |
| ## Inference | |
| Below is an example of inference code using JiuZhou-Instruct-v0.2. | |
| ```python | |
| import torch | |
| from transformers import AutoTokenizer, AutoModelForCausalLM | |
| device = torch.device("cuda:0") if torch.cuda.is_available() else torch.device("cpu") | |
| model_path = "itpossible/JiuZhou-Instruct-v0.2" | |
| tokenizer = AutoTokenizer.from_pretrained(model_path) | |
| model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=torch.bfloat16, device_map=device) | |
| text = "What is geoscience?" | |
| messages = [{"role": "user", "content": text}] | |
| inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(device) | |
| outputs_id = model.generate(inputs, max_new_tokens=600, do_sample=True) | |
| outputs = tokenizer.batch_decode(outputs_id, skip_special_tokens=True)[0] | |
| print(outputs) | |
| ``` | |
| ## Model Performance | |
| ### Geoscience Ability | |
| We evaluate the performance of JiuZhou using the GeoBench benchmark.<br> | |
| JiuZhou outperforms GPT-3.5 in objective tasks: | |
| <p align="center"> | |
| <br> | |
| <img src="image/objective_score.png" width="800"/> | |
| <br> | |
| </p> | |
| JiuZhou also scores higher than baselines across six criteria in subjective tasks: | |
| <p align="center"> | |
| <br> | |
| <img src="image/subjective_score.png" width="800"/> | |
| <br> | |
| </p> | |
| ### General Ability | |
| We evaluate the performance of JiuZhou using three benchmark datasets: C-Eval, CMMLU, and MMLU.<br> | |
| Compared to other variants of Llama and Mistral models, JiuZhou shows outstanding performance: | |
| <p align="center"> | |
| <br> | |
| <img src="image/general_score.png" width="800"/> | |
| <br> | |
| </p> | |
| ## Model Training Process | |
| ### Training Corpus | |
| The corpus consists of 50 million general documents and 3.4 million geoscience-related documents. | |
| <p align="center"> | |
| <br> | |
| <img src="image/JiuZhou-Corpus.png" width="800"/> | |
| <br> | |
| </p> | |
| ### Training Framework | |
| We use the JiuZhou-Framework proposed in this study. | |
| <p align="center"> | |
| <br> | |
| <img src="image/JiuZhou-Framework.png" width="800"/> | |
| <br> | |
| </p> | |
| ### Two-stage Pre-adaptation Pre-training (TSPT) | |
| TSPT improves the efficiency of using limited geoscience data and overcomes some of the technical bottlenecks in continual pretraining for LLMs.<br> | |
| The difference between TSPT and single-stage training algorithms: | |
| <p align="center"> | |
| <br> | |
| <img src="image/TSPT.png" width="800"/> | |
| <br> | |
| </p> | |
| Comparison of TSPT and one-stage pre-training algorithm performance: | |
| <p align="center"> | |
| <br> | |
| <img src="image/TSPT_score.png" width="800"/> | |
| <br> | |
| </p> | |
| ## Model Training Code | |
| We use [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) to fine-tune JiuZhou. | |
| ### Project Deployment | |
| ```bash | |
| git clone https://github.com/THU-ESIS/JiuZhou.git | |
| cd JiuZhou | |
| pip install -e ".[torch,metrics]" | |
| ``` | |
| ### Model Training | |
| Pre-training： | |
| ```bash | |
| llamafactory-cli train examples/train_lora/JiuZhou_pretrain_sft.yaml | |
| ``` | |
| Instruction-tuning： | |
| ```bash | |
| llamafactory-cli train examples/train_lora/JiuZhou_lora_sft.yaml | |
| ``` | |
| Chat with the fine-tuned JiuZhou:： | |
| ```bash | |
| llamafactory-cli chat examples/inference/JiuZhou_lora_sft.yaml | |
| ``` | |
| Merge the instruction-tuned LoRA weights with the original JiuZhou weights: | |
| ```bash | |
| llamafactory-cli export examples/merge_lora/JiuZhou_lora_sft.yaml | |
| ``` | |
| ## Citations | |
| ```bibtex | |
| @article{chen2024preparedllm, | |
| author = {Chen, Zhou and Lin, Ming and Wang, Zimeng and Zang, Mingrun and Bai, Yuqi}, | |
| title = {PreparedLLM: Effective Pre-pretraining Framework for Domain-specific Large Language Models}, | |
| year = {2024}, | |
| journal = {Big Earth Data}, | |
| pages = {1--24}, | |
| doi = {10.1080/20964471.2024.2396159}, | |
| url = {https://doi.org/10.1080/20964471.2024.2396159} | |
| } | |
| ``` | |
| ## Acknowledgments | |
| - [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) | |
| - [OpenCompass](https://github.com/open-compass/opencompass) | |
| - [K2](https://github.com/davendw49/k2) | |
| - [GeoGalactica](https://github.com/geobrain-ai/geogalactica) | |
| - [BB-GeoGPT](https://github.com/AGI-GIS/BB-GeoGPT) | |