sovaz1997/zevra-selfplay

Zevra Self-Play Training Data

Raw self-play positions generated by the Zevra 2 chess engine, used to train its NNUE evaluation network.

Provenance — trained from scratch on own play only. Every position here was produced by Zevra playing itself. No hand-crafted-evaluation labels and no games from any other engine were used. The data is the product of a reinforcement loop seeded by the Zevra 2.6 network: each generation's champion generates self-play games, a fresh network is trained from scratch on them, and the stronger network then generates the next, better games (g0 → g1 → g2 → g3).

Contents

• zevra-selfplay-raw.tar.zst — the full cumulative corpus, zstd-compressed (~20 GB compressed, ~110 GB / ~1.8 billion positions uncompressed).
• SHA256SUMS.txt — checksum for integrity.

The public Zevra 2.7 network was trained on the ~1.5B-position subset available at the g3 generation; this archive is the full cumulative set.

Format

Plain text, one position per line:

<FEN> | <eval_cp> | <result>
rn1qkbnr/pb1pp1pp/2p5/1p3p2/2P5/1Q5N/PP1PPPPP/RNB1KBR1 w KQkq - 6 7 | 123 | 0.0

• FEN — the position.
• eval_cp — the engine's search evaluation in centipawns.
• result — the game's final result (1.0 win / 0.5 draw / 0.0 loss).

Positions were generated by fixed-node self-play with light opening randomness. For the exact sign/perspective conventions see the datagen code in the engine repository.

Usage

Decompress:

# modern GNU tar auto-detects zstd:
tar xf zevra-selfplay-raw.tar.zst
# or explicitly:
zstd -d --long=27 zevra-selfplay-raw.tar.zst && tar xf zevra-selfplay-raw.tar

This unpacks a collected/ tree of .txt shards in the format above. To train an NNUE the same way Zevra does, convert the text to the bullet bulletformat and train with bullet:

# using bullet-utils (from the bullet repo)
bullet-utils convert --from text --input <shards.txt> --output data.bin

License

Released under CC0-1.0 (public domain dedication) — use it for anything.

Citation

If you use this data, a link back to the engine is appreciated: https://github.com/sovaz1997/Zevra2