second-data-1.0 / README.md
joebos's picture
Upload folder using huggingface_hub
4065a5a verified
|
Raw History Blame Contribute Delete
8.21 kB
metadata
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - chess
  - stockfish
  - reinforcement-learning
  - grpo
  - sft
  - qwen3.5
pretty_name: Second Chess Explanation Dataset (second-data-1.0)
size_categories:
  - 10K<n<100K

Dataset Card for second-chess/second-data-1.0

1. Dataset Summary

second-data-1.0 provides the training, validation, test-pool, and reinforcement learning (RL) prompt splits for Second. Second is a family of small language models based on Qwen 3.5. These models turn Stockfish evaluations and board facts into clear human explanations.

Every record uses the Phase 1 prompt format from core/python and core/ts. Stockfish calculates all tactical and positional facts at fixed depth with MultiPV=3. The language model does not calculate variations on its own. Instead, the model explains the Stockfish facts for beginner (under 1200 Elo) and club (1200–2000 Elo) players. Explanations use short (2–4 sentences, up to 80 words) or long (80–200 words) detail.

2. Data Sources and Licenses

The pipeline follows the license rules in docs/data_sources.md:

Data Source License Redistribution Policy Public HF Export (data/processed/public/) Local Rebuild (data/processed/)
Lichess Open Game Database (data/sources/lichess_games.py) CC0 1.0 Universal Full public domain rights. Yes (rl_prompts.*) Yes (rl_prompts.*)
MATE Dataset (data/sources/mate.py, Wang et al., NAACL 2025) MIT License Permissive redistribution with MIT notice. Yes (sft_*.parquet, sft_*.jsonl) Yes (sft_*.parquet, sft_*.jsonl)
Lichess Studies (Open License) (data/sources/lichess_studies.py) CC0 / CC-BY / Public Domain Redistribute when an open license exists. Yes (open-license subset) Yes
Lichess Studies (Standard ToS) (data/sources/lichess_studies.py) Lichess Terms of Service Author holds copyright. Do not redistribute text. No (rebuild script only) Yes
GameKnot Annotated Games (data/sources/gameknot.py) Proprietary (GameKnot EULA) Do not redistribute commentary text. No (rebuild script only) Yes

MIT License Notice for MATE Dataset Portion

The supervised fine-tuning (SFT) splits include records from OutFlankShu/MATE_DATASET (Wang et al., NAACL 2025). The authors release the MATE dataset under the MIT License:

Copyright (c) 2024 OutFlankShu (MATE Dataset Authors) Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

3. Pipeline Filter Counts (Phase 2.3 & 2.4)

The collectors extracted 16,000 raw human commentary records from GameKnot (6,000), Lichess Studies (6,000), and the MATE dataset (4,000). Stockfish analyzed unique positions at depth=12 and MultiPV=3. The pipeline stored 47,876 unique FEN evaluations in data/cache/stockfish_eval_cache.parquet.

The filter module (data/pipeline/filter_comments.py) runs seven checks in order on each raw comment:

Filter Stage Description Items Removed Share of Raw Input
Total Raw Input Raw extracted (position, move, comment) records 16,000 100.0%
1. cleaned_empty_or_short Fewer than 8 words remain after stripping PGN tags, move numbers, and glyphs 4,165 26.0%
2. not_english langdetect (seed=0) identifies non-English text 134 0.8%
3. no_chess_content Text lacks piece names, square coordinates, SAN moves, or chess terms 931 5.8%
4. illegal_move_or_wrong_piece Text names an illegal SAN move or puts a piece on the wrong square (rewards/checks.py) 4,164 26.0%
5. tone_disagrees_with_verdict Praise appears on an error move, or criticism appears on a good move 191 1.2%
6. links_usernames_offensive Text contains URLs, @username handles, or profanity 69 0.4%
7. duplicate Text matches an earlier fingerprint or 4-gram shingle set (Jaccard >= 0.8) 152 1.0%
Total Kept Examples Clean human commentary (6,107 short + 87 long) 6,194 38.7%

4. Dataset Splits and Distributions (Phase 2.5 & 2.6)

4.1 SFT Splits (Split Strictly by Game)

The split module (data/pipeline/splits.py) groups records by the SHA-256 hash of game_id. If a 4-field normalized FEN in val or test_pool also appears in train, the split module removes that record (66 colliding opening records removed). Therefore, train, val, and test_pool share zero games and zero board positions.

Partition Total Records (Local Rebuild) Public Redistributable Records (public/)
sft_train 5,659 973
sft_val (~2.3%) 138 16
sft_test_pool (~5.4%) 331 86
Total SFT 6,128 1,075

4.2 RL Prompt Set (rl_prompts)

The Lichess collector (data/sources/lichess_games.py) sampled 25,000 positions from rated CC0 Lichess games. More than 80% of these games have an average player Elo between 800 and 2000. The RL prompt generator (data/pipeline/rl_prompts.py) removed all games and 4-field FENs that appear in sft_val or sft_test_pool. Stockfish analyzed the remaining positions at depth=12 and MultiPV=3 to build 20,000 RL prompts:

Category Count Share Target in PLAN.md
Total RL Prompts 20,000 100.0% 20k–50k
Task: explain_move (Inaccuracy / Mistake / Blunder) 9,000 45.0% ~45%
Task: explain_move (Good / Best / Only move) 7,000 35.0% ~35%
Task: explain_position 4,000 20.0% ~20%
Level: beginner 10,000 50.0% 50%
Level: club 10,000 50.0% 50%
Detail: short 16,000 80.0% 80%
Detail: long 4,000 20.0% 20%

5. Rebuilding the Full Dataset Locally

To rebuild the full dataset on your local machine or Google Colab, run these commands:

# 1. Collect raw sources
python -m data.sources.gameknot --output data/raw/gameknot.jsonl --max-items 6000
python -m data.sources.lichess_studies --output data/raw/lichess_studies.jsonl --max-items 6000 --include-all-studies
python -m data.sources.mate --output data/raw/mate.jsonl --max-items 4000
cat data/raw/gameknot.jsonl data/raw/lichess_studies.jsonl data/raw/mate.jsonl > data/raw/sft_combined.jsonl

# 2. Sample CC0 Lichess games for RL
python -m data.sources.lichess_games --output data/raw/lichess_games.jsonl --max-positions 25000

# 3. Run Stockfish analysis, filtering, splitting, and export
python -m data.pipeline.run_pipeline \
  --sft-input data/raw/sft_combined.jsonl \
  --rl-input data/raw/lichess_games.jsonl \
  --depth 12 \
  --workers 12 \
  --rl-target-count 20000

6. Known Limitations

  1. Human Commentary Grounding: Human authors wrote their comments without seeing our Stockfish Facts: lines. Phase 2.3 removes comments with illegal moves, wrong piece squares, or mismatched verdict tone. However, SFT mainly teaches explanation style and format. Phase 6 GRPO reinforcement learning enforces strict factual grounding.
  2. Long Commentary Scarcity: Most human move annotations are short (15–50 words). Only 1.4% of filtered human comments exceed 80 words (detail: long). To balance this, the RL prompt set includes 20% detail: long prompts (4,000 prompts). GRPO RL trains the long mode directly with the length reward check.
  3. Engine Depth: The bulk pipeline runs Stockfish 18 at depth=12 and MultiPV=3. It caches evaluations by FEN in Parquet format. If you run the pipeline with --depth 18, the cache updates the shallow entries automatically.