abdelstark's picture
Add CodeLeWM public shard dataset card
57cd128 verified
|
Raw
History Blame Contribute Delete
3.22 kB
metadata
pretty_name: CodeLeWM Public Code-Edit Shard
license: other
tags:
  - code
  - python
  - world-model
  - code-edit
  - research-artifact

CodeLeWM Public Code-Edit Shard

This dataset repository hosts the public-safe CodeLeWM code-edit transition shards used by the scaled May 2026 CodeLeWM training runs. It is a diagnostic research artifact, not a benchmark leaderboard dataset and not a source of broad model-quality claims.

What Is Hosted Here

Run family Path pattern Role
Scaled first public run runs/codelewm-scaled-20260520-9699b53/pack/** scaled code-edit transition pack
Action-use margin run runs/codelewm-action-use-20260520-6650183/pack/** action-use margin experiment data
Action-use retrieval run runs/codelewm-action-use-retrieval-20260520-7895d18/pack/** margin plus retrieval follow-up data
v0.2 action-swap/inverse-action run runs/codelewm-v0-2-action-swap-rerun-20260520-7c7cb0b/pack/** action-contrast remediation data

Each pack includes manifest-backed metadata used by the CodeLeWM CLI. The corresponding model checkpoints are in abdelstark/codelewm-transition-model, and result reports are in abdelstark/codelewm-runs.

Dataset Information

The public shard represents Python code-edit transitions extracted for CodeLeWM. Records are designed for transition-model training and evaluation: before-state features, action/instruction views, after-state targets, split metadata, and manifest checksums. The dataset is used by the package-native training, retrieval, ablation, surprise, scorer-quality, and reranking diagnostics.

Candidate code and source data are treated as untrusted inputs. CodeLeWM parsing and scoring paths do not execute candidate patches by default.

Claim Boundary

The action-use experiments trained from these shards did not establish a positive action-conditioned model-quality claim. In the documented runs, no-action controls remained stronger on headline retrieval gates or later claim gates failed. Use this repository for reproducible negative and diagnostic evidence, not for claims of broad coding improvement.

Verification

Download one run pack and verify it locally:

hf download abdelstark/codelewm-public-shard \
  --repo-type dataset \
  --include 'runs/codelewm-v0-2-action-swap-rerun-20260520-7c7cb0b/pack/**' \
  --local-dir .artifacts/hf-download/codelewm-public-shard

uv run codelewm manifest verify \
  --manifest .artifacts/hf-download/codelewm-public-shard/runs/codelewm-v0-2-action-swap-rerun-20260520-7c7cb0b/pack/manifest.json \
  --json

uv run codelewm secret-scan \
  .artifacts/hf-download/codelewm-public-shard/runs/codelewm-v0-2-action-swap-rerun-20260520-7c7cb0b/pack \
  --json

Expected result: manifest verification returns ok=true and the secret scan returns ok=true with zero findings.

Related Repositories

  • Model checkpoints: https://huggingface.co/abdelstark/codelewm-transition-model
  • Current execution dataset: https://huggingface.co/datasets/abdelstark/codelewm-execution-pack
  • Run artifacts: https://huggingface.co/datasets/abdelstark/codelewm-runs
  • Code repository: https://github.com/AbdelStark/CodeLeWM