avewright/chessbot-sf19-policy

ChessBot SF19 Policy

MultiPV policy labels from Stockfish 19, stored in ChessBot's I/O, not the project's compact-1968 vocab.

23,899,367 unique positions. Honor split (1 = 5% holdout). bucket is soft / opening / middlegame / endgame / puzzles. Value CE only when wdl_source is 1 (live official UCI) or 3 (imported SF19 teacher WDL). 0 is policy-only (bestline / puzzles).

Why this shape

Published ChessBot / exp287 expect:

piece	size	notes
input	19 planes × 64	token 0 = a8 (rank-flipped). Use fens_to_planes(fen)
policy	1929	first 1858 = ChessFENS LC0 moves, last = padding_token
value	3	[black_win, draw, white_win]

Do not train this pack through a 1968-way head. Do not run stm_wdl_to_chessbot again — wdl is already ChessBot order. Both colors are stored as-is; do not assume white-perspective like raw ChessFENS.

Knight promotions drop the trailing n. King-to-rook castle (e1h1) is stored as e1g1. Illegal / unlabelled logits are -1 after expand (ChessFENS convention). Only MultiPV support has mass.

Load

import numpy as np
from datasets import load_dataset
from chess_chessbot import CHESSBOT_VOCAB_SIZE, expand_chessbot_policy, fens_to_planes

ds = load_dataset("avewright/chessbot-sf19-policy", split="train")
train = ds.filter(lambda r: int(r["split"]) == 0)

row = train[0]
planes = fens_to_planes([row["fen"]])          # [1, 64, 19]
policy = expand_chessbot_policy(row["policy_idx"], row["policy_p"])  # [1929]
wdl = np.asarray(row["wdl"], dtype=np.float32)  # [black, draw, white]
assert policy.shape == (CHESSBOT_VOCAB_SIZE,)
assert abs(float(policy.clip(min=0).sum()) - 1.0) < 1e-5

Loss: hard CE on hard_idx + soft CE on the expanded 1929 vector (clamp -1 → 0, renormalize). Value CE / KL on wdl. Drop wdl_source != 1 if you only want official UCI WDL.

Columns

column	meaning
fen	full FEN (clocks from the self-play game)
policy_idx / policy_p	up to 8 ChessBot-1929 slots, pad -1 / 0
hard_idx	best ChessBot index
wdl	ChessBot [black, draw, white]
wdl_source	1 live official UCI, 3 imported teacher, 0 policy-only
bucket	soft / opening / middlegame / endgame / puzzles
wdl_white	same triple White-absolute [W, D, B]
wdl_raw	official per-mille, White-absolute, sum 1000
n_soft	how many MultiPV moves mapped
split	0 train, 1 holdout

Teacher

• Stockfish 19, binary SHA-256 a18534ee5eab7de770692a1219e477b31a98f90555431df44e16ea2373a0a6e1
• 100,000 nodes / MultiPV=8 / tau=120
• Official WDL required. No sigmoid fallback.
• ECO starts: 7,852 unique (A 817 / B 772 / C 1,250 / D 614 / E 357)

Do not mix with avewright/local-wdl (value-only, White-absolute, compact). Do not mix with compact chess-soft-sf19 without remapping.