abdelstark/codelewm-runs

CodeLeWM Run Artifacts

This dataset repository is the public artifact registry for CodeLeWM training, evaluation, demo, and publication runs. It is organized as manifest-backed run trees rather than as a conventional train/test dataset.

What Is Hosted Here

Artifact family	Example path	Role
v0.2 and earlier code-edit runs	runs/codelewm-action-use-retrieval-20260520-7895d18/**	negative action-use and retrieval evidence
v0.7/v0.8 execution runs	codelewm-v0-8-short-execution-20260605-1b737e4-seed-42/**	execution-transition diagnostics
v0.9 execution runs	codelewm-v0-9-short-execution-20260606-69f798a-seed-{42,1729}/**	final execution-training and reranking artifacts
Visual/demo bundles	visual-observability/** and related run paths	UI and harness diagnostics

Artifacts include manifests, configs, reports, checkpoint metadata, benchmark outputs, and selected checkpoint files. The primary current dataset is abdelstark/codelewm-execution-pack; historical code-edit shards are in abdelstark/codelewm-public-shard.

Dataset And Model Context

• Current execution dataset: abdelstark/codelewm-execution-pack@v0.9.0-rc1, 2,188 tokenized (code, input, output) records with deterministic sandbox provenance.
• Historical code-edit dataset: abdelstark/codelewm-public-shard, used by the scaled/action-use experiments.
• Historical transition-model checkpoints: abdelstark/codelewm-transition-model.
• v0.9 execution checkpoints and run reports are stored directly in this repository under their run ids.

Claim Boundary

CodeLeWM is claim-gated. These artifacts support reproducibility, inspection, and diagnostic comparisons. They do not support a broad claim that CodeLeWM generally improves coding. The final public evidence supports negative action-use conclusions plus a narrow HumanEval WS-D reranking slice; the aggregate downstream claim remains closed.

Verification

Download a v0.9 run tree and verify it locally:

hf download abdelstark/codelewm-runs \
  --repo-type dataset \
  --include 'codelewm-v0-9-short-execution-20260606-69f798a-seed-42/**' \
  --local-dir .artifacts/hf-download/codelewm-runs

uv run codelewm manifest verify \
  --manifest .artifacts/hf-download/codelewm-runs/codelewm-v0-9-short-execution-20260606-69f798a-seed-42/manifest.json \
  --parent-manifest .artifacts/v0_9/hf-pack-download/artifact_manifest.json \
  --json

uv run codelewm secret-scan \
  .artifacts/hf-download/codelewm-runs/codelewm-v0-9-short-execution-20260606-69f798a-seed-42 \
  --json

Expected result: manifest verification returns ok=true and the secret scan returns ok=true with zero findings. The parent-manifest path should point to a local download of abdelstark/codelewm-execution-pack@v0.9.0-rc1.

Related Repositories

• Code repository: https://github.com/AbdelStark/CodeLeWM
• Current execution dataset: https://huggingface.co/datasets/abdelstark/codelewm-execution-pack
• Historical code-edit shard: https://huggingface.co/datasets/abdelstark/codelewm-public-shard
• Historical transition-model checkpoints: https://huggingface.co/abdelstark/codelewm-transition-model