Pre-to-Post-2/680M_alpha1.0_SFT_datasets

680M chess SFT CoT datasets (puzzle sequences)

Generator model: checkpoints/C_6p5e19/6p5e19_680m_alpha0.750/final (the 6.5e19-compute 680M pretrained model with alpha = 0.750). Every shard's config.model_path records this. The repo name says alpha1.0; the data itself comes from alpha 0.750.

Search-based chain-of-thought traces for chess puzzles, generated by sampling move trajectories from the pretrained model and labelling them with a Stockfish-backed minimax evaluation. Used as SFT data for the pre-to-post compute-allocation sweep.

Configurations

Each subfolder is one search budget (nc candidates, nt trajectories, nta trajectories for the answer move). Shards are 500 puzzles each, named datapuzzle_v4_shard<start>-<end>.json. Layout mirrors the local folder data/sft_ablation/cot_data_puzzle_seq_qwen/680m_ce19_alpha0_75/.

Subfolder	Search budget	Shards	Puzzles	Trajectories	GiB	Missing shards
hf_model_seed0_evalminimax_candpolicy_nc4_nt8_nta8_d20_dmrandom_wt300_pm8_inj	nc4_nt8_nta8	40	16,378	340,307	0.54	none
hf_model_seed0_evalminimax_candpolicy_nc8_nt10_nta10_d20_dmrandom_wt300_pm8_inj	nc8_nt10_nta10	82	31,786	1,452,538	1.69	24500-25000, 25500-26000, 36000-36500, 36500-37000, 37000-37500

Shared settings: source puzzles training_splits_v4/sft_dataset.csv, policy candidate selection, DFS-policy traversal, trajectory depth 20 (random depth mode), minimax evaluation with Stockfish 0.1 s and a 300 cp win threshold, answer injection on, max 8 puzzle plies at generation time (the answer tails have since been restored to the full solution; see Changelog), temperature 1.0, seed 0.

File format

Each shard is a JSON object with config (generation arguments), summary (per-shard counts, leaf win/draw/loss rates, ground-truth match rate) and results (500 records). Record fields include pgn, moves_list, opponent_move, ground_truth_move, target_move, target_move_san, is_match, best_val, minimax_values and the rendered CoT formats under cot_by_method.

Use in SFT

The chess SFT trainer reads the sorted shard list, caps it at max_train_files, and holds out the last 2 files in sorted order for eval. Note the sort is lexicographic, so the holdout is shard9000-9500 and shard9500-10000.

Changelog

2026-10-20: full solutions restored in the answer tails. The data was generated with max_puzzle_moves: 8, which the generator counts in plies, so every answer (the text after </T>) stopped after the 4th solver move even when the Lichess solution has 5 or 6. The tails were rebuilt from each record's moves_list (the full, untruncated solution) in every multi-step CoT field (cot_format, cot_format_with_context, cot_format_no_labels, cot_format_with_context_no_labels, puzzle_sequence). The <T> ... </T> traces, the first-move fields and all other record fields are unchanged, and each shard's config.solution_tail_patch notes the change.

Subfolder	Records changed	5-move	6-move
..._nc4_nt8_nta8_..._pm8_inj	115 of 16,378	104	11
..._nc8_nt10_nta10_..._pm8_inj	495 of 31,786	414	81

The previous (truncated) shards are at commit 42f95fbdd3e3aca7802c56b5513d6c6fd7d945db.