artist/gcc2025-chess-datasets

GCC 2025 Chess Policy Datasets (searchless + gold)

This repo contains the datasets used to train and evaluate a text-only chess policy model for the Global Chess Challenge 2025.

Files

• —searchless.jsonl: 2,048,000 searchless teacher-labeled positions.
• —gold.jsonl: 208,832 Stockfish-labeled positions (gold).
• —labels_merged.jsonl: merged labels (searchless + gold).
• —policy_sft_v2.jsonl: SFT training set with gold_repeat=3 applied.
• —labels_eval_5000.jsonl: fixed 5k canary set for offline evaluation.
• —notes/data_assets_20251231T220044Z.md: dataset inventory and backup notes.

Schema

Common fields

• —fen: FEN string
• —move_history_uci: optional move history (UCI)
• —legal_moves_uci: list of legal moves (UCI)
• —best_move: labeled best move (UCI)
• —label_source: searchless or stockfish

Searchless-only fields

• —teacher_model, teacher_checkpoint_step
• —top_moves, top_win_probs
• —bottom_moves, bottom_win_probs
• —best_win_prob

Gold-only fields

• —stockfish_depth

Intended use

• —Supervised policy SFT for chess move selection.
• —Preference/ranking training using top/bottom move lists and win probabilities.

Notes

• —Prompt templates are stored in the training repo; this dataset only includes labels.
• —The canary set is fixed to enable paired comparisons across checkpoints.