Pre-to-Post-2/stockfish_augment_opt_type1_SFT

Stockfish-augmented chess puzzle SFT data (option type 1)

Chain-of-thought SFT data for chess puzzles in which the thinking block lists the puzzle's solution line together with a few Stockfish alternative lines, in shuffled order, followed by the complete puzzle solution. Each record also carries the action-only (no CoT) target for the same puzzle.

• Puzzles: 86,078 in three groups (see Layout): 77,717 base puzzles that match the 200M SFT data, 6,772 hard puzzles (Elo 2200–2499), and 1,589 base puzzles the 200M data lacks
• Files: 198 shards of up to 500 CSV rows, sf_d16_mpv8_tempt3_max5_min2_r0123_cutrandom_seed0/datapuzzle_v4_shard{start}-{end}.json (392 MB)
• Engine: Stockfish, depth 16, single thread, top-8 lines, hash cleared per puzzle (deterministic)

Layout

Location in sf_d16_mpv8_tempt3_max5_min2_r0123_cutrandom_seed0/	Shards	Puzzles	Source CSV (training_splits_v4/)	Ratings
top level, shard0-500 … shard91000-91500	180	77,717	sft_dataset.csv	800–2199
top level, shard91619-92119 … shard98119-98391	14	6,772	sft_hard_2200_2499.csv	2200–2499
not_in_200M_data/	4	1,589	sft_dataset.csv	800–2199

• Base shards are named by CSV row range and hold exactly the same puzzles, shard for shard, as the 200M SFT data (Pre-to-Post-2/200M_SFT_dataset).
• Hard shards continue the row numbering: row k of sft_hard_2200_2499.csv has index 91,619 + k (91,619 = the number of rows in sft_dataset.csv). The hard puzzles share no PuzzleId, position or source game with the SFT, RL or test splits; 6,772 is 20% of the 33,858 puzzles in the 85-file SFT set. The CSV itself is not in this repo.
• not_in_200M_data/ holds shards 14500-15000, 24500-25000, 81500-82000 and 91500-91619, which the 200M data never generated. They sit outside the top level so the trainer's file selection matches the 200M runs (see File selection below).

Fields used for training

Format	Field (cot_field)
CoT, no labels (main)	cot_by_method.sf_multipv.cot_format_no_labels
CoT, each line tagged with Stockfish's bucketed eval	cot_by_method.sf_multipv.cot_format
Action-only (no CoT)	cot_by_method.solution_continuation.cot_format

The prompt is pgn (the game up to and including the opponent's trigger move), identical to the existing puzzle-sequence SFT data. Both targets contain the complete puzzle solution in multi-turn form (move <call_env> reply move <call_env> …); there is no 4-move cap.

Example (cot_format_no_labels):

<T> Pa2a3 Bb4d6 Ra1e1 <sep> Ra1e1 Pe6e5 Pc2c3 <sep> Bc4xe6 Nf6e4 Ng5xe4 <sep> Pc2c3 Bb4d6 Bf4xd6
<sep> Qf3xc6+ Pb7xc6 Bc4a6# <sep> Ng5xe6 Nf6e4 Ne6xd8 </T> Qf3xc6+ <call_env> Pb7xc6 Bc4a6# <call_env>

Generation rule

1. Solution line: the puzzle's own moves, extended with Stockfish's best continuation where more plies are needed.
2. Alternatives: Stockfish's top-8 lines minus the solution move.
3. Tempting alternatives: b_j = first ply at which the solution line is ahead of alternative j on material (solver's view, mate = ±100) at that ply and the next. Alternatives with b_j ≥ 3 (or never refuted) are tempting. Keep up to 5 tempting alternatives in engine order; if fewer than 2, top up with the best-ranked others.
4. Cut points: a = min(puzzle solution length in plies, first ply at which the solution stably leads every chosen alternative). The solution line is cut at a + r and each alternative at a + its own r, with r drawn uniformly from {0, 1, 2, 3} and limited to how far the line can continue. If the solution line reaches mate, every line is cut at the same ply, so the mate line is never the odd-length one.
5. Order: lines are shuffled per puzzle (seeded by puzzle index).

Checks on the generated data: a length-only guesser ("pick the shortest/longest line") does no better than chance; the solution's position among the lines is uniform; every record contains the complete puzzle.

Statistics

Base puzzles (all 79,306 from sft_dataset.csv, including not_in_200M_data/). Lines per puzzle:

Lines	Puzzles
2	562
3	30,130
4	5,286
5	5,382
6	37,946

Puzzle solution length (plies, solver + replies):

Plies	Puzzles
1	8,552
3	36,664
5	26,956
7	5,756
9	1,128
11	250

Mean cut ply of the solution line: 4.42. Mean thinking length: 21.8 moves. Puzzles with a mating solution (shared cut): 23,073.

Hard puzzles (6,772 from sft_hard_2200_2499.csv):

Lines	Puzzles
2	83
3	2,018
4	523
5	545
6	3,603

Plies	Puzzles
3	2,885
5	3,135
7	719
9	28
11	5

Mean cut ply of the solution line: 5.29. Mean thinking length: 26.3 moves. Puzzles with a mating solution: 1.

Notes

• Skipped rows: 12,300 CSV rows have no ctx (game record) and 13 have a trigger move that is illegal in their game record; both are skipped, as in the existing SFT data. No hard-CSV row was skipped.
• Stockfish crash fallback: this Stockfish build occasionally crashes on a top-8 search of specific positions. For those 10 base puzzles and 1 hard puzzle the alternatives come from searching each legal move separately (sf_meta.multipv_fallback = true). 53 shards were written before this fallback existed; they contain no crashed puzzles but lack the multipv_fallback and summary.error_counts fields.
• Metadata: sf_meta holds every line in UCI, its engine rank and eval, the cut points (a, r, H, per-line lengths), and each alternative's b_j and tempting flag.
• File selection: the SFT trainer reads only the top level of the folder, sorts file names as text, keeps the first max_train_files and holds out the last 2 of those for eval. With this layout max_train_files=85 selects exactly the 85 files the 200M runs used (32,858 training puzzles; shard48000-48500 and shard48500-49000 held out). In text order the hard shards fill places 180–194 of the 194 top-level files, except that the base shard shard9500-10000 sorts between shard94619-95119 and shard95119-95619. So the 85-file selection never includes them, and with no cap the last two hard shards would become the eval holdout. To train on the 85 files plus the hard shards, pass an explicit file list (select hard shards by their start index, 91619 and up, not by sort position).

Generation code is in code/. The hard shards were made with DATA_PATH=.../sft_hard_2200_2499.csv INDEX_OFFSET=91619 bash submit_sharded_jobs_sf_multipv.sh --submit; check_sf_multipv_shards.py checks a shard folder for missing or incomplete shards.

Changelog

• 2026-10-23: added the 14 hard shards (6,772 puzzles); moved the 4 shards the 200M data lacks into not_in_200M_data/; generator, sbatch and submitter gained --index_offset / INDEX_OFFSET.
• 2026-10-22: first release (184 shards, 79,306 puzzles).