LittlePlyResearch notebook

Reproduction / CM-7

Follow the exact lineage.

Source, selected-weight identities, evaluation PGNs and sanitized evidence are versioned separately from host-specific launch machinery.

  1. Read training.json and evaluation.json. CM-6 → selected finite-curriculum weights → selected follow-up update 2,232 is the lineage. Terminal weights are distinct.
  2. Download source.zip; verify each rendered hash against the manifest. Original hashes identify the local research files before path redaction. Extract into a new directory.
  3. Inspect the state codec/model, dataset builders, finite-schedule trainers and evaluation adapter. Dependencies include Python, PyTorch, Transformers, NumPy and python-chess. The browser release separately pins its ONNX Runtime and chess.js distributions.
  4. For a source-only codec check, place the state snapshot on PYTHONPATH and run python -m unittest discover -s state-v2-cm7 -p test_codec.py with its imports installed. Other tests require corresponding source directories and fixture/data dependencies; read each test before execution.
  5. Training requires reconstructing the protected datasets, manifests, schedule, checkpoint metadata and reviewed configuration. Set local paths explicitly; do not run redacted research launchers unchanged. The source hashes and sampler/optimizer counters are part of the run contract.
  6. Replay all three PGN collections and retain their separate checkpoint identities. Minis overlap the full evaluation; do not combine them as 120 independent strength trials.

Availability boundaries

This release publishes the implementation, sanitized counts/hashes and all 120 final diagnostic/evaluation PGNs. It is not a turnkey bit-identical training replay: the full state-curriculum payload, historical optimizer/resume artifacts and host launch configuration are not bundled. Earlier downloadable full-history datasets are not substitutes for the CM-7 curriculum. Browser weights and their exact manifest are linked from browser play.

Licenses

Original LittlePly source is Apache-2.0. Lichess-derived source data and official-stockfish/books opening sources are CC0; dependency licenses remain separate. No Syzygy files or third-party engine binaries are included in this source bundle. The separate Rust mate-helper component and its dependencies retain their existing GPL notices. License index.

Inspect and reproduce

Training counts and checkpoint lineage · Evaluation evidence and hashes · Download allowlisted source · Original/rendered source hashes · Execution-pin matches

State codec and model · Data preparation · First-stage trainer · Follow-up trainer · Evaluation harness · Separate GPL Rust helper

Reproduction instructions and availability boundaries. Public source replaces host-specific paths with /work placeholders; both hashes are recorded. Historical snapshots remain unchanged.