SWE-Lego Live / swegen

Motivation#

Why we built SWE-gen

SWE-gen turns real GitHub pull requests into verified SWE-style coding tasks. It is the task-curation stage of the SWE-Lego-Live pipeline, sitting before trajectory generation, SFT, and RL.

GitHub PRs -> swegen -> trajgen -> sft -> rl

SWE-gen exists because high-quality agent training data needs more than a patch and a repository URL. Each task must have a reproducible container, a clear instruction, a bug-introducing patch, a ground-truth fix, and a test script that separates solved from unsolved attempts.

SWE-gen provides:

• Multi-language PR collection across Python, JavaScript, TypeScript, Go, C, C++, Java, and Rust
• LLM-assisted PR filtering and task instruction generation
• Harbor-compatible task directories with Docker environments and tests
• NOP and Oracle validation before a task is exposed downstream
• verifiable_tasks.txt manifests that downstream blocks can trust
• Batch state and resume support for long-running data construction
• Static difficulty scoring and task metadata for dataset analysis

Where to go next#

• Getting Started - set up the block and run a smoke generation
• Core Concepts - PR pools, task skeletons, validation, and manifests
• Run Generation - run per-language generation scripts safely
• Outputs - understand task directories, manifests, logs, and state packages
• Dashboard - monitor SWE-gen progress online