README.md

---
license: llama3.2
datasets:
- OctoThinker/MegaMath-Web-Pro-Max
- LLM360/MegaMath
language:
- en
base_model:
- meta-llama/Llama-3.2-3B
pipeline_tag: text-generation
---

# [OctoThinker: Mid-training Incentivizes Reinforcement Learning Scaling](https://arxiv.org/abs/2506.20512)



## OctoThinker-3B-Hybrid-Zero


The OctoThinker family is built on carefully studied mid-training insights, starting from the Llama-3 family, to create a reinforcement learning–friendly base language model.

OctoThinker-3B-Hybrid-Zero is trained using the R1-Zero-style reinforcement learning technique, starting from OctoThinker-3B-Hybrid-Base without any supervised fine-tuning (SFT).


### Training Recipe for OctoThinker-3B-Hybrid-Base

<div style="display: flex; justify-content: left; gap: 20px;">
<img src="https://cdn-uploads.huggingface.co/production/uploads/62cbeb2d72dfd24b86bdf977/XSSllxdLr3dcw250dFm7e.png" alt="Data Pipeline" style="width:90%;">
</div>




### Evaluation Results of OctoThinker-3B-Base Series

Note that we adopt the few-shot prompting evaluation for these base language models.


<div style="display: flex; justify-content: left; gap: 20px;">
<img src="https://cdn-uploads.huggingface.co/production/uploads/62cbeb2d72dfd24b86bdf977/UCZ9MahRYqLY0iKjiWMqS.png" alt="Data Pipeline" style="width:80%;">

</div>



### RL Training Dynamics of OctoThinker-3B-Zero Series

<div style="display: flex; justify-content: left; gap: 20px;">
<img src="https://cdn-uploads.huggingface.co/production/uploads/62cbeb2d72dfd24b86bdf977/e21Eg8jj_ITxC4YcIJUmx.png" alt="Data Pipeline" style="width:80%;">
</div>



### More about OctoThinker


<div style="display: flex; justify-content: left; gap: 20px;">
<img src="https://cdn-uploads.huggingface.co/production/uploads/62cbeb2d72dfd24b86bdf977/bn85CEB_DW6azJ7KJp11Q.png" alt="Data Pipeline" style="width:100%;">
</div>


## Citation

Check out our [paper](https://arxiv.org/abs/2506.20512) for more details. If you use our models, datasets or find our work useful, please cite

```
@article{wang2025octothinker,
  title={OctoThinker: Mid-training Incentivizes Reinforcement Learning Scaling},
  author={Wang, Zengzhi and Zhou, Fan and Li, Xuefeng and Liu, Pengfei},
  year={2025},
  journal={arXiv preprint arXiv:2506.20512},
  note={Preprint}
}
```