You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

LiteCoder-SFT-Terminal

Paper | Code | Blog Post

LiteCoder-SFT-Terminal is a dataset of 11,255 agent trajectories in terminal environments, introduced in the paper LiteCoder-Terminal: Scaling Long-Horizon Terminal Environments for Learning Language Agents.

Fine-tuned on this data, the LiteCoder-Terminal-30b-a3b-sft model achieves 31.5% Pass@1 on Terminal Bench Pro, while the LiteCoder-Terminal-4b-sft model shows distinct gains over its baseline.

Released Artifacts

Date Type Link
2026/04/13 Model LiteCoder-Terminal-30b-a3b-sft
2026/04/13 Model LiteCoder-Terminal-4b-sft
2026/04/13 Dataset LiteCoder-Terminal-SFT
2026/04/13 Dataset LiteCoder-Terminal-World-Model-SFT
2026/04/13 Dataset LiteCoder-Terminal-RL-preview
2026/04/13 Code icip-cas/LiteCoder

What’s New

  • Taxonomy expansion: Introducing new task categories to cover a broader range of terminal interactions.
  • Data scale expansion: Scaling from a sub-1k setup to 11,255 total trajectories.
  • Scaffold expansion: Moving from Terminus-only training to a multi-scaffold setting including OpenHands and Claude Code.
  • Improved results: Higher performance on Terminal Bench 1.0 / 2.0 / Pro.

Taxonomy Expansion

We expanded the task taxonomy to cover a broader range of real-world terminal interactions, introducing 3 new categories: coding, scientific/numerical computing, and games. This ensures the model is exposed to highly diverse scenarios, ranging from rigorous developer workflows to the dynamic, state-driven environments typically found in terminal games.

Scaffolding Extension

Given the vast array of agent scaffolds, a robust model must be able to adapt to diverse setups. To improve cross-scaffold generalization, we expanded our trajectory collection beyond the initial Terminus-only setup. The updated training pipeline now integrates trajectories from popular frameworks like Claude Code and OpenHands.

Dataset Statistics

The LiteCoder-SFT dataset comprises 11,255 trajectories spanning 10 task categories, with an average of 27.4 turns per trajectory. The dataset also incorporates trajectories from three distinct agent scaffolds: Terminus‑2 (86.6%), OpenHands (7.1%), and Claude Code (6.3%).

combined_crest_p2_4

Results

To ensure the reliability of our evaluation, the reported Pass@1 scores represent the average performance across four independent runs in Terminal Bench 1.0 and 2.0.

Terminal Bench 1.0 Performance

Model Agent pass@1 pass@4
LiteCoder-Terminal-30b-a3b-sft Terminus 2 24.38% 40%
Qwen3-30B-A3B-Nex-N1 Openhands 18.44% 32.5%
LiteCoder-30a3b-Terminal-preview Terminus 2 16.56% 27.5%
Qwen3-30B-A3B-Instruct Terminus 2 16.56% 28.75%
LiteCoder-Terminal-4b-sft Terminus 2 14.69% 28.75%
OpenThinker-Agent-v1 Terminus 2 11.25% 25%
LiteCoder-4b-Terminal-preview Terminus 2 9.38% 20%
Qwen3-4B-Instruct Terminus 2 6.25% 15%

Terminal Bench 2.0 Performance

Model Agent pass@1 pass@4
LiteCoder-Terminal-30b-a3b-sft Terminus 2 12.36% 23.60%
Qwen3-30B-A3B-Nex-N1 Openhands 12.36% 23.60%
LiteCoder-30a3b-Terminal-preview Terminus 2 6.18% 13.75%
Qwen3-30B-A3B-Instruct Terminus 2 5.34% 11.24%
LiteCoder-Terminal-4b-sft Terminus 2 4.78% 10.11%
OpenThinker-Agent-v1 Terminus 2 4.49% 10.1%
LiteCoder-4b-Terminal-preview Terminus 2 4.78% 12.36%
Qwen3-4B-Instruct Terminus 2 1.12% 3.37%

Terminal Bench Pro Performance

Model Agent pass@1
LiteCoder-Terminal-30b-a3b-sft Terminus 2 31.5%
LiteCoder-30a3b-Terminal-preview Terminus 2 22.0%
LiteCoder-Terminal-4b-sft Terminus 2 21.5%
Qwen3-30B-A3B-Nex-N1 Openhands 21.0%
Qwen3-30B-A3B-Instruct Terminus 2 20.5%
OpenThinker-Agent-v1 Terminus 2 19.5%
LiteCoder-4b-Terminal-preview Terminus 2 15.0%
Qwen3-4B-Instruct Terminus 2 3.5%

Citation

@article{peng2026litecoderterminal,
  title={LiteCoder-Terminal: Scaling Long-Horizon Terminal Environments for Learning Language Agents},
  author={Peng, Xiaoxuan and Zhang, Kaiqi and Lu, Xinyu and Cao, Boxi and Lu, Yaojie and Lin, Hongyu and Han, Xianpei and Sun, Le},
  journal={arXiv preprint arXiv:2605.29559},
  year={2026}
}
Downloads last month
12

Paper for alucent/mirror-LiteCoder-Terminal-SFT