---
library_name: pytorch
license: other
tags:
- code
- python
- world-model
- transition-model
- reranking
- research-artifact
datasets:
- abdelstark/codelewm-public-shard
- abdelstark/codelewm-execution-pack
---

# CodeLeWM Transition Model Artifacts

This repository hosts CodeLeWM transition-model checkpoints and checkpoint
manifests for code-edit and execution-world-model experiments. CodeLeWM is a
research artifact for scoring and reranking candidate code states; it is not a
code generator and this repository is not an inference endpoint.

## What Is Hosted Here

| Artifact family | Repository path | Dataset surface | Claim posture |
| --- | --- | --- | --- |
| Early scaled code-edit transition run | `checkpoints/codelewm-scaled-20260520-9699b53` | `abdelstark/codelewm-public-shard` | smoke/scaled diagnostic |
| Action-use margin run | `checkpoints/codelewm-action-use-20260520-6650183` | `abdelstark/codelewm-public-shard` | negative action-use result |
| Action-use retrieval run | `checkpoints/codelewm-action-use-retrieval-20260520-7895d18` | `abdelstark/codelewm-public-shard` | negative against no-action control |
| v0.2 action-swap/inverse-action run | `checkpoints/codelewm-v0-2-action-swap-rerun-20260520-7c7cb0b` | `abdelstark/codelewm-public-shard` | negative action-use and representation result |
| v0.8 execution checkpoints, seeds 42 and 1729 | `checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-{42,1729}` | `abdelstark/codelewm-execution-pack` | mixed diagnostic execution evidence |

The final v0.9 seed-42 and seed-1729 execution runs are published in
`abdelstark/codelewm-runs`, not in this model repository. They are documented
by the CodeLeWM release cards and final public artifact index.

## Dataset Information

CodeLeWM uses two public dataset surfaces:

| Dataset | Role |
| --- | --- |
| `abdelstark/codelewm-public-shard` | Historical public-safe Python code-edit transition shards used by the scaled/action-use runs. |
| `abdelstark/codelewm-execution-pack` | Current execution-substrate pack of 2,188 tokenized `(code, input, output)` records at revision `v0.9.0-rc1`. |

The execution pack records a deterministic sandbox policy, source provenance,
split policy, checksums, and a claim boundary. It contains tokenized code,
inputs, outputs, and metadata; training and scoring do not execute candidate
code.

## Intended Use

- Reproduce CodeLeWM checkpoint loading, retrieval, surprise, scorer-quality,
  and reranking diagnostics.
- Compare transition-model scores against no-action, shuffled-action, lexical,
  random, and LLM-order controls.
- Inspect checkpoint manifests and trust-gate metadata before using a
  checkpoint in local scoring.

## Out Of Scope

- Generating code.
- Claiming broad coding improvement or live patch utility.
- Treating a green manifest, demo, or checkpoint load as a model-quality claim.
- Loading checkpoints without the CodeLeWM checkpoint trust gates and manifest
  verification.

## Claim Boundary

The tested code-edit action-use interventions are negative. The v0.9/v1.0
release evidence supports a narrow HumanEval WS-D diagnostic reranking slice,
while the aggregate downstream claim remains closed because MBPP-Plus is
saturated against no-action and lexical controls. This repository therefore
supports reproducible diagnostic research, not a general claim that CodeLeWM
improves coding.

## Verification

Download a hosted checkpoint family and verify its manifests:

```bash
hf download abdelstark/codelewm-transition-model \
  --repo-type model \
  --include 'checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-42/**' \
  --local-dir .artifacts/hf-download/codelewm-transition-model

uv run codelewm manifest verify \
  --manifest .artifacts/hf-download/codelewm-transition-model/checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-42/manifest.json \
  --json

uv run codelewm secret-scan \
  .artifacts/hf-download/codelewm-transition-model/checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-42 \
  --json
```

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

## Primary References

- Code repository: `https://github.com/AbdelStark/CodeLeWM`
- Execution dataset: `https://huggingface.co/datasets/abdelstark/codelewm-execution-pack`
- Historical code-edit shard: `https://huggingface.co/datasets/abdelstark/codelewm-public-shard`
- Run artifacts: `https://huggingface.co/datasets/abdelstark/codelewm-runs`
- Final artifact index: `docs/benchmark/PUBLIC_ARTIFACT_INDEX_2026-06-08.md`