LegumMagister/chsm8-post-training

CHSM8 post-training: style labels (30M)

Five natural-language style descriptions (terse, terse, short, medium, long) for one side of one game, for 30.08M side-appearances of the CHSM8 pre-training corpus. Rows join to the games on (month, row_idx), the stable game id used by all CHSM8 datasets.

Exactly the labels chsm8-sft and chsm8 were trained on: every Apertus-70B label set is replaced by the clean distilled labeler (gemma-3-1b-distill-clean), and the data comes as the train / validation / test splits those models used (split by game, both sides of a game in the same split).

split	rows
train	27,076,507
validation	1,503,626
test	1,503,412

from datasets import load_dataset
val = load_dataset("LegumMagister/chsm8-post-training", split="validation")

Which games

A subset of the pre-training corpus: game sides where the described player is rated ≥ 2200 on Lichess (opponents ≥ 1800, the corpus floor; 35% of opponents are below 2200). All 162 months (2013-01 – 2026-06), both colours equally (15.04M white / 15.04M black). By speed: 75% blitz, 24% rapid, 1% classical (base + 40 × increment). Sides of games longer than CHSM8's 512-position context were removed (1,674 rows, 0.006%), leaving 30,083,545 rows.

Columns

column	meaning
month, row_idx	game id in chsm8-pre-training
side	white or black: the player described
labels	{terse_1, terse_2, short, medium, long}
labeler	who wrote the label set (see below)
correct	per-band mask: 1 = no claim in that band is contradicted by the measured card (rule check), 0 = contradicted
clean	all five bands correct

How the labels were made

labeler	meaning
gemma-4-31b	written by Gemma-4-31B
gemma-3-1b-distill-clean	an Apertus-v1.5-70B set relabelled from scratch by the clean distilled labeler (3,847,766 rows)
gemma-3-1b-distill	written by chsm8-labeler-gemma3-1b, distilled from Gemma-4-31B on chsm8-labeler-distill
*-patched	one band contradicted the card and was regenerated by the distilled labeler, conditioned on the game and the earlier bands

When two or more bands were contradicted, the whole set was regenerated by the distilled labeler (gemma-3-1b-distill).

A second pass then re-fixed every set that still had a contradicted band (1.97M sets) with the clean distilled labeler: up to 5 rounds, each regenerating the first still-contradicted band (8 candidates, first one passing the check is kept) and re-checking the set. Contradicted bands: 4.7% (distilled), 4.6% (Gemma), 16.5% (Apertus) before repair → 4,868 of about 150M bands (0.003%) after (before the Apertus relabel below). Those few are marked 0 in correct; mask or drop them.

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") labeler that labelled the corpus; the clean labeler contradicts the card 9× less.

Because of this, every Apertus-70B set was relabelled with the clean distilled labeler (100% parsed, 0.40% of bands contradicted vs 16.5% for raw Apertus). This release contains no Apertus labels. Labelers: gemma-3-1b-distill 19,273,741; gemma-3-1b-distill-patched 4,010,380; gemma-3-1b-distill-clean 3,847,766; gemma-4-31b 2,405,017; gemma-4-31b-patched 546,641. Contradicted bands overall: 78,472 of 150.4M (0.05%), in 75,366 rows (clean: false).

The model input

Every label was written from the same input: the side's moves plus a card of features measured from the board. scripts/build_prompts.py rebuilds that exact input from chsm8-pre-training:

pip install datasets python-chess
python scripts/build_prompts.py 2016-01 --split validation --limit 100   # -> prompts_2016-01.jsonl: labeler input + labels

Raw teacher labels

The unrepaired labels of both teacher LLMs (Gemma-4-31B, and Apertus-v1.5-70B in thinking and non-thinking mode) are in chsm8-labeler-distill, with fixed train / validation / test splits per teacher. Only the Gemma-4-31B labels were used to train the distilled labeler.

Labeling scripts

scripts/labeling/ holds the exact pipeline that wrote these labels:

• production_label.py + autolabel/: teacher labeling (Gemma-4-31B, Apertus-70B) with the s10-mini prompt (autolabel/genome.py holds the prompt, autolabel/corpus.py and features.py build the measured card). API keys are read from environment variables (autolabel/providers.py).
• build_unlabeled.py, label_shards.py: inputs for the remaining sides and labeling with the distilled labeler (chsm8-labeler-gemma3-1b) in vLLM; train.py holds the labeler's prompt format and output parser.
• grounding.py: the rule check behind correct; repair_labels.py, fix_todo.py, fix_loop.py: band repair; merge_final.py: merge; build_post_training_ds.py: this release.

Paths inside these scripts point to our cluster; adjust them before running. The training and inference code for the CHSM8 models will be released later; this repository contains the data and the scripts that built it.

License

Apache-2.0. The labels were written by Gemma-4-31B and Apertus-v1.5-70B (both Apache-2.0) and by chsm8-labeler-gemma3-1b, a fine-tune of Gemma 3 distributed under the Gemma Terms of Use (Google claims no rights in its outputs). The games come from the Lichess open database (CC0). You are welcome to use this dataset to train or evaluate your own models; please cite the CHSM8 technical report.

Acknowledgments

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