Built from 3.8M Lichess games (players ≥2400 vs ≥2200), this pack contains five themed opening books plus a combined Allbook. Goal: practical, Lichess-style play, broad choice, no junk. All moves are equally weighted, so the engine doesn’t bias for popularity and stays more diverse.
Books (games used):
lc-open.bin – 1.e4 e5 → 324,769
lc-semiopen.bin – Sicilian/French/Caro-Kann … → 1,342,110
lc-closed.bin – 1.d4 d5 → 306,022
lc-indian.bin – 1.d4 Nf6 (King’s Indian, Nimzo/… ) → 283,735
lc-flank.bin – 1.c4 / 1.Nf3 / 1.g3 … → 579,000
lc-allbooks.bin merges all five; duplicate entries are averaged, preserving the flat weighting.
What sets them apart
Flat weighting: several sound candidates per node.
Quality filters: only moves with sufficient games/share; minimum continuation depth; gentle transposition smoothing to include near-threshold candidates without noise.
Practical depth: typically around 20–22 ply, depending on family.
Safety checks: Polyglot verification and backward check (legality, parent/child hashes).
Usage
Format: Polyglot .bin (Zobrist-sorted).
Works with PicoChess, SCID, Lucas Chess, and many UCI GUIs.
Load the book and enable it—moves are chosen with equal probability until out of book.
Thanks: PGN data courtesy of https://database.nikonoel.fr/
Hello Johan,
Thanks for your feedback!
You’re right—there’s an issue with the allbooks build:
allbooks.bin was generated from the finished individual .bin books rather than directly from the PGN files.
The individual .bin books themselves are fine.
I’m currently working on V2 with a larger dataset, more recent games, and improved filters.
I’ll upload them here over the next few days, and the allbooks build will be error-free as well.
Best regards
Tobias