|
|
--- |
|
|
license: mit |
|
|
base_model: |
|
|
- Qwen/Qwen3-4B-Instruct-2507 |
|
|
--- |
|
|
|
|
|
## LiteCoder-4b-Terminal-preview |
|
|
|
|
|
**LiteCoder-4b-Terminal-preview** is part of our series of models specialized in terminal-based interactions and stems from our recent efforts to develop capable small and medium-sized code agent models. The model is fine-tuned from ` |
|
|
Qwen3-4B-Instruct-2507` on the [LiteCoder-SFT-Terminal-preview](https://huggingface.co/datasets/Lite-Coder/LiteCoder-SFT-Terminal-preview) dataset. |
|
|
|
|
|
**Notably, this model achieves competitive results using fewer than 1,000 training samples.** By relying entirely on a fully synthetic pipeline—without converting any existing datasets—we were able to secure significant gains on the challenging Terminal Bench, matching the performance of leading open-source models with extreme data efficiency. |
|
|
|
|
|
## Released Artifacts |
|
|
|
|
|
| 2025/12/17 | | | |
|
|
| --- | --- | --- | |
|
|
| LiteCoder-4b-Terminal-preview | Model | https://huggingface.co/Lite-Coder/LiteCoder-4b-Terminal-preview | |
|
|
| LiteCoder-SFT-Terminal-preview | Dataset | https://huggingface.co/datasets/Lite-Coder/LiteCoder-SFT-Terminal-preview | |
|
|
|
|
|
## Results |
|
|
|
|
|
Our models achieve competitive results on **Terminal Bench**, significantly outperforming general-purpose models of similar (and even larger) sizes. |
|
|
|
|
|
**Terminal Bench 1.0 Performance** |
|
|
|
|
|
| **Model** | **Agent** | **Results** | |
|
|
| --- | --- | --- | |
|
|
| **LiteCoder-30a3b-Terminal-preview** | Terminus 2 | **18.75%** | |
|
|
| Qwen3-30B-A3B-Nex-N1 | Terminus 2 | 18.75% | |
|
|
| **LiteCoder-4b-Terminal-preview** | Terminus 2 | **13.75%** | |
|
|
| Qwen3-30B-A3B-Instruct | Terminus 2 | 12.5% | |
|
|
| Qwen3-4B-Instruct | Terminus 2 | 5.0% | |
|
|
|
|
|
**Terminal Bench 2.0 Performance** |
|
|
|
|
|
| **Model** | **Agent** | **Results** | |
|
|
| --- | --- | --- | |
|
|
| **LiteCoder-30a3b-Terminal-preview** | Terminus 2 | **5.6%** | |
|
|
| **LiteCoder-4b-Terminal-preview** | Terminus 2 | **3.3%** | |
|
|
| Qwen3-32B | Terminus 2 | 1.9% | |
|
|
| InternLM3-8B-Nex-N1 | Terminus 2 | 0% | |
|
|
| Qwen3-8B | Terminus 2 | 0% | |
|
|
|
|
|
## Citation |
|
|
|
|
|
```latex |
|
|
@misc{LiteCoder Team, |
|
|
title={LiteCoder: Advancing Small and Medium-sized Code Agents}, |
|
|
author={Xiaoxuan Peng and Xinyu Lu and Kaiqi Zhang and Taosong Fang and Boxi Cao and Yaojie Lu}, |
|
|
year={2025}, |
|
|
} |
|
|
``` |
|
|
|
|
|
## Future Directions |
|
|
|
|
|
- **Scaling Environments:** Expanding the diversity of Docker environments and teacher models to improve generalization. |
|
|
- **Agentic RL:** Implementing Reinforcement Learning specifically for multi-turn agentic workflows. |
|
|
|
|
|
## Team & Contributions |
|
|
|
|
|
- **Xiaoxuan Peng:** Main Contributor |
|
|
- [Xinyu Lu](https://scholar.google.com/citations?user=_OsLG8EAAAAJ&hl=zh-CN)**:** Project Lead |
|
|
- **Kaiqi Zhang:** Contributor |
|
|
- **Taosong Fang**: Contributor |
|
|
- **Boxi Cao:** Contributor |
|
|
- **Yaojie Lu:** Contributor |
|
|
|
|
|
## Acknowledgements |
|
|
|
|
|
LiteCoder builds upon multiple open-source projects, including [Harbor](https://github.com/laude-institute/harbor). The models are trained using [AutoAlign](https://github.com/icip-cas/AutoAlign). |
|
|
|
|
|
## Join Us |
|
|
|
|
|
Join the discussion on our [Discord](https://discord.gg/EX9qZe8B). |
|
|
|
