17 lines
1.2 KiB
Markdown
17 lines
1.2 KiB
Markdown
|
---
|
||
|
|
license: mit
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
library_name: transformers
|
||
|
|
tags:
|
||
|
|
- qwen3
|
||
|
|
- reasoning
|
||
|
|
- llm
|
||
|
|
---
|
||
|
|
|
||
|
|
### Co-rewarding-I: Qwen3-8B-Base trained on DAPO-14k
|
||
|
|
|
||
|
|
This model is the Qwen3-8B-Base, trained by Co-rewarding-I using the DAPO-14k training set. It was presented in the paper [Co-rewarding: Stable Self-supervised RL for Eliciting Reasoning in Large Language Models](https://huggingface.co/papers/2508.00410).
|
||
|
|
|
||
|
|
Co-rewarding is a novel self-supervised reinforcement learning (RL) framework designed to improve the training stability of large language models (LLMs) for reasoning tasks. This particular model utilizes **Co-rewarding-I**, a data-side instantiation that derives reward signals from contrastive agreement across semantically analogous questions. This approach aims to mitigate the training collapse and reward hacking issues often encountered in single-view self-rewarding methods, thereby enhancing the LLM's reasoning abilities for complex challenges like mathematical reasoning.
|
||
|
|
|
||
|
|
For more details on the Co-rewarding framework, access to the code, and other trained checkpoints, please refer to the official GitHub repository: [https://github.com/tmlr-group/Co-rewarding](https://github.com/tmlr-group/Co-rewarding).
|