CHSM8-SFT: chess moves conditioned on a natural-language style description

chsm8-pt fine-tuned to play one side of a game in the style a text describes (e.g. "trade everything and grind the endgame") or like a named top player ("play like Carlsen").

Highlights

Conditioning works: the right description makes the described side's moves more predictable. Test set (8,000 games), cross-entropy on the described side's moves:

	right description	wrong description (another game's)	no description
all 9 labels	2.880	2.929	2.905
5 LLM style descriptions only	2.855
player prompts ("play like {player}", val)	2.977	2.980	2.987

Style descriptions carry most of the signal (about 0.05 nats below no text); player prompts help only slightly so far.

Prompts change how it plays (model's own moves vs Stockfish 2000, no prompt in brackets):

• "leave your king in the center and go for an open fight": castles in 45% of games [94%]
• "castle queenside and storm the enemy king with your pawns": 39% of castling is queenside [10%]
• "trade queens as early as possible": queens off by move 20 in 40% of games [25%]
• "push your pawns aggressively up the board and grab space": pawn moves 0.289 per move [0.248]
• "bring your queen out early and push random pawns for no reason": early queen moves 16% [7%], fewer Sicilians
• "attack the king at all costs...": most checks per move (0.082 [0.072]), shorter games (89 plies [98])
• every prompt plays more varied openings than no prompt (0.77–0.96 distinct lines per game [0.75])

Prompts don't change how well it plays. No prompt: 1828 Elo (1808–1847, 1,000 games). Strong prompts ("a 2800-rated super grandmaster...") 1765–1836 and weak prompts ("you just learned how the pieces move") 1778–1839 overlap within the confidence intervals: every training game came from a ≥ 2200 player, so no description was ever paired with weak play. Prompts that push risky behaviour cost strength (king in the center 1735, "1 minute" 1756).

Data

• chsm8-post-training: 27.08M train / 1.50M val / 1.50M test game sides (described player rated ≥ 2200), each with 5 LLM style descriptions (weight 1.0) and 4 metadata descriptions (rating, time control, opening, trading; weight 0.05).
• 542,109 train games of 99 FIDE top-100 players (chsm8-top-players) with one "play like {player}" prompt each (49 templates × aliases); about 1% of games for the first 80% of training, ramping to about 9.5% at the end.
• One pass, every training game exactly once, order random across years.

Conditioning

• The description is encoded by a frozen google/embeddinggemma-300m (up to 128 tokens). All its token states, projected to the decoder width, are the keys and values of a cross-attention in the top 6 of 12 decoder layers.
• Each cross-attention output is added as x + tanh(gate) · attention, gates initialised at 0 (training starts exactly at the pretrained model). With no description the model is the pretrained policy.
• Loss: next-move cross-entropy on the described side's moves only; the opponent's moves are context.

Training

• Length-bucketed batches: up to 128 games per GPU, games × 5 descriptions × longest game ≤ 98,304 tokens. 2 nodes × 4 GH200 (8 GPUs, DDP), about 2,400 games/s, peak 53 GB per GPU, 3.1 h of training.

Final metrics (cross-entropy on the described side's moves)

matched = the game's own description, mismatched = another game's description, zero = no text; p* = player prompts.

metric	value
pval_games	8000.0000
pval_matched	2.9782
pval_matched_ce	2.9772
pval_matched_ce_llm	0.0000
pval_matched_ce_meta	2.9772
pval_matched_minus_mismatched	-0.0022
pval_matched_minus_zero	-0.0092
pval_mismatched_ce	2.9795
pval_zero_ce	2.9865
test_games	8000.0000
test_matched	2.8802
test_matched_ce	2.8797
test_matched_ce_llm	2.8546
test_matched_ce_meta	2.9110
test_matched_minus_mismatched	-0.0488
test_matched_minus_zero	-0.0251
test_mismatched_ce	2.9285
test_zero_ce	2.9048
val_every	1000.0000
val_games	8000.0000
val_matched	2.8776
val_matched_ce	2.8770
val_matched_ce_llm	2.8532
val_matched_ce_meta	2.9069
val_matched_minus_mismatched	-0.0477
val_matched_minus_zero	-0.0239
val_mismatched_ce	2.9248
val_zero_ce	2.9010

Post-SFT eval

vs Stockfish 17.1 UCI_Elo 2000, 0.05 s/move, greedy legal moves. Behaviour is measured on the model's own moves.

group	prompt	games	Elo (95% CI)	game length (plies)	captures / own move	checks / own move	pawn moves / own move	queens traded by ply 40	castled	castled queenside (of castled)	queen moved in first 5 own moves	distinct first-8-own-move lines / game	Sicilian as Black (1.e4 c5)	mean centipawn loss
strong	You are Magnus Carlsen in his prime: squeeze every tiny edge until it becomes a win	300	1836 (1797–1871)	101	0.219	0.068	0.246	23%	93%	16%	9%	0.83	100%	200
strong	a 2800-rated super grandmaster with flawless technique	300	1829 (1792–1862)	98	0.216	0.070	0.242	17%	95%	8%	5%	0.88	100%	199
strong	think like an engine: precise, accurate and ruthless	300	1827 (1789–1862)	95	0.218	0.072	0.227	21%	96%	6%	4%	0.78	100%	186
strong	play like a 2750 player against a similar-rated player	299	1818 (1779–1852)	98	0.213	0.068	0.248	18%	94%	12%	7%	0.82	100%	206
strong	mature, patient, grandmaster-level positional chess	299	1813 (1776–1847)	102	0.208	0.054	0.245	16%	96%	8%	7%	0.80	100%	201
strong	calculate deeply, never blunder, and convert every advantage cleanly	300	1809 (1770–1845)	97	0.231	0.078	0.246	29%	95%	12%	10%	0.90	100%	194
strong	play like a strong GM against an opponent of the same strength	299	1791 (1749–1828)	98	0.215	0.070	0.244	24%	95%	8%	7%	0.82	100%	231
strong	play the way a reigning world champion would in a title match	300	1765 (1721–1803)	96	0.211	0.060	0.247	15%	95%	11%	5%	0.77	100%	210
weak	play like an 800 rated player against a much stronger opponent	300	1839 (1802–1873)	97	0.220	0.064	0.245	18%	96%	9%	5%	0.82	100%	172
weak	play badly on purpose	300	1827 (1789–1862)	98	0.221	0.063	0.252	20%	93%	8%	5%	0.84	100%	194
weak	a careless beginner who leaves pieces hanging every few moves	300	1817 (1779–1851)	96	0.217	0.060	0.248	27%	93%	8%	5%	0.88	100%	221
weak	play like a 1900 player against someone better than you	300	1817 (1777–1853)	98	0.215	0.064	0.242	27%	95%	15%	8%	0.89	100%	197
weak	you just learned how the pieces move	300	1814 (1775–1849)	95	0.221	0.069	0.244	21%	95%	8%	6%	0.88	100%	216
weak	move instantly without thinking and ignore every threat	300	1792 (1753–1826)	98	0.219	0.061	0.239	25%	95%	8%	5%	0.81	100%	204
weak	a distracted casual player blundering in a noisy bar	300	1790 (1750–1825)	94	0.220	0.057	0.242	23%	96%	9%	5%	0.79	100%	207
weak	bring your queen out early and push random pawns for no reason	300	1778 (1736–1816)	90	0.234	0.070	0.237	23%	85%	22%	16%	0.95	0%	247
style	keep the position full and avoid every exchange	200	1847 (1800–1888)	97	0.225	0.075	0.243	20%	96%	7%	6%	0.86	100%	202
style	play like it's a long classical game with plenty of time	200	1834 (1786–1876)	100	0.219	0.073	0.254	21%	96%	11%	4%	0.81	100%	213
style	grind out a long endgame, play on for as long as it takes	199	1813 (1764–1855)	102	0.218	0.059	0.256	24%	97%	8%	7%	0.90	100%	205
style	keep the queens on the board no matter what	200	1811 (1764–1853)	99	0.220	0.064	0.244	28%	94%	13%	5%	0.90	100%	204
style	give lots of checks and keep harassing the enemy king	200	1809 (1762–1850)	97	0.221	0.076	0.242	28%	96%	9%	9%	0.91	100%	209
style	trade everything you can and simplify into an endgame	200	1804 (1759–1844)	97	0.219	0.060	0.245	32%	95%	15%	8%	0.91	100%	203
style	castle queenside and storm the enemy king with your pawns	200	1800 (1748–1844)	93	0.222	0.067	0.276	20%	92%	39%	6%	0.84	100%	225
style	go for a short, sharp game and finish it fast	200	1785 (1736–1828)	99	0.224	0.071	0.238	22%	94%	6%	2%	0.92	0%	204
style	go for the Sicilian Defense	199	1781 (1728–1827)	96	0.219	0.071	0.244	28%	92%	12%	6%	0.87	100%	230
style	attack the king at all costs and sacrifice material for the initiative	200	1780 (1724–1828)	89	0.232	0.082	0.251	17%	95%	11%	8%	0.80	100%	211
style	solid, safe, risk-free chess: wait for your opponent to make mistakes	200	1780 (1729–1824)	100	0.216	0.060	0.243	24%	96%	14%	10%	0.90	100%	222
style	push your pawns aggressively up the board and grab space	200	1775 (1725–1818)	93	0.223	0.055	0.289	20%	90%	9%	14%	0.92	100%	249
style	trade queens as early as possible	200	1772 (1722–1816)	102	0.214	0.061	0.246	40%	92%	22%	8%	0.92	100%	196
style	you have 1 minute, play like it	200	1756 (1701–1803)	94	0.215	0.063	0.249	20%	94%	11%	4%	0.82	100%	219
style	leave your king in the center and go for an open fight	199	1735 (1676–1783)	86	0.215	0.050	0.268	18%	45%	34%	5%	0.96	100%	254
none	(no prompt)	1000	1828 (1808–1847)	98	0.221	0.072	0.248	25%	94%	10%	7%	0.75	100%	196

Limitations

Research checkpoint, no search. The model imitates how ≥ 2200 Lichess players and top players played; prompts describing much weaker play are outside the training data.