Spaces:
Running
Upload r1_responses_cleaned.csv
Browse files---
license: apache-2.0
---
# MasterControl AIML Team 🚀
## Overview
The **MasterControl AIML team** supports the **Hugging Face** initiative of re-creating **DeepSeek R1 training**, recognizing it as one of the most **impactful open-source projects** today.
We aim to contribute to **reasoning datasets**, specifically those where:
- A **real-world problem** involves **generating complex structured output**
- It is accompanied by **step-by-step reasoning and unstructured input**
## Challenges in Integrating Generative AI into Systems of Record (SoR)
Integrating **Generative AI** into **Systems of Record (SoR)** for **health, life sciences, and manufacturing quality** is challenging due to:
- These systems rely on **strictly structured data formats** (e.g., **JSON, XML, templates**).
- **LLM outputs** are **unstructured** and do not conform to **regular expressions** or **context-free grammars**.
## Techniques for Structured Output Generation
To enforce structured output generation, we explore:
- **Strict schema prompting**
- **Post-processing and output validation**
- **Reformulating text generation** into transitions between **finite-state machine states**
## DeepSeek R1 Approach
A key challenge is **fitting hybrid structured and unstructured historical manufacturing production records** to **master templates**.
We aim to leverage the **DeepSeek R1 model**, which uses:
- **Pure reinforcement learning** to train a **base language model**
- Learning to **reason without human supervision**
## Model Used
- We used deepseek's distilled 7b to created reasoning responses to go from unstructured to structured
## Purpose of Reasoning Responses
- The reasoning responses are created in such a way, that if the model is presented with unstructured text and a schema of rules, it needs to convert it into a structured schema. These responses can be used for any unstructured to structured creation.
## Next Step: Reasoning Data
Our **first step** is curating and contributing to **reasoning datasets** that facilitate structured output generation.
---
- .gitattributes +1 -0
- r1_responses_cleaned.csv +3 -0
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
r1_responses_cleaned.csv filter=lfs diff=lfs merge=lfs -text
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:dbf828ffcde1b06da0e82a9178b49e0a33df9c11cdcd992787abe99aa1887106
|
3 |
+
size 161177437
|