Pradheep1647/kibitzer-sft-elo-1129-step-016000
Kibitzer checkpoint for a chess policy/value model trained on game sequences. Each board is encoded as 64 square tokens plus auxiliary state, summarized by a square-level encoder, then processed by a causal transformer over the game timeline.
Status
checkpoint:step_016000.ptstep:016000elo_rating: 1129 estimated vs stockfish-elo-1320 (-190.8 Elo diff)github_repo: https://github.com/Mantissagithub/kibitzereval_opponent:stockfish-elo-1320elo_diff: -190.8eval_games: 20eval_score: 5.00elo_error: 217.1
Architecture
- Input: 64 chess-square piece tokens plus 7 auxiliary scalars covering side to move, castling rights, en-passant file, and halfmove clock.
- Position encoder: bidirectional square-level transformer that compresses each board into one timeline token.
- Timeline trunk: causal transformer with
d_model=384, sequence length256, 12 layers, and 8 attention heads. - Transformer blocks: pre-norm RMSNorm, RoPE on causal Q/K attention, PyTorch scaled dot-product attention, and SwiGLU feed-forward layers.
- Outputs: policy head over 4,672 AlphaZero-style moves (
64 * 73) plus a tanh value head for bounded outcome prediction.
Files
step_016000.pt: PyTorch checkpoint.training_metadata.yaml: training config and checkpoint metrics.post_eval.yaml: uploaded after local Stockfish/cutechess eval.
Usage
git clone https://github.com/Mantissagithub/kibitzer.git
cd kibitzer
uv run python scripts/uci.py --checkpoint <path-to>/step_016000.pt
This checkpoint is one artifact in a sequence. Repos with
elo-pending in the name have not been evaluated yet; rated repos are
renamed after scripts/eval_and_rename_hf.py --from-hf completes.