DeepScaleR-7B-GRPO-8k
DeepScaleR-7B-GRPO-8k Overview
DeepScaleR-7B-GRPO-8k is a language model fine-tuned from DeepSeek-R1-Distilled-Qwen-7B using distributed reinforcement learning (RL) to scale up to long context lengths.
Note that we follow the training process of
original DeepscaleR and further extend the model size up to 7B model.
This model is a 8K Context version, corrspoding to the stage-1 training of DeepScaleR-1.5B-Preview. Our full training process of this 8K-context stage consist of 400 training steps before the learning curves converge. This is a snapshot of step 240 where the testing result of AIME2023/2024 is relativly better.
Data
Same to DeepScaleR-1.5B-Preview, the training dataset consists of approximately 40,000 unique problem-answer pairs compiled from:
- AIME problems (1984-2023)
- AMC problems (prior to 2023)
- Omni-MATH dataset
- Still dataset
Training Recipe
We employ Deepseek's Group Relative Policy Optimization (GRPO), a simplified RL algorithm that extends PPO by:
- Normalizing advantage function over all samples generated from the same prompt.
- Applying KL divergence regularization on top of PPO's surrogate loss to prevent significant policy drift.
Reward Function: Our reward function is simple but effective:
- 1 for correct answers passing LaTeX/Sympy checks
- 0 for incorrect or improperly formatted answers
- Note: No partial rewards (such as PRMs) or intermediate feedback.
Iterative Context Lengthening:
- Initial 8K Context (0-400 steps):
- 38% -> ~50% Pass@1 on AIME 2023
- Trained on 4 H20 GPUs, BS= (Prompts) * (Samples/Prompt) = 128 * 6 = 768
We will share our foundings, recipe and wandb logs in our upcoming blog post.
Evaluation
We report Pass@1 accuracy averaged over 16 samples for each problem.
| Model | AIME 2024 | MATH 500 | AMC 2023 | Minerva Math | OlympiadBench | Avg. |
|---|
| Qwen-2.5-7B-Instruct | 13.3 | 79.8 | 50.6 | 34.6 | 40.7 | 43.8 |
| rStar-Math-7B | 26.7 | 78.4 | 47.5 | - | 47.1 | - |
| Eurus-2-7B-PRIME | 26.7 | 79.2 | 57.8 | 38.6 | 42.1 | 48.9 |
| Qwen2.5-7B-SimpleRL | 26.7 | 82.4 | 62.5 | 39.7 | 43.3 | 50.9 |
| DeepSeek-R1-Distill-Qwen-1.5B | 28.8 | 82.8 | 62.9 | 26.5 | 43.3 | 48.9 |
| Still-1.5B | 32.5 | 84.4 | 66.7 | 29.0 | 45.4 | 51.6 |
| DeepScaleR-1.5B-Preview | 43.1 | 87.8 | 73.6 | 30.2 | 50.0 | 57.0 |
| O1-Preview | 40.0 | 81.4 | - | - | - | - |
Serving DeepScaleR
Our model can be served using popular high-performance inference systems:
- vLLM
- Hugging Face Text Generation Inference (TGI)
- SGLang
- TensorRT-LLM
All these systems support the OpenAI Chat Completions API format.
License
This project is released under the MIT License, reflecting our commitment to open and accessible AI development.
We believe in democratizing AI technology by making our work freely available for anyone to use, modify, and build upon.
This permissive license ensures that researchers, developers, and enthusiasts worldwide can leverage and extend our work without restrictions, fostering innovation and collaboration in the AI community.
Acknowledgement
- Our training experiments are powered by our heavily modified fork of Verl, an open-source RLHF library.
- Our model is trained on top of
DeepSeek-R1-Distill-Qwen-1.5B.
- Our work is done as part of Berkeley Sky Computing Lab and Berkeley AI Research.
Citation
1@misc{deepscaler2025,
2 title={DeepScaleR: Surpassing O1-Preview with a 1.5B Model by Scaling RL},
3 author={Michael Luo and Sijun Tan and Justin Wong and Xiaoxiang Shi and William Y. Tang and Manan Roongta and Colin Cai and Jeffrey Luo and Tianjun Zhang and Li Erran Li and Raluca Ada Popa and Ion Stoica},
4 year={2025},
5 howpublished={\url{https://pretty-radio-b75.notion.site/DeepScaleR-Surpassing-O1-Preview-with-a-1-5B-Model-by-Scaling-RL-19681902c1468005bed8ca303013a4e2}},
6 note={Notion Blog}
7 year={2025}
8}
