---
license: apache-2.0
base_model: Qwen/Qwen3.5-4B
library_name: transformers
pipeline_tag: text-generation
language:
- en
- zh
tags:
- learning-environment-generation
- slide-generation
- interactive-html
- education
---

# CogEvol-4B

**CogEvol-4B** is an open post-trained model for **Learning Environment Generation (LEG)**: given a natural-language course brief, it generates complete, usable learning artifacts in a single pass — a structured-JSON slide page, or a self-contained interactive HTML page that runs directly in the browser (simulations, visualizations, interactive exercises, and more).

It is the small open member of the CogEvol family, trained with a three-stage recipe (mix SFT → Slide RL → interactive-HTML RL) on 53,687 verified samples, using a hybrid rule + VLM reward system hardened against reward hacking — interactivity is *measured* by automated probes, never just judged. Paper: *CogEvol: Towards Efficient and Reliable Learning Environment Generation* (link below).

## Quick start

- **Run it on your laptop**: the [GitHub repository](https://github.com/CogEvol/CogEvol-4B) is the complete deployment guide — a step-by-step llama.cpp setup, a `serve.sh` with the validated flags, the [OpenMAIC](https://github.com/THU-MAIC/OpenMAIC) app integration (runs fully offline), evaluation scripts, and a troubleshooting table. A quantized **Q4_K_M GGUF (~2.4 GB)** is available at [CogEvol/CogEvol-4B-Q4_K_M-GGUF](https://huggingface.co/CogEvol/CogEvol-4B-Q4_K_M-GGUF).
- Serving this BF16 checkpoint: sglang / vLLM / llama.cpp all work. Two rules from the serving guide: disable thinking mode at inference, and when generating slides use the system prompt template shipped in the GitHub repository.

## Links

- GitHub: <https://github.com/CogEvol/CogEvol-4B>
- Tech report: [arXiv:2608.30968](https://arxiv.org/abs/2608.30968)
- Contact: <contact@cogevol.com>

## Citation

If you find CogEvol useful, please cite:

```bibtex
@misc{tu2026cogevolefficientreliablelearning,
      title={CogEvol: Towards Efficient and Reliable Learning Environment Generation}, 
      author={Shangqing Tu and Daniel Zhang-Li and Yucheng Wang and Shiyu Gan and Yanpeng Wang and Huiqiang Rong and Mofei Chen and Shen Yang and Yini Chen and Yinuo Duan and Binglin Liu and Ye He and Danqi Zheng and Zhanxin Hao and Yuxuan Wu and Mengting Tao and Yuqiu Liu and Jifan Yu and Juanzi Li and Bin Xu and Lei Hou and Huiqin Liu and Yu Zhang},
      year={2026},
      eprint={2608.30968},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2608.30968}
}
```