Reproduce / November 3, 2026 tournament

Build the tournament bot from random weights.

A concrete recipe, archived inputs and inspectable code for CM-14. Follow the verified boundaries as carefully as the commands.

Recipe reproduction, not a guaranteed identical result. The original checkpoint is identified below. A clean-machine, sixteen-hour retraining of this package has not been completed. CCRL move data is intentionally not mirrored because a redistribution grant was not established: obtain its original PGNs locally, then use the supplied exact-key reconstruction. On November 3, automated checks of the recorded CCRL archive URLs returned HTTP 403; do not bypass that restriction. This is an access gap as well as a mirroring-rights gap. Account deployment also needs your own bot credentials and local configuration. Do not call missing inputs a successful reproduction.

This is the bot used in the November 3 arena, not just the unassisted browser model. The tournament combined CM-14, legal move selection, asynchronous Stockfish 19 mate checks and the separately enabled endgame conversion rule. Its 7 wins and 29 losses are a deployment observation, not a repeatable strength target.

A prompt for your coding agent

Copy this prompt together with this page's URL. The agent should inspect the package and report gaps before consuming a training budget.

Reproduce LittlePly's November 3, 2026 tournament system using this guide and its hash manifests. Start CM-14 from random weights, not CM-13 or downloaded pretrained weights. Inspect the licenses, source manifest and explicit reproduction gaps first. Download and hash-check the 66 raw/tokenized human archives, the frozen row selections, and the permitted engine records. Obtain CCRL originals from their listed upstream sources under their terms; reconstruct all 54,515 required CCRL sequences by exact hash. Never substitute another dataset or silently skip missing games.

Use the pinned two-GPU recipe: 12 layers, 768 width, 12 heads, 512 context, 1,975 vocabulary entries; seed 20261103; clean full-UCI-move history; separate fixed-White WDL normalization; hybrid Muon/AdamW. Run the eight-hour mixed-data phase, then the eight-hour human-only continuation, restoring model and both optimizer states. Verify initialization, dataset identity, masks, sampler continuity, legal promotions, finite losses and checkpoint integrity before long runs. Document any portability edits with old/new hashes.

Evaluate a newly frozen checkpoint with the provided 100-game Sunfish schedule and the recorded Stockfish 19 mate-only configuration. Endgame takeover is OFF for this benchmark. Preserve all games, errors, clocks, legal replay audits and unfinished outcomes. Report W/D/L and limitations; do not invent Elo or promise the historical 72% score.

For the tournament serving configuration, enable the separate <=7-piece endgame conversion only after Stockfish's bot-relative win estimate exceeds 97% on two consecutive bot turns, under the existing one-second and clock safeguards. Use the exact-history asynchronous mate helper. Never publish credentials or automatically create an account, enter an event or send messages. Ask the operator to authorize their own account deployment. Report verified artifacts, deviations and remaining blockers. Do not claim bit-identical weights, calibrated WDL or tablebase-proven wins.

Download the actual ingredients

Open the complete artifact directory: source ZIP, checksums, all monthly raw download links, token archive chunks, selection references and engine-source provenance. Each archive is independently hashed. Twenty-MiB chunks keep large datasets compatible with the static host; the downloader rejoins and verifies them.

IngredientWhat is suppliedBoundary
Human games66 monthly Elite archives, June 2020–November 2025; original ZIP URLs and hashes; token and record archivesLichess CC0 source games. 7.14 GB raw ZIPs; 8.20 GB uncompressed tokens/records.
Training selectionsStage A ranked references and move prefix; Stage B frozen referencesThese are the original selections, not a new scrape.
Engine mix1,282 TCEC and 116,859 Fishtest sequences, book masks/outcomes and attribution; exact selection metadata for 54,515 CCRL gamesCCRL moves are omitted; reconstruct locally from the listed originals. TCEC is CC BY-SA 3.0; Fishtest is LGPL-3.0.
CodeTraining, full-move adapter, preprocessing, evaluation, live serving and Stockfish integration; original and published file hashesHost paths relocated; remaining WORKSPACE markers identify account/runtime integration work. No credentials, private state or compiled Stockfish binary.
Reference browserONNX weights, source and CPU/WebGPU parity testsUnassisted inference; not the tournament helper configuration.

Rights and source references: Lichess CC0, Elite archive provider, TCEC archive, official Fishtest dataset, Stockfish 19 and corresponding source. Component licenses remain in the bundle; the project's Apache license does not replace them.

What was trained

The 86,966,016-parameter GPT-2-style decoder has 12 layers, width 768 and 12 heads, zero dropout and tied embedding/output weights. A game begins with BOS 1971; full UCI moves use IDs 0–1967, including every promotion piece. WDL classes 1968–1970 mean White win, draw and White loss. PAD 1972, unused QUERY 1973 and EOS 1974 complete the vocabulary. Outcome labels never enter the input.

Move and outcome logits are normalized separately. Every eligible move of both players is supervised; engine book moves are context only. A deterministic 15% sample, stratified over game thirds, receives WDL supervision. The summed move and selected WDL losses are divided by the globally accumulated eligible-move count. The outcome values are estimates, not calibrated probabilities or move-search supervision.

  1. Stage A: eight hours from random initialization. Seed 20261103, sampler 1301. 23,691,842 selected human games and 172,656 engine games. Sampling targets 20% actual eligible engine moves using accumulated sampling counts, not tenfold record repetition. Lower-player rating, mean rating and sequence hash define human ranking. Exact duplicates and protected opening families are excluded. The observed terminal update was 105,373.
  2. Stage B: eight-hour continuation. Restore Stage A weights, both optimizers and rank RNGs. Switch to the frozen 8,137,577 human-game view, with 689,547,458 unique move targets across 53 nonempty months; balance exposure by supervised moves per month. Sampler seed 20261103. Repetition is expected. The observed terminal update was 210,395, after 1,550,814,060 continuation move targets.

Both phases use FP32 master parameters and optimizer states with BF16 autocast, two-GPU DDP, eight packed 512-token rows per rank and two accumulation steps. Block-causal attention keeps packed games isolated. Two-dimensional matrices inside transformer blocks use Muon at 4.5e-4; embeddings, the tied output weights and other parameters use AdamW at 3e-4. Matrix weight decay 0.1, vector decay zero, gradient clipping 1.0; exact partitions and numerical settings are in the configuration and optimizer code.

Each phase has a 28,800-second wall budget: 28,500 seconds of training plus 300 seconds for final save. Warmup is 1,440 seconds, plateau ends at 24,900 seconds, and cooldown lasts 3,600 seconds. The continuation begins its ramp at one tenth of the peak. Since the schedule is time-based, different load or hardware changes the number of updates. Deterministic seeds do not make the final weights bit-identical across machines.

Recorded source checkpoint SHA-256: 31451d0d425ef14e020b847a3d1fefcbceacd32b0e1123adeab5f62877c379c2. Inference export: 4cee1bf258f6061843f6073cd21d65f8981153ae47108551df617cb3053eea72. Tournament release: 4cee1bf258f6-2a405de5e368. Container hashes differ; the export preserved all model tensors.

Prepare and run

The recorded environment used Linux under WSL, Python 3.12, PyTorch 2.9.0+cu128, Transformers 5.21.0, NumPy 2.5.4 and two RTX 4090 GPUs (24 GB each). Budget sixteen wall-clock training hours plus preparation/evaluation. Reserve at least 60 GB of working disk for data, archives and multiple optimizer checkpoints; additional rollback snapshots need more. The code currently requires two ranks; other hardware and smaller GPUs have not been validated.

unzip source.zip -d littleply-reproduction
cd littleply-reproduction
python3.12 -m venv .venv
. .venv/bin/activate
python -m pip install torch==2.9.0+cu128 --index-url https://download.pytorch.org/whl/cu128
python -m pip install -r requirements.txt
# Install zstd through your operating system's package manager.
# Download human-datasets.json from the artifact directory.
python fetch_data.py --manifest human-datasets.json
# Optional independent raw-source reconstruction/audit:
python fetch_data.py --manifest human-datasets.json --raw
# Download and verify the auxiliary chunks using release.json.
# Concatenate each listed sequence in order and verify its full SHA-256.
cat selection-refs.tar.zst.part* > selection-refs.tar.zst
zstd -dc selection-refs.tar.zst | tar -xf -
cat engine-selection.jsonl.gz.part* > engine-selection.jsonl.gz
# Extract the hash-verified CCRL originals into your local PGN directory.
python rebuild_engine.py --selection engine-selection.jsonl.gz \
  --ccrl-pgn-dir /path/to/local-ccrl-pgns --output engine
python prepare_run.py a
torchrun --standalone --nproc_per_node=2 stage-a/train.py --run-dir stage-a
python prepare_run.py b --parent stage-a/final.pt
torchrun --standalone --nproc_per_node=2 stage-b/train.py --run-dir stage-b

Run from the extracted bundle root. The engine rehydrator refuses to continue if any selected sequence is missing; compact offsets intentionally change container hashes while preserving sequence keys, labels and book masks. Use --resume only with that run's saved state and original clock; it is not a fresh eight-hour extension. Check hardware temperatures and available memory independently: unavailable sensors were not treated as safe readings in the original run.

The source includes the original preprocessing code and historical hashes for inspection. The archived token path is the practical replay path; a full raw-PGN-to-token rebuild and full sixteen-hour retraining of this relocated package remain unverified. The frozen continuation selection is part of this artifact, rather than a promise that a new run will have consumed exactly the same Stage A games by its time cutoff.

Separate the model, benchmark and live bot

Model-only inference: serving/cm13/cm13_worker.py is the full-move adapter used for CM-14 despite its historical filename. Supply your checkpoint SHA and vocabulary SHA explicitly. It reads JSON lines containing an id and complete UCI history, and returns legal ranks and fixed-White WDL. It rejects context overflow instead of silently cropping history.

python serving/cm13/cm13_worker.py --checkpoint stage-b/final.pt \
  --sha256 YOUR_CHECKPOINT_SHA256 --vocab data/vocab.json \
  --vocab-sha256 9a537cc1d879330301dbcec81128bb36423f649809d26363612f367cfced8f7f
# stdin example: {"id":"opening","history":["e2e4"]}

Benchmark: obtain the official Stockfish 19 universal Linux binary and corresponding source. The recorded binary SHA is 0f83d24cc46d2c66c60f16001af5444873bc112b7d028594513426894c12da19. The supplied preparation command checks that binary, legally decodes CPU fixtures, then freezes the relocated harness hashes. Use a fresh result directory; do not merge old and new configurations.

python prepare_eval.py --checkpoint stage-b/final.pt \
  --stockfish /path/to/stockfish-linux-x86-64-universal
python evaluation/run_suite.py --subject CM13

The internal subject label CM13 selects the shared full-move adapter; the config records the actual checkpoint and display name. The 100-game schedule uses 50 paired openings, Sunfish at 1,000 requested nodes, 300+2 clocks, a 340-ply cap and symmetric legal draw claims. Caps remain unfinished outcomes, never invented draws. Endgame conversion and outcome search are OFF. This is a development suite, not an independently held-out general-strength measurement.

Tournament serving: the source snapshot includes lichess_service, ponder.py, stockfish_mate.py, conversion.py and their policy files. Persistent workers use two Stockfish engines, two threads and 64 MiB hash per engine for mate checks, with an aggregate CPU-slot limit. They ponder on opponent time; a fast reply can use the residual of a one-second minimum opportunity on our clock, clipped by clock safety. Each candidate search has a one-second cap. Exact-history identity, cancellation and stale-result rejection are essential. An explicit UCI mate report can change a move; an ordinary centipawn advantage cannot. A UCI mate report is not an independently checked proof certificate.

The separate conversion path uses one Stockfish thread and 64 MiB hash, no tablebases, and a maximum one-second analysis. It activates only at seven or fewer pieces including kings, with bot-relative estimated win probability strictly above 97% on two consecutive bot turns, under exact-history and clock safeguards. This is an engine estimate, not a guaranteed win.

Account serving is an integration step, not a one-command public enrollment. Configure your own account name, private token-file path, state directory, Python interpreter and release manifest; set the scorer checkpoint digest to your verified export rather than retaining the tournament digest. Inspect every remaining WORKSPACE marker. Preserve server backoff, quota state, uncertain requests and the one-outgoing-attempt rule. Test local play first, then explicitly authorize online play. Do not copy LittlePly's account identity or tournament presence configuration. The archive contains no working credential or private cooldown history.

What to verify, and what remains open

The package supplies the original ingredients and tested preparation components, but does not establish bit-identical training or a clean-machine end-to-end rerun. Raw-data preprocessing, unattended account setup and a new full training/evaluation cycle require independent verification. If a pinned package or original archive is unavailable, report that exact gap instead of substituting silently. Read the packaging and smoke-test receipt.