RepoFactory / README.md
aaabiao's picture
Update RepoFactory to the 200-environment release: readme_and_instances
8946d6c verified
|
Raw History Blame Contribute Delete
6.44 kB
metadata
pretty_name: RepoFactory-200
language:
  - en
size_categories:
  - n<1K
task_categories:
  - text-generation
tags:
  - sft-environment
  - pre-evolution
  - harbor
  - code
  - software-engineering
  - repository-generation
  - long-horizon
  - python
  - docker
viewer: false

RepoFactory

200 verifiable environments for full-repository generation.

RepoFactory constructs training environments for long-horizon coding agents from real-world software projects. Each task asks an agent to reconstruct a complete, working repository from a natural language specification and API guide. The agent works in a fresh workspace with dependencies installed, and its implementation is evaluated through executable tests.

This release provides 200 environments before applying RepoFactory's Environment Evolution Loop, together with their corresponding Docker images. The broader RepoFactory collection described in the paper contains 10,159 environments. The released tasks primarily cover Python projects and can be used to collect coding trajectories, train agents with executable feedback, and study repository reconstruction.

The environments use the Harbor task format. Each package includes the task specification, runtime configuration, oracle implementation, and verifier. Oracle implementations and reference tests are included for reproduction and validation and must remain separate from the solving agent's workspace.

Release contents

Item Contents
Environments 200
Task packages instances/<id>/instance.tar.gz
Runtime images images/<id>.tar
Image tags repofactory/task:<id>
Agent workspace /workspace

Each instance can be downloaded and used independently. Start with one task before downloading the complete image collection. Additional disk space is needed for extracted task files, imported Docker layers, and trial workspaces.

.
├── README.md
├── assets/
│   └── ledgerlens_example.png
├── instances/
│   ├── 000001/
│   │   └── instance.tar.gz
│   ├── 000002/
│   │   └── instance.tar.gz
│   └── ...
└── images/
    ├── 000001.tar
    ├── 000002.tar
    └── ...

The files under images/ are Docker image archives exported for docker load.

Quickstart

The example below runs instance 000001 locally. It requires Docker with Linux/amd64 container support and a Python environment compatible with Harbor. See the Harbor installation guide for host requirements.

1. Install the tools

python -m pip install --upgrade huggingface_hub harbor

2. Download one instance and its image

Download an instance and its image from IQuestLab/RepoFactory.

export HF_DATASET_REPO="IQuestLab/RepoFactory"

hf download "$HF_DATASET_REPO" \
  instances/000001/instance.tar.gz \
  images/000001.tar \
  --repo-type dataset \
  --local-dir ./repofactory

cd repofactory

The Hugging Face download CLI supports downloading individual files without fetching the complete release.

3. Extract the task and load the image

mkdir -p tasks
tar -xzf instances/000001/instance.tar.gz -C tasks
docker load --input images/000001.tar
docker image inspect repofactory/task:000001

The extracted task is located at tasks/000001/. Its task.toml refers to the image loaded above. Harbor can use this prebuilt image directly. Rebuilding environment/Dockerfile instead requires its separately referenced repofactory/base:000001 image.

4. Check the oracle and initial workspace

harbor run --path tasks/000001 --agent oracle --env docker
harbor run --path tasks/000001 --agent nop --env docker

The oracle agent copies the reference implementation into the workspace. The nop agent leaves the initial workspace unchanged. The intended validity criterion is full reward for the oracle and zero reward for the unchanged workspace; inspect both trial results and verifier logs when validating a local setup.

5. Run a coding agent

For example, run an agent through the Claude Code harness:

harbor run \
  --path tasks/000001 \
  --agent claude-code \
  --model "YOUR_MODEL_ID" \
  --env docker

Replace YOUR_MODEL_ID and configure credentials for the selected agent and model provider. Agent selection, model configuration, and parallel execution are described in the Harbor CLI documentation.

Instance format

Every task archive expands into a directory named after its six-digit instance ID:

000001/
├── task.toml
├── instruction.md
├── start.sh
├── environment/
│   ├── Dockerfile
│   └── start.md
├── solution/
│   ├── solve.sh
│   └── oracle/
└── tests/
    ├── test.sh
    ├── config.json
    └── reference/
Path Purpose
task.toml Task metadata, timeouts, resource limits, and prebuilt image tag
instruction.md Agent-facing instructions for reconstructing the project
environment/start.md Functional specification, repository structure, public interfaces, and usage examples; copied to /workspace/start.md
environment/Dockerfile Runtime image recipe and initial workspace setup
start.sh Startup configuration, including restrictions on common code-hosting domains
solution/oracle/ Reference implementation for oracle validation
solution/solve.sh Oracle entrypoint that copies the implementation to /workspace
tests/config.json Reference test paths, test commands, fixtures, package files, and configured test count
tests/test.sh Verifier entrypoint that restores evaluation resources and runs the configured tests
tests/reference/ Reference tests and supporting evaluation files

Resource limits and timeouts are specified in each task's task.toml. For instance 000001, the configuration allows 2 CPUs, 8 GiB of memory, and 20 GiB of storage, with timeouts of 60 minutes for the agent, 20 minutes for verification, and 30 minutes for environment construction.