adamtwiss/incipit-data

Incipit self-play training data

Self-play games played by Incipit, a chess engine with NNUE evaluation, and used to train its networks. About 7.1 billion scored positions from about 66 million games, in viriformat.

Every game was played by Incipit itself. The first generation was bootstrapped from a hand-crafted material-and-mobility evaluation (no other engine's networks, evaluations or data were used). Each later generation was played by the best network trained on the generations before it.

Format

Each file (<set>/<set>_NN.vf) is a sequence of games in viriformat:

• a 32-byte marlinformat board for the game's start position (castling rooks are marked as unmoved rooks, so Chess960/DFRC starts are representable), carrying the game result;
• one 4-byte record per move: a 16-bit move (from | to << 6 | promo << 12 | type << 14, type 0 normal, 1 en passant, 2 castling written king-takes-rook, 3 promotion) and the 16-bit search score of the position before the move, white-relative, in centipawn-like units; i16::MAX (32767) means the position has no usable score (mate found, or unscored);
• a 4-byte zero terminator.

Readers such as the viriformat Rust crate and Bullet's viriformat loader read these files directly.

How the data was generated

• Games on an OpenBench instance with fastchess, both sides the same Incipit build and network. Fixed nodes per move: 15,000 unless noted (5,000 for e12).
• Openings: Incipit's genfens (8-9 random plies from the start position, or from a random DFRC start for the d* sets), kept only if the position's search score is within 500.
• PGNs converted to viriformat with Incipit's datatools pgn; every game converted (no skips) unless noted below.

Sets

cN is generation N of standard chess. A letter suffix (c4b, c5a, ...) marks a run stopped early, usually because a stronger network arrived. dN are DFRC (double Fischer random) sets. e12 was a search-depth experiment. "Played by" is the network (by its file-hash prefix) that played the games.

Set	Played by	Games	Positions	Notes
c1	material+mobility bootstrap eval	~170k	16.7M	older 5k-node datagen; converted from a legacy per-position format by rebuilding game chains (98.8% of records kept)
c2	first trained net		34.3M	as c1
c3	second trained net		45.3M	as c1
c4	C3	1.51M	138.2M
c4b	C3	231k	21.2M	stopped early
c5a	F0359B05	233k	21.5M	stopped early
c5	765C79CD	1.51M	140.5M
c6a	F8CA057A	462k	43.2M	stopped early
c6	7F72F973	1.51M	140.3M
c7a	2CBCC7E0	505k	47.2M	stopped early
c7	849C3939	1.51M	139.4M
c8	849C3939	1.50M	138.1M
c8b	5448A4FD	1.50M	137.8M
c9a	4C56329E	395k	36.4M	stopped early
c9	39D2FF42	1.50M	135.4M
c10	3638FF8F	1.50M	137.4M
c11	DF8B55A5	1.50M	139.6M
c12	BE9A5723	1.50M	138.2M
c13	85B5CDD6	1.50M	138.4M
c14	A617C0B4	1.50M	138.9M
c15	A617C0B4	1.50M	139.6M
c16	43F3DD60	1.50M	139.5M
c16t	43F3DD60	1.09M	134.3M	decisive games played out (no win adjudication), 5-man tablebase adjudication
c17	AE339E48	792k	72.4M	stopped early
c17t	AE339E48	1.50M	184.1M	decisive games played out, 5-man tablebase adjudication
c18a	4A59F07A	2.01M	241.9M	8,000 nodes per move; first part of c18
c18b	4A59F07A	1.62M	194.3M	8,000 nodes per move; rest of c18
c19	3AB5F594	1.51M	181.2M	8,000 nodes per move
c20	3AB5F594	1.50M	180.6M	8,000 nodes per move
c21	ABE5E000	1.50M	180.6M	8,000 nodes per move
c22	ABE5E000	4.00M	481.2M	8,000 nodes per move
c23	ABE5E000	4.01M	479.4M	8,000 nodes per move
c24	B300A84A	4.00M	481.2M	8,000 nodes per move; first set from a hidden-layer net
c25	B300A84A	4.00M	481.2M	8,000 nodes per move
c26a	B300A84A	419k	50.2M	8,000 nodes per move; stopped early when CA63487B was promoted
c27a	CA63487B	2.22M	268.2M	8,000 nodes per move; stopped early when 18FE9B0A was promoted
c28	18FE9B0A	4.00M	483.7M	8,000 nodes per move
e12	BE9A5723	3.05M	254.3M	5,000 nodes per move; stopped at the same fleet time as c12; 96% of its games (an incomplete PGN download)
d1	DF8B55A5	301k	25.2M	DFRC
d2	BE9A5723	300k	25.3M	DFRC
d3	85B5CDD6	1.00M	83.8M	DFRC
d4	85B5CDD6	600k	50.9M	DFRC
d5	A617C0B4	600k	52.7M	DFRC
d6	AE339E48	997k	113.7M	DFRC, 8,000 nodes per move, decisive games played out with 5-man tablebase adjudication
d7	ABE5E000	1.00M	113.9M	DFRC, 8,000 nodes per move, decisive games played out with 5-man tablebase adjudication
d8	B300A84A	1.00M	114.7M	DFRC, 8,000 nodes per move, decisive games played out with 5-man tablebase adjudication
d9	B300A84A	1.00M	114.8M	DFRC, 8,000 nodes per move, decisive games played out with 5-man tablebase adjudication
d10a	B300A84A	173k	19.8M	DFRC, as d9; stopped early when CA63487B was promoted
d11	CA63487B	1.00M	115.4M	DFRC, 8,000 nodes per move, decisive games played out with 5-man tablebase adjudication
d12	18FE9B0A	1.00M	115.3M	DFRC, as d11
d13	EBDBE4F3	1.00M	116.0M	DFRC, as d11
d14	EBDBE4F3	1.00M	115.8M	DFRC, as d11

Later sets were played by progressively stronger networks: roughly +180 Elo from c3 to c4's player, then +10 to +55 per generation.

Licence

CC BY 4.0. Please credit "Incipit self-play data (Adam Twiss)".