Datasets:
metadata
license: other
pretty_name: ProgramLab
configs:
- config_name: default
data_files: instances.jsonl
task_categories:
- text-generation
tags:
- code
- agents
- reinforcement-learning
- programbench
size_categories:
- 1K<n<10K
1,261 environments for training agents to rebuild real programs. In each task the agent can run a compiled program but not read it, reads its documentation, and writes the whole program again in any language. Hidden behavioral tests compare its program with the reference. Tasks use the ProgramBench format, so the official evaluator scores them unchanged.
| Tasks | Languages | Difficulty | Hidden tests | Code coverage |
|---|---|---|---|---|
| 1,261 | 566 Go · 415 Rust · 280 C/C++ | 205 easy · 888 medium · 168 hard | 187,073 (median 127 per task) | median 76.6% |
Quickstart
Start from the setup in the code repository.
Files
ProgramLab/
├── instances.jsonl one row per task: repo, commit, language, difficulty, size,
│ tests, coverage, license, image
├── tasks.tar.gz task definitions, installed into ProgramBench by setup.sh
│ └── <instance_id>/
│ ├── task.yaml
│ └── tests.json test ids per suite; excluded tests are marked
├── build/ to rebuild the images yourself instead of pulling them
│ ├── build.sh
│ ├── base.v6.tar.gz base image, for docker load
│ ├── base.v6.Dockerfile
│ ├── contexts/<instance_id>.tar.gz Docker build context of each task
│ ├── isolation_manifest.json hashes checked before each build
│ ├── verify_context.py
│ └── probe.py isolation check run inside each built image
└── assets/logo.png
The hidden tests live in ProgramLab-Tests.
Validation
Every task passed these checks before release:
- Cleanroom. The reference runs as an execute-only binary (mode
0111) with no network, no source, no build cache, and no recoverable Git history. Images are probed for these properties as the agent user. - Real and empty programs. The original program scores 100% under the official evaluator, and an empty program that does nothing scores 0. Tests that the original program fails (4,047) or that the empty program passes (1,735) are excluded from scoring.
Citation
@misc{programlab2026,
title={ProgramLab: Train single and multi agents to rebuild programs with RL},
author={Yao Tang and Daixuan Cheng and Li Dong and Zhihao Zhan and Shaohan Huang and Jiatao Gu and Furu Wei},
howpublished={Blog Post},
url={https://programlab.blog/},
year={2026}
}