harbor run -d LiteCoder/LiteCoder-rllicense: mit task_categories:
- text-generation configs:
- config_name: default
data_files:
- split: train path: data/train.parquet tags:
- terminal
- rl
- code
LiteCoder-Terminal-RL-preview
This dataset contains 602 standardized Harbor terminal environments and was released as part of the paper LiteCoder-Terminal: Scaling Long-Horizon Terminal Environments for Learning Language Agents.
Unlike static text-only instructions, these environments are fully executable and are designed to support the training of terminal-based agents.
Environment Generation Pipeline
The lack of high-quality, executable training environments is currently a major challenge for training terminal agents. Raw task descriptions lack the execution feedback required for techniques like rejection sampling and reinforcement learning, so an executable environment is essential. We implemented a five-stage synthesis pipeline to convert sampled tasks into the Harbor format using the Claude Agent SDK:
- Task refinement: rewrite raw task descriptions into clear and testable instructions
- Environment setup: prepare the execution environment and required resources
- Reference solution generation: produce a working solution for the task
- Verifier creation: construct test cases and evaluation logic
- Harbor packaging: assemble everything into the standard Harbor format

Environment Structure
Each environment represents a standardized terminal task instance organized in the Harbor format, which is as the following:
├── instruction.md
├── task.toml
├── environment
│ ├── Dockerfile
│ └── ...
├── solution
│ ├── solve.sh
│ └── ...
└── tests
├── test.sh
└── ...
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}
}Displaying 0 of 602 tasks