SWE-Lego Live / swegen

Outputs#

Files and directories produced by SWE-gen

SWE-gen outputs are designed for downstream blocks to consume verified tasks without guessing which generated directories are complete.

Task directories#

Per-language task directories live under:

artifacts/swe_tasks/<lang>-cc/

Each verified task directory contains:

File	Purpose
`instruction.md`	Natural-language problem statement
`environment/Dockerfile`	Reproducible build environment
`environment/bug.patch`	Patch that introduces the failing behavior
`solution/fix.patch`	Ground-truth fix
`solution/solve.sh`	Applies the fix patch
`tests/test.sh`	Verification entrypoint
`task.toml`	Metadata, scoring, timeouts, and tags

Verified manifest#

The most important output is:

artifacts/swe_tasks/<lang>-cc/verifiable_tasks.txt

Downstream consumers should filter through this manifest. It is the difference between "a task directory exists" and "the task is validated and safe to use".

Batch state#

Batch state lives next to each language output:

artifacts/swe_tasks/<lang>-cc/.swegen-create-batch/

This directory is needed for resume, deduplication, failure analysis, and progress accounting. It should be included in state packages, but not committed to Git.

Logs#

Create logs live under:

artifacts/logs/swegen-create/

These logs are useful for debugging model failures, Docker failures, and timeout tuning. They are audit artifacts rather than source files.

Merged verified tasks#

The block can optionally flatten verified tasks into:

outputs/<task_id>/

using:

python scripts/extract_verified_tasks.py

This is convenient for tools that expect one task root, but manifest-gated in-place consumption is preferred for large runs.

State packages#

When moving to another node, package runtime state separately from source code. A complete resume package should include:

artifacts/swe_tasks/<lang>-cc/
verifiable_tasks.txt
.swegen-create-batch/*.json
PR input files in artifacts/collected_prs/
optional logs for audit

Do not commit large runtime state to Git. Store it in Cloud storage, a release artifact, or a shared filesystem.

Dataset handoff#

The trajgen block consumes tasks after SWE-gen writes manifests. The handoff contract is simple:

1. SWE-gen writes task directories and verifiable_tasks.txt. 2. Trajgen copies or reads only task IDs listed in the manifest. 3. Trajgen rolls out agents and captures trajectories.

This keeps incomplete, failed, or partially generated task skeletons out of training data.