Pre-to-Post-2/stockfish_augment_opt_beam_SFT

Stockfish beam-search chess puzzle SFT data

Chain-of-thought SFT data for chess puzzles in which the thinking block is a beam search over the solver's moves, run with Stockfish. The opponent always plays Stockfish's best reply; the solver's moves branch. Every line that competed in the beam is written: lines the beam pruned show the failed move and the reply that refutes it, and the 4 final survivors are written to one shared horizon. Each record also carries the action-only (no CoT) target. Companion to Pre-to-Post-2/stockfish_augment_opt_type1_SFT (same puzzles, same prompts).

• 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, sfbeam_d16_mpv8_r0123_beam4_br2_steps2_seed0/datapuzzle_v4_shard{start}-{end}.json (482 MB)
• Engine: Stockfish, depth 16, single thread, hash cleared per puzzle (deterministic)

Layout

Location in sfbeam_d16_mpv8_r0123_beam4_br2_steps2_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_beam.cot_format_no_labels
CoT, each line tagged with Stockfish's bucketed eval	cot_by_method.sf_beam.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). 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, puzzle 9OVqg; solution Rf5 Rxf5 c8=Q):

<T> Ke2d2 Rc5xc7 <sep> Rf4f3 Rc5xc7 <sep> Rf4f7 Rc5c2+ Ke2d1 Rc2c6 Rf7f2 Rc6xc7 <sep> Rf4f7 Rc5c2+ Ke2d3 Rc2c6
<sep> Rf4f5 Rc5xf5 Pc7c8=R Pa4a3 <sep> Rf4f5 Rc5xf5 Pc7c8=Q Rf5c5 Qc8d7+ Kb5b4 <sep> Rf4f8 Rc5xc7 Ke2d2 Kb5b4
<sep> Rf4f8 Rc5xc7 Rf8b8+ Kb5a5 <sep> Ke2d3 Rc5xc7 <sep> Rf4f1 Rc5xc7 Ke2d2 Kb5b4 Rf1c1 Rc7h7
<sep> Rf4f1 Rc5xc7 Rf1b1+ Kb5a5 Ke2d2 Pa4a3 <sep> Ke2d1 Rc5xc7 </T> Rf4f5 <call_env> Rc5xf5 Pc7c8=Q <call_env>

Generation rule

1. Pool: Stockfish's top 8 first moves, each followed by Stockfish's best reply. The puzzle's own first move is always in the pool, with the puzzle's own reply.
2. Beam start: rank the pool by Stockfish eval; keep 4 (the puzzle line always survives). The other lines are written as pruned lines.
3. Expansion (at most 2 steps, only at solver moves the puzzle itself asks for): each surviving line's next solver move branches into Stockfish's top 2 moves, each followed by the best reply. Of the 8 children, 4 survive (the puzzle line always survives); the rest are written as pruned lines.
4. Horizon: the 4 survivors are extended with best play and cut at one shared horizon H = max(a + r, beam depth, L), where L is the puzzle length in plies, a = min(L, first ply at which the puzzle line is ahead on material of every other survivor at that ply and the next), and r is drawn from {0, 1, 2, 3}. If the puzzle line mates, all survivors are cut at the mating ply.
5. Order: lines are grouped by root move; root order and order within each root are shuffled per puzzle (seeded by puzzle index).
6. Check: a record is written only if its thinking contains the complete puzzle solution exactly once.

Lines per puzzle = 4 pruned roots + 4 per expansion step + 4 survivors: 8, 12 or 16 for 89% of puzzles. Other counts occur when a position has fewer than 8 legal moves or a line ends the game before a step.

Statistics

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

Lines	Puzzles
2	4
3	81
4	217
5	166
6	391
7	146
8	9,040
9	542
10	1,024
11	1,551
12	33,518
13	746
14	1,281
15	2,342
16	28,257

Mean thinking length: 51.6 moves. Puzzles with a mating solution (survivors cut at mate): 24,584. The answer's root group comes first in 13.3% of puzzles (chance 13.3%); the puzzle line's mean relative position in the thinking block is 0.50 (0 = first, 1 = last).

Hard puzzles (6,772 from sft_hard_2200_2499.csv; 2,885 with 1 expansion step, 3,887 with 2):

Lines	Puzzles
3	10
4	28
5	13
6	32
7	18
8	79
9	56
10	97
11	119
12	2,626
13	85
14	138
15	268
16	3,203

Mean thinking length: 60.6 moves. Puzzles with a mating solution: 1. The answer's root group comes first in 13.8% of puzzles (chance 13.7%); the puzzle line's mean relative position is 0.50.

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.
• Length cue (by design): outside mating puzzles the 4 survivors are the longest lines, so the answer is always among the 4 longest lines.
• Stockfish crash fallback: this Stockfish build occasionally crashes on a multi-PV search of specific positions; for those 14 base puzzles and 2 hard puzzles the moves at that node were ranked by searching each legal move separately (sf_meta.multipv_fallback = true).
• Metadata: sf_meta holds every line in UCI with its root, engine rank, eval and the beam step at which it was pruned (null for survivors), plus a, r, H and the beam depth.
• 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/ (sf_beam_generator.py imports helpers from sf_multipv_generator.py). The hard shards were made with DATA_PATH=.../sft_hard_2200_2499.csv INDEX_OFFSET=91619 bash submit_sharded_jobs_sf_beam.sh --submit.

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/; generators, sbatch, submitter and checker gained --index_offset / INDEX_OFFSET.
• 2026-10-22: first release (184 shards, 79,306 puzzles).