LegumMagister/chsm8-game-labels

CHSM8 game labels: style labels from two teacher LLMs

Label sets for one side of one Lichess game: five natural-language style descriptions of how that player played. Every label set the two teacher LLMs wrote, raw and after our repair passes, so that every stage of the labelling can be compared and reproduced. Only the clean part of the raw gemma labels (2.40M sets) trained the distilled labeler (chsm8-labeler-gemma3-1b); that exact training set is chsm8-distillation. The labels the chess models were trained on are in chsm8-post-training.

subset	written by	label sets	contradicted bands
gemma (default)	Gemma-4-31B-it, raw	3,021,638	4.6%
gemma-patched	the same game sides after our repair passes	3,021,418	0.004%
apertus	Apertus-v1.5-70B, raw, thinking on and off	4,753,034	16.5%
apertus-patched	the same game sides after our repair passes	4,752,759	0.001%
apertus-distill	the same game sides relabelled from scratch by the clean distilled labeler	4,753,034	0.40%

"Contradicted" means a claim in the band is contradicted by features measured from the board (correct, see Format).

from datasets import load_dataset
gemma = load_dataset("LegumMagister/chsm8-game-labels")                     # subset "gemma"
student = load_dataset("LegumMagister/chsm8-game-labels", "apertus-distill")
gemma["train"], gemma["validation"], gemma["test"]

The gemma rows also contain the exact input the teacher saw (moves plus a card of features measured from the board).

Why these labels (out of the 7.77M we had)

Our label pool had 7.77M label sets from two teachers: 4.75M from Apertus-v1.5-70B and 3.02M from Gemma-4-31B. Apertus was chosen originally because its labels read well: the voice was vivid and varied. So our first training set mixed both teachers. Measuring grounding showed that Apertus invents events heavily, so the labeler was distilled from Gemma only. Both teachers' labels are published here so the comparison can be reproduced:

• Apertus contradicts the game. On the same games, 18.2% of Apertus labels contradict the measured card, vs about 4.6% for Gemma. Its most common error is describing sacrifices that never happened (in 58 of 76 audited games).
• Students inherit it. Students trained on the mixed pool contradicted the card in 14–19% of labels. The same students trained on Gemma-only labels dropped to about 4–5%, the teacher's own level. Validation loss also fell from 0.336 to 0.143 nats/byte, because one consistent voice is easier to learn.
• Filtering Apertus did not pay. A "clean mix" (Gemma plus the 42% of Apertus label sets that pass the automatic grounding check) trained to a worse validation loss (0.269 vs 0.143) with no measurable grounding gain. The automatic check covers only 5 claim types, so Apertus errors outside them still get through.
• Gemma is not perfect either. About 4.6% of its labels (in about 20% of games) contradict the card. These are kept in this dataset as the teacher wrote them.
• Records with any missing band, duplicates, or unparseable outputs were dropped before this (the pool counts above are after that cleaning).

Does distillation keep the signal? (downstream test)

We trained the same style-conditioned chess model (CHSM8 decoder + frozen Qwen3-Embedding-0.6B text encoder) on 1.75M games per labeler, using only that labeler's own, never-repaired label sets with no band contradicted by the card, and tested on 4,000 held-out games per labeler. Metric: cross-entropy on the described player's moves with the right label minus with another game's label (more negative = the label says more about how that game was played). Compare within a column (same test games):

trained on \ tested on	Gemma-4-31B labels	Apertus-70B labels	distilled Gemma-3-1B labels
Gemma-4-31B (teacher, 31B)	−0.061	−0.015	−0.068
Apertus-v1.5-70B	−0.014	−0.013	−0.015
distilled Gemma-3-1B (student)	−0.062	−0.017	−0.072

• Distillation keeps the teacher's signal. On the teacher's own test labels, a model trained on the 1B student's labels does as well as one trained on the 31B teacher's (−0.062 vs −0.061), and vice versa on the student's labels (−0.072 vs −0.068). The student decodes greedily, so it writes the teacher's most likely description without the teacher's sampling noise, at about 80× the per-GPU throughput.
• Apertus-70B labels carry 4–5× less signal, in every length band and in both directions, despite reading well.
• The student here is the first ("dirty") version that labelled the corpus; the clean version contradicts the card 9× less.

Format

gemma: parquet files in data/ (train, validation, test); one row per label set:

field	meaning
key	month:row_idx:side, unique id
month, row_idx	position in our monthly Lichess game shards
side	white or black (the player described)
elo	the described player's rating
user	the exact teacher user message (the game block is inside it); the system message is the same for every row and is in system_prompt.txt
labels	{terse_1, terse_2, short, medium, long}
correct	per-band mask in band order, e.g. [1, 1, 0, 1, 1]: 1 = no claim in that band is contradicted by the measured card (rule check), 0 = contradicted
clean	true if all five bands are correct (about 80% of records); the clean model was trained only on clean: true rows

The other subsets have key, month, row_idx, side and labels, plus:

• apertus: thinking, whether Apertus ran with thinking on (2,195,200 sets) or off (2,557,834). Different serving endpoints of the same model are merged.
• gemma-patched, apertus-patched: labeler and the correct / clean mask. Repairs regenerated contradicted bands with the distilled labeler, conditioned on the game and the other bands (*-patched); sets with two or more contradicted bands were regenerated whole (gemma-3-1b-distill: 69,760 Gemma and 904,993 Apertus sides). A second pass re-fixed what was still contradicted. These are the labels of the first chsm8-post-training release; a few hundred sides that are not in chsm8-post-training are missing.
• apertus-distill: the correct / clean mask. 1.93% of sets have at least one contradicted band.

These rows do not store the labeler input; scripts/build_prompts.py rebuilds it for any game side (see below).

The student input is the game block of user (from You are looking at... to before Return ONLY) followed by ### STYLE LABELS.

Two models trained on this set

• dirty (the dirty/ folder of the model repo): trained on 100% of the training split, flawed rows included. It reproduces the teacher's error rate (about 4.5% of labels contradicted).
• clean (the default model): trained only on rows with clean: true. This is the recommended model. Both are kept so that results can be compared or reproduced.

Splits

Fixed splits by game, the same rule for both teachers: bucket = blake2b("{month}|{row_idx}", 8 bytes) mod 10000; bucket < 5 is validation, 5 to 9 is test, the rest is train. Both sides of a game always fall in the same split, and a game is in the same split in both subsets, so results stay comparable across teachers and between papers. (A random split at load time, such as train_test_split, would put the two sides of one game in different splits.)

split	gemma	gemma-patched	apertus	apertus-patched	apertus-distill
train	3,018,597	3,018,377	4,748,318	4,748,043	4,748,318
validation	1,524	1,524	2,327	2,327	2,327
test	1,517	1,517	2,389	2,389	2,389

Label diversity: teacher vs distilled

Share of distinct label strings in random samples of 10k to 10M label sets, per band. Dashed = Gemma-4-31B teacher (measured up to 1M of its 2.4M sets in the 30M training labels), solid = distilled model (measured up to 10M), hollow dots = projection to 30M from a power-law fit (unique = k·n^β) on the points from 100k up. Short tags keep repeating as the corpus grows, so the curves bend down; the projection is an upper bound.

Band	Teacher at 1M	Distilled at 1M	Distilled at 10M	Teacher at 30M (projected)	Distilled, all 19.3M (actual)	All 30.1M labels (actual)
terse_1	15.5%	6.2%	1.9%	5.1%	1.4% (266k distinct)	9.8% (2.94M distinct)
terse_2	22.9%	9.9%	3.4%	8.8%	2.5% (482k)	11.7% (3.51M)
short	86.1%	57.5%	35.0%	74%	29.4%	40.8%
medium	99.8%	94.0%	83.7%	99.5%	79.5%	83.2%
long	100%	99.5%	98.1%	100%	97.4%	98.0%

The last two columns are exact counts, not projections. "Distilled, all 19.3M" is every set the distilled labeler wrote for the 30M training labels; "All 30.1M" is the whole label pool the chess models were trained on (chsm8-post-training): 19.3M distilled, 4.0M distilled-repaired, 2.4M + 0.5M Gemma-4-31B (raw + repaired) and 1.98M + 1.87M Apertus-v1.5-70B (raw + repaired). The teacher's own 2.4M sets in that pool are 11.2% / 17.1% / 80.8% / 99.6% / 100% distinct. Measured at 19.3M, the distilled shares are already close to the 30M projection (1.2% / 2.2% / 30% / 82% / 98%) and below it for short to long, consistent with the projection being an upper bound.

The distilled labeler writes less varied text than its teacher, as expected from a 1B model fine-tuned on a narrow 3M-example set: its short tags behave like a vocabulary of a few hundred thousand style phrases, while the long descriptions stay near-unique per game. Label lengths are unchanged. For style conditioning this is acceptable: different games still get descriptions that match them. Sampling at a moderate temperature (for example 0.7) recovers some variety.

Prompts and scripts

• system_prompt.txt: the teacher's system prompt; the user column holds the full user message the teacher saw.
• scripts/build_prompts.py (+ autolabel/): rebuilds that user message for any game side of chsm8-post-training from the games in chsm8-pre-training:

  pip install datasets python-chess
  python scripts/build_prompts.py 2016-01 --split validation --limit 100
  

• scripts/production_label.py: the teacher labeling run; build_data.py: joins labels with the teacher's input; grounding.py: the rule check behind correct; filter_clean.py: the clean subset; export_distill.py: this release.

Paths inside these scripts point to our cluster; adjust them before running. Inference code and demos: github.com/nov3o/llm_chess-demo; the training code will be released later.

License

Apache-2.0, the license of both teachers: Gemma-4-31B-it (Gemma 4 license, Apache-2.0) and Apertus-v1.5-70B (Apache-2.0). The games come from the Lichess open database (CC0). You are welcome to use these labels to train or evaluate your own models; please cite the CHSM8 technical report.

Source

Games: chsm8-pre-training (Lichess open database, rated, blitz and slower), described player rated ≥ 2200. Labels: Gemma-4-31B (instruction-tuned, thinking off) and Apertus-v1.5-70B.

Acknowledgments

Thanks to Vladimir Melnikov for the idea of a distilled labeler.