cetusian/chess-sft-lichess-2200

Lichess Chess SFT (≥2200 Elo)

A 200,000-position supervised fine-tuning dataset for training language models to play chess by predicting the next move from a PGN prefix. Derived from strong-player Lichess games, stripped to a minimal input/output format.

Summary

• Task: given a partial game in PGN notation, predict the next move in Standard Algebraic Notation (SAN).
• Source: Lichess/standard-chess-games (all games downloadable from lichess.org; CC0).
• Strength filter: both players ≥ 2200 Elo.
• Size: 195,999 train · 4,000 validation · 199,999 total.
• One position per game — chosen uniformly at random over non-terminal plies — to prevent the model from memorising long chunks of individual games and to keep openings/middlegame/endgame all represented.

Fields

Each row is a single JSON object:

{"input": "1. e4 e5 2. Nf3 Nc6 3. Bb5 a6 4. Ba4", "output": "Nf6"}

Field	Type	Description
input	string	PGN move prefix in standard algebraic notation (e.g. "1. e4 e5 2. Nf3"). Always contains at least one move.
output	string	The next move in SAN (e.g. "Nf6", "O-O", "e8=Q", "Qe6+").

No system prompt, no role/chat wrappers, no <think> block, no Stockfish annotation, no board description — pure prefix → next move.

Construction

Games are streamed from Lichess/standard-chess-games and filtered:

1. Elo ≥ 2200 for both White and Black.
2. Termination == "Normal" (excludes time forfeits, abandons, rules-violation terminations).
3. Game length ≥ 20 ply (drops ultra-short games that finish in the opening).

For each game that survives, a single ply k is sampled uniformly from [0, len(game) − 2], the position is replayed with python-chess, and (PGN[:k], move[k]) is emitted. A row is dropped if:

• The position's FEN is already in the dataset (exact-duplicate de-dup).
• The first six plies of the game are already represented ≥ 300 times (opening cap — prevents the Sicilian et al. from dominating).
• Accepting it would push the White / Black count imbalance past ±500 (colour-balance cap).

After collection the rows are shuffled with a fixed seed, then split 98 / 2 into train / validation.

Statistics

Size & balance

Split	Rows	White	Black
train	195,999	97,776	98,223
validation	4,000	1,989	2,011

Ply depth of input

stat	min	p10	median	mean	p90	max
train	1	11	34	40.0	77	275

Median game position is mid-middlegame; the long tail reaches deep endgames.

Opening diversity

Top White first moves across train:

Move	Count
e4	74,825
d4	60,940
Nf3	19,828
c4	9,318
g3	7,957
b3	6,070
e3	4,806
d3	3,800

Output shape

	Count
Castles (O-O / O-O-O)	5,833
Checks (SAN ending in +)	10,952
Promotions (SAN containing =)	449
Mates (#)	0
Unique output tokens	2,382

Mate-in-one is absent because the one-position-per-game sampler picks a non-terminal ply; the final checkmating move is therefore never selected.

Label quality (Stockfish agreement)

We evaluated 2,000 random positions from the train split against Stockfish 16 at depth 18 (multi-PV = 3). The table shows how often the dataset's label — the move actually played by the human grandmaster — agrees with Stockfish's recommendation.

Metric	Value
Top-1 agreement (GM move == Stockfish best)	41.1%
Top-3 agreement (GM move in Stockfish top-3)	68.3%
Within 30cp of Stockfish's best	64.6%
Within 100cp of Stockfish's best	85.7%
Blunders (≥ 200cp loss)	7.5%
Median centipawn loss	13
p95 centipawn loss	298

Interpretation: labels are strong but imperfect — what you expect from 2200+ human play. The median position is within 13cp of the engine's top choice; the long tail of large cp losses consists mostly of lost tactical shots or missed mates.

Script: stockfish_probe.py in the companion repo.

Token lengths

Computed on the full 196k train split. input_tok is the PGN prefix; total is input_tok + output_tok.

Tokenizer	Vocab	Field	Mean	Median	p95	p99	Max	% > 512	% > 1024
Qwen3 / Qwen2.5	151,643	input	173.5	144	418	553	1,306	1.72%	0.00%
Qwen3 / Qwen2.5	151,643	output	2.7	3	4	4	6	0%	0%
Qwen3 / Qwen2.5	151,643	total	176.2	147	421	556	1,309	1.79%	0.00%
SmolLM2	49,152	total	181.5	151	433	573	1,357	2.15%	0.00%
Phi-4-mini	200,019	total	160.6	135	377	498	1,153	0.80%	0.00%

Recommendation: max_length=1024 covers the full dataset across all tokenizers tested (≤3 rows exceed it). max_length=512 covers ~98% — the 2% of truncated rows lose opening context but keep the position relevant to the labelled move, which is usually acceptable.

Usage

Load with 🤗 datasets

from datasets import load_dataset

ds = load_dataset("cetusian/chess-sft-lichess-2200")
print(ds)
# DatasetDict({
#     train: Dataset({features: ['input', 'output'], num_rows: 196000}),
#     validation: Dataset({features: ['input', 'output'], num_rows: 4000})
# })
print(ds["train"][0])
# {'input': '1. e4 e5 2. Nf3 Nc6 3. Bb5 Nge7 ...', 'output': 'Be6'}

Wrap into a chat template for SFT

Because this dataset is in a raw completion format, you'll typically wrap each row into your model's chat template before training. For a minimalist system prompt:

SYSTEM = (
    "You are a chess grandmaster. Given a game in PGN notation, play the "
    "best move for your colour. Respond with exactly one move in SAN."
)

def format_for_chat(row):
    user = row["input"] if row["input"] else "Start of game."
    return {
        "messages": [
            {"role": "system", "content": SYSTEM},
            {"role": "user", "content": user},
            {"role": "assistant", "content": row["output"]},
        ]
    }

ds = ds.map(format_for_chat, remove_columns=["input", "output"])

Or train directly on the raw input → output format (completion-style SFT) if you plan to serve with a text-completion interface.

Limitations & biases

• No Stockfish ground truth. The label is the move the human grandmaster played, which is not always the objectively best move. Strong but fallible.
• Move, not move + commentary. No reasoning traces or candidate-move evaluations. Suitable for behaviour cloning, not chain-of-thought.
• Opening bias. Despite the 300-per-opening cap, well-known lines still appear more often than rare ones.
• No mate labels (see above): the sampler never emits the final move of a game, so models trained solely on this dataset may under-produce #.
• Colour is implicit. Whose turn it is can be inferred from parity of the move count. The dataset does not include an explicit colour field.

Source & license

All games are sourced from the public Lichess database via the Lichess/standard-chess-games dataset, which Lichess releases under Creative Commons Zero v1.0 Universal. This derivative dataset is released under the same licence.

Citation

If you use this dataset, please credit Lichess:

@misc{lichess-database,
  title  = {Lichess Open Database},
  author = {{Lichess}},
  year   = {2024},
  url    = {https://database.lichess.org/},
  note   = {CC0 1.0 Universal}
}