LegumMagister/chsm8-distillation

CHSM8 distillation set

The exact data the style labeler chsm8-labeler-gemma3-1b was trained on: 2,402,472 label sets written by Gemma-4-31B in which no band is contradicted by features measured from the board, each with the full input the teacher saw.

split	rows
train	2,402,472	Gemma-4-31B sets with clean: true (about 80% of its 3.02M sets)
validation	1,524	unfiltered
test	1,517	unfiltered

from datasets import load_dataset
ds = load_dataset("LegumMagister/chsm8-distillation")

All label sets, of both teachers and after every repair pass, are in chsm8-game-labels. The labels the chess models were trained on are in chsm8-post-training.

Format

One row per label set (one side of one game):

field	meaning
key	month:row_idx:side, unique id
month, row_idx	the game in chsm8-pre-training
side	white or black (the player described)
elo	the described player's rating
user	the exact teacher user message: moves plus a card of features measured from the board; 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: 1 = no claim in that band is contradicted by the card (rule check), 0 = contradicted
clean	all five bands correct (always true in train)

The student input is the game block of user (from You are looking at... to before Return ONLY) followed by ### STYLE LABELS; the target is the five labels, one per line.

Why only clean sets

About 4.6% of Gemma-4-31B's labels (in about 20% of its sets) contradict the measured card. A student trained on all 3.02M sets copied that rate (4.5%); trained on this clean subset, it contradicts the card in 0.5% of labels. "Contradicted" is a rule check (scripts/grounding.py) for five claim types: a sacrifice when there was none, a queen trade when the queens stayed on, a king attack with no king pressure, slow development when it was fast, and "never castle" when the side castled.

Splits

Fixed by game: 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 fall in the same split.

Label diversity: teacher vs distilled

Share of distinct label strings in random samples of label sets, per band. Dashed: Gemma-4-31B teacher (measured up to 1M), solid: the distilled labeler (measured up to 10M), circles: power-law projection to 30M, an upper bound. The student's short tags repeat more than the teacher's; medium and long descriptions stay nearly unique per game.

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.

Scripts

• system_prompt.txt: the teacher's system prompt.
• scripts/build_prompts.py (+ autolabel/): rebuilds the teacher's user message for any game side from chsm8-pre-training.
• 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: the export.

Paths inside these scripts point to our cluster; adjust them before running.

License

Apache-2.0, the license of the teacher Gemma-4-31B-it. The games come from the Lichess open database (CC0). Please cite the CHSM8 technical report.

Acknowledgments

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