Chess-Nut-Engine/chess-reasoning-traces-seed

Chess reasoning traces (seed)

11905 verified reasoning traces for choosing a chess move, written to train a small model that reasons before it moves. Each trace explains 2-4 candidate lines that Stockfish chose and scored, then picks the best one. The language model wrote only the explanations, and every trace passed mechanical checks and two critic reviews before it was accepted.

Split	Rows	Contents
train	11905	accepted traces
rejected	396	traces that never passed review (kept for analysis)

Source	Accepted	Rejected
games	8520	277
forced	1728	51
chess960	1657	68

How it was built

1. Fresh positions. Training-split positions from three sources:

   • rated 2025 Lichess games, both players 1800+, 3+ minute games;
   • positions in check or with at most 3 legal moves, from 2025 games (1600+);
   • Lichess Chess960 games (1600+).

   Every position identity and every source game used in our earlier datasets (1.7M positions from 836k games) was excluded, and at most one position per game was kept.

2. Engine analysis. Stockfish 19 with Syzygy 3-5 piece tablebases:

   • 100k nodes, MultiPV 4;
   • two tempting captures or checks searched separately;
   • a 5k-node quick search.

3. Fixed skeleton. Candidates, their lines (up to 4 plies) and scores come from the analysis. The chosen move is the robust label: among moves within 0.03 expected score of Stockfish's best, the one the quick search ranks highest. Stockfish's own move is in engine_best_*. Candidates valued above the chosen move are left out, so it is always the highest stated score.

4. Teacher. openai/gpt-6-luna at high reasoning effort read a JSON report of the position (teacher_input) and wrote a position summary, one note per candidate and a plan (teacher_output). The report holds:

   • material, pieces, king safety, pawn structure, files, activity and tactics;
   • an attack map for every piece;
   • move-by-move facts along each line;
   • privileged engine analysis.

   The text uses an explicit voice that names sides and pieces by colour.

5. Review. A trace is accepted only when the mechanical checks pass and neither critic finds a false statement. Otherwise the teacher revises it, up to 3 times. Every round is in review.

   • Mechanical checks: every move written must occur on the board or a shown line; piece-placement and colour claims must be true; no engine talk.
   • Two critic reviews by openai/gpt-6-luna: one with the verified facts, and one that can also look at positions with board tools.

An independent referee (anthropic/claude-sonnet-5.5 ($12.09)) re-checked 800 randomly chosen accepted traces with the critic prompt: 732 (92%) had no false statement. Those rows carry referee_checked, referee_clean and referee_claims.

Teacher and critic calls cost $82.94 in total.

Fields

Field	Meaning
id, game_id, source_set	Position id, source game (lichess:<id>) and source set
fen, start_fen, history_uci, chess960, side_to_move, phase	The position and how it was reached
candidates	List of {move, line, score}: the fixed skeleton. Moves in SAN; scores in pawns for the side to move, #n for mate
chosen_move, chosen_move_uci	The move the trace picks
engine_best_san, engine_best_uci, robust_label	Stockfish's own top move, and whether the chosen move differs from it
teacher_input	The exact JSON report the teacher read
teacher_output	The teacher's JSON answer: position, notes (per candidate) and plan
trace	The rendered reasoning trace
prompt, reasoning, answer	A ready SFT example: the move prompt (with varied context), the trace and the move
revisions, review, accepted	Review rounds: mechanical flags, critic objections and costs
teacher_model, teacher_effort, voice, critic_model, critic_effort, critic_mode, engine, cost_usd	Provenance
referee_checked, referee_clean, referee_claims, referee_model	The independent spot check, where it was run

The trace format (Candidates: / - line: score / Best:) is checked by chess_llm.tasks.trace. The teacher's sentences are indented lines that this parser ignores.

from datasets import load_dataset
ds = load_dataset("Chess-Nut-Engine/chess-reasoning-traces-seed", split="train")

Limitations

• Correctness is checked, not proven. Critics checked every trace and a referee spot-checked a sample. Some accepted traces still contain a false statement (see the referee rate above).
• Coverage. Explanations cover the shown lines, not every alternative.
• Engine strength. Lines and scores are Stockfish's at 100k nodes: strong but not perfect.
• Sources and terms:
  • positions from the Lichess open database (CC0);
  • engine analysis by Stockfish 19 (GPL);
  • text written by openai/gpt-6-luna through OpenRouter. Check the model provider's terms for your use.