33 lines
1.7 KiB
Markdown
33 lines
1.7 KiB
Markdown
---
|
|
library_name: transformers
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# Composition-RL-8B
|
|
|
|
[**Paper**](https://huggingface.co/papers/2602.12036) | [**Code**](https://github.com/XinXU-USTC/Composition-RL)
|
|
|
|
**Composition-RL-8B** is a large language model fine-tuned for enhanced reasoning using the Composition-RL framework. It was initialized from the **Qwen3-8B-Base** architecture and trained on the **MATH-Composition-199K** dataset.
|
|
|
|
## Description
|
|
|
|
Composition-RL is a data-efficient Reinforcement Learning with Verifiable Rewards (RLVR) approach presented in the paper [Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models](https://huggingface.co/papers/2602.12036).
|
|
|
|
The method addresses the issue of "too-easy" prompts (where pass rates reach 1) that occur as training progresses, which reduces the effective training signal. Composition-RL automatically composes multiple verifiable problems into a single, more challenging compositional prompt, ensuring the model continues to receive informative rewards throughout the reinforcement learning process.
|
|
|
|
- **Base Model:** Qwen3-8B-Base
|
|
- **Training Method:** Reinforcement Learning with Verifiable Rewards (RLVR)
|
|
- **Training Dataset:** [MATH-Composition-199K](https://huggingface.co/datasets/xx18/MATH-Composition-199K)
|
|
|
|
## Citation
|
|
|
|
If you find this work helpful for your research, please consider citing:
|
|
|
|
```bibtex
|
|
@article{xu2026composition-rl,
|
|
title={Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models},
|
|
author={Xu, Xin and Bai, Clive and Yang, Kai and Chen, Tianhao and Chen, Yangkun and Liu, Weijie and Chen, Hao and Wang, Yang and Yang, Saiyong and Yang, Can},
|
|
journal={arXiv preprint arXiv:2602.12036},
|
|
year={2026}
|
|
}
|
|
``` |