james77mill/chessbench-encoded-npz

ChessBench Encoded NPZ (dense policy + WDL + moves-left)

Pre-encoded training shards for the chess-rl transformer (Phase 3, ~100M positions). These are the model-ready tensors — loading them skips the download + preprocess step entirely.

• Train: ~100.27M positions across 585 shards (train_*.npz)
• Val: ~0.20M positions across 195 shards (val_*.npz)
• Total: ~37 GB, 780 .npz files, flat layout.

Provenance (regenerable)

Derived deterministically (no Stockfish at encode time) from the public dataset prdev/chessbench-full-policy-value, itself built on DeepMind's ChessBench. This repo is a convenience cache of the encoded form so a fresh GPU run does not have to re-preprocess.

Schema (per shard)

Each .npz holds one shard of N positions (train shards: N=250000):

key	shape	dtype	meaning
square_tokens	(N, 64)	int8	per-square piece token (board encoding)
state_features	(N, 18)	float32	side-to-move / castling / counters state vector
wdl	(N, 3)	float32	win/draw/loss target distribution
moves_left	(N,)	float32	moves-left (game-length) target
counts	(N,)	int32	number of legal moves for each position
legal_indices	(sum(counts),)	int32	CSR-flattened legal-move policy indices
legal_probs	(sum(counts),)	float32	dense policy prob aligned with legal_indices

The policy target is stored CSR-style: position i owns the slice of legal_indices / legal_probs of length counts[i], starting at cumsum(counts)[i-1].

Loading example

import numpy as np
d = np.load("train_00_00000.npz")
sq   = d["square_tokens"]      # (250000, 64) int8
st   = d["state_features"]     # (250000, 18) f32
wdl  = d["wdl"]                # (250000, 3)  f32
ml   = d["moves_left"]         # (250000,)    f32
cnt  = d["counts"]             # (250000,)    i32
li   = d["legal_indices"]      # ragged, i32
lp   = d["legal_probs"]        # ragged, f32
offs = np.concatenate([[0], np.cumsum(cnt)])
# position i policy: li[offs[i]:offs[i+1]] -> lp[offs[i]:offs[i+1]]