LittlePlyModel + harness

Historical material

Preserved research material.

Earlier formats, artifacts and limitations remain available; current entry points are in the main navigation.

Return to current pipeline guide

Archived notebook content

Historical CM-14 checkpoint

The 86,966,016-parameter transformer reads clean, complete move history: one token per UCI move, including promotions. It has 12 layers, 12 heads, width 768, 512 context positions and 1,975 vocabulary entries. Move logits and fixed-White win/draw/loss logits are normalized separately. Outcome labels are not supplied as input, and the WDL diagnostic does not choose the browser move.

CM-14 began from fresh random initialization with a hybrid Muon/AdamW optimizer. This promoted checkpoint continues that CM-14 run; it is not a continuation of CM-13 weights.

Evaluation · Historical inference record

Historical notebook material and artifacts

Current final snapshot

Fresh random initialization with AdamW; clean game history, one token per complete UCI move (one ply), including promotions. The 86,966,016-parameter transformer has 12 layers, 12 heads and width 768, with a 512-token context and 1,975 vocabulary entries. No masking and no outcome labels in inputs. Move outputs and fixed-White WIN/DRAW/LOSS outputs are separately normalized from the shared transformer and tied output matrix.

Current input and serving interfaces · Final evaluation · Published browser package. Earlier explanations below retain their historical context; CM-10 is no longer the browser default.

Earlier notebook material and artifacts

Online on Lichess: CM-13, audited eight-hour snapshot. Play LittlePly on Lichess, with a different tokenizer and helper. CM-13 results and limitations.

Current online model: CM-13

CM-13 starts from random initialization with AdamW. Clean game history is encoded as one token per complete UCI move (one ply), including promotions, in a 512-token context. There is no masking and no outcome label in the inputs.

A shared 12-layer, 768-wide, 12-head transformer and tied output matrix produce two separately normalized distributions: move outputs and fixed-White WIN/DRAW/LOSS outputs. Outcome predictions are diagnostics, not policy conditioning or a search value function.

The model selects ordinary legal moves. A separate asynchronous Stockfish 19 helper can intervene for explicit offensive or defensive mates; centipawn scores do not trigger changes. This deployed system is not search-free. Engine mate reports are not independently verified proof certificates.

Eight-hour checkpoint, results and training context; serving metadata. The CM-10 explanations and artifacts below describe the separate browser model and earlier experiment.

Source map / CM-10

Trace the current release.

CM-10 uses the full-history codec and an inference export that preserves the approved training tensors.

CM-10 checkpoint · Evaluation and limitations · Retired browser artifact · Reproduction instructions

Historical inference source and full-history tokenizer lineage. Historical training and service snapshots below retain their original checkpoint identities; they are not an exact CM-10 training bundle.

Historical CM-7 and earlier source maps

Source map / current CM-7

Trace a move back to its training.

Follow the current implementation first. Historical pipelines remain below with their original source versions.

  1. Prepare examples and curriculum: ordinary games, book-supported moves and exact endgame alternatives; EXP-10 and EXP-11 record actual exposures.
  2. Encode current state: 120 prefix slots and three target tokens. Read the format.
  3. Train the initial curriculum, then the follow-up. CM-7 is the selected checkpoint, not the terminal update.
  4. Evaluate against Sunfish: protocol, outcomes and limitations, plus all games.
  5. Serve on Lichess or run in the browser. Their helper limits differ; the policy diagram separates model and search.

Check artifact availability and reproduction instructions before running code. Public snapshots redact host-specific paths and exclude credentials.

Historical full-history pipeline

The following map documents earlier model families. It remains useful for reproducing those experiments; use the current links above for CM-7.

  1. Data preparation and tokenizer: verified dataset-manifest code hashes; source-square, destination-square, promotion triples. Obtain Lichess source data separately.
  2. Training: complete-history causal objective. Read the recorded run configuration and results.
  3. Sunfish harness and model adapter: legal-move selection and separately labeled bounded forced-mate search. All 20 PGNs and replays.
  4. Historical GPU Lichess service: immutable release 6133648b2926-c205ec8f7f77, supervised locally with systemd. Casual standard-start chess from bullet to classical, including zero-increment games. Up to four simultaneous games when capacity allows; bullet games run alone. UltraBullet, correspondence and variants are unsupported. Availability depends on the local host remaining awake and running. Open the editable challenge dialog. Multi-game admission/controller, scheduler and inference owner are tied to the release manifest. This source subset excludes host configuration and launcher; it is not a complete deployment bundle. The CPU-service snapshot and initial bridge snapshot retain their historical limits. That historical snapshot retains its original checkpoint and evaluation; see the current release above.

Clean website build

npm ci
npm run check
npm run preview

Use Node 24.22.1. The lockfile pins dependencies. The build needs only the checked-in, sanitized notebook content and performs no training, match play, artifact retrieval or deployment. It emits static files to notebook/dist.

Reproducing model experiments

The exact recorded configuration, checkpoint digests, source hashes and PGNs are provided. Model-only exports and filtered raw/tokenized datasets are available. Resume checkpoints are excluded, and byte-exact pilot code is unavailable. Therefore this release is an inspectable experiment record, not yet a one-command reproduction bundle. Path-redacted source derivatives are labeled explicitly. Before a rerun, obtain licensed dependencies and data, restore local paths, and validate hashes against the recorded receipts. Do not interpret historical scripts with machine-specific launchers as portable run instructions.

Rust implementations

Preprocessing and token codec · Buildable codec source ZIP · Parallel mate search. Versioned source, tests, dependency locks and build guides; separate component licenses retained.