TMLR-Group-HF/GT-Llama-3.2-3B-Instruct-MATH
TMLR-Group-HF/GT-Llama-3.2-3B-Instruct-MATH is a 3.2 billion parameter instruction-tuned Llama-3.2 model developed by TMLR-Group-HF, trained using the GRPO Ground Truth method with the MATH training set. This model is specifically optimized for eliciting reasoning in large language models through the novel Co-rewarding self-supervised reinforcement learning framework. It is designed to address training stability issues in self-rewarding methods, making it suitable for complex mathematical and reasoning tasks. The model features a 32768 token context length, enhancing its ability to process longer problem descriptions.
Loading preview...
Model Overview
TMLR-Group-HF/GT-Llama-3.2-3B-Instruct-MATH is a 3.2 billion parameter instruction-tuned model based on the Llama-3.2 architecture, developed by TMLR-Group-HF. It was trained using the GRPO Ground Truth method, specifically leveraging the MATH training set. This model is a key checkpoint from research on the Co-rewarding framework, detailed in the paper "Co-rewarding: Stable Self-supervised RL for Eliciting Reasoning in Large Language Models" (arXiv:2508.00410).
Key Capabilities
- Enhanced Reasoning: Utilizes the novel Co-rewarding self-supervised reinforcement learning (RL) framework to improve reasoning capabilities.
- Training Stability: Addresses the training collapse issue common in single-view self-rewarding methods by seeking complementary supervision from multiple views.
- Mathematical Proficiency: Specifically trained on the MATH dataset, indicating a strong focus on mathematical problem-solving.
- Instruction Following: Instruction-tuned to better understand and respond to user prompts.
Good For
- Mathematical Problem Solving: Ideal for applications requiring robust mathematical reasoning.
- Research in RL and Reasoning: Useful for researchers exploring stable self-supervised reinforcement learning techniques for LLMs.
- Complex Reasoning Tasks: Suitable for scenarios where eliciting detailed reasoning steps is crucial.