base_model:
- Qwen/Qwen2.5-32B-Instruct
datasets:
- codeparrot/apps
- BAAI/TACO
- AI-MO/NuminaMath-CoT
language:
- en
library_name: transformers
license: apache-2.0
pipeline_tag: text-generation
Model Details
Model Description
This is a 32B reasoning model trained from Qwen2.5-32B-Instruct with 17K data. The performance is on par with o1-preview model on both math and coding. Please see our blog post for more details.
- Developed by: NovaSky Team from Sky Computing Lab at UC Berkeley.
Training Details
Training Data
17K verified correct responses from Qwen/QwQ-32B-Preview on coding, math. In addition, we add the science portion from the Still-2 paper.
Training Procedure
We perform supervised fine tuning on the data, with a batch size of 96.
Speeds
We use Llama-Factory for training. On 8 H100, the training takes 19 hours with DeepSpeed Zero-3 Offload.
Evaluation
| Sky-T1-32B-Preview | Qwen-2.5-32B-Instruct | QwQ | o1-preview | |
|---|---|---|---|---|
| Math500 | 82.4 | 76.2 | 85.4 | 81.4 |
| AIME2024 | 43.3 | 16.7 | 50.0 | 40.0 |
| LiveCodeBench-Easy | 86.3 | 84.6 | 90.7 | 92.9 |
| LiveCodeBench-Medium | 56.8 | 40.8 | 56.3 | 54.9 |
| LiveCodeBench-Hard | 17.9 | 9.8 | 17.1 | 16.3 |
| GPQA-Diamond | 56.8 | 45.5 | 52.5 | 75.2 |
Acknowledgement
We would like to thanks the compute resources from Lambda Lab and AnyScale. We would like to thanks the academic feedback and support from the Still-2 Team, and Junyang Lin from the Qwen Team.
Citation
Please considering citing our paper if you found it useful for your research. Thank you!
@misc{zhu2025llms,
author = {Wenxuan Zhu and Xiangru Tang and Ziyang Ma and Hongbo Zhang and Tianqi Chen},
title = {LLMs Can Easily Learn to Reason from Demonstrations Structure, not content, is what matters!},
year = {2025},
eprint={2502.07374},
archivePrefix={arXiv},
primaryClass={cs.CL},
url = {https://arxiv.org/abs/2502.07374}
}