unparallelled/BluScriptCP-aligned-mixes

BluScriptCP aligned mixes (2026-10-15)

Training mixes for long-context extension of the 1.70B fullchess model (GPT-2 tokenizer), built by a new two-stage mixer:

1. mkmix_bucket.py selects whole documents to hit TARGET domain shares and a TARGET document-length profile at the same time (iterative proportional fitting under per-cell supply caps; documents drawn uniformly at random within each domain x length cell).
2. mkpack_aligned.py lays those documents out so that no document of length <= T is cut by a training-window edge. Documents longer than T start a window and are chunked at multiples of T from their own start; all others are placed whole, best-fit decreasing. No padding: every window is exactly T tokens.

Contents

folder	T	train tokens	windows (= T-token batches)	steps at 128 windows/step	what
p1b_16k_aligned/	16,384	1,000,439,809	61,062	477	ready to train: 4k->16k extension mix, aligned packing (--float-tails)
p1b_16k_unaligned/	16,384	1,000,589,883	--	477	the same documents, plain concatenation (for comparison)
p2d_32k_aligned/	32,768	600,473,601	18,325	143	ready to train: 16k->32k extension mix, aligned packing
p2c_32k_unaligned/	32,768	600,732,695	--	143	the same documents, plain concatenation (for comparison)
tools/					the mixer scripts and the statistics scripts

Each mix folder holds train_000000.bin, val_000000.bin (raw uint16 GPT-2 tokens; EOT 50256 marks the START of every document), mix_report.json (provenance and statistics) and selected.npz where present (keys of the source documents used, to exclude them from later mixes). MANIFEST.txt lists sha256 and byte size of every file.

Domain shares and length profile

mix	code	math	commoncrawl	chess (Lichess)	stockfish
P1b	37.0%	29.0%	28.0%	3.0%	3.0%
P2c / P2d	40.0%	29.1%	24.9%	3.0%	3.0%

Share of training tokens by whole-document length:

mix	<4k	4-16k	16-32k	32-64k	64-128k	>=128k
P1b	18.2	0.0	56.1	16.6	4.5	4.6
P2c / P2d	15.6	3.6	17.0	22.4	8.5	32.9

P1b copies the length profile of the 2026-10-05 ladder's P1 mix (including its empty 4-16k bucket) with the 2026-10-13 ladder's domain shares. P2c/P2d copy the 2026-10-05 P2 profile for the text part, with 6% chess added in <4k.

What aligned packing buys

Share of tokens with at least d tokens of same-document context inside their window:

mix	packing	>= T/4	>= T/2	>= 3T/4
P1b (T=16k)	plain	52.1%	28.5%	11.2%
P1b (T=16k)	aligned	56.4%	35.4%	16.9%
P2c/P2d (T=32k)	plain	51.6%	28.1%	11.7%
P2c/P2d (T=32k)	aligned	56.6%	33.4%	14.8%

Packing cost: P2d drops 259,095 tokens (0.043%), P1b 180,274 (0.018%), all from short "reserve" documents (<= 227 tokens) cut to close the last gaps. P1b additionally inserts one EOT before each of 30,199 floated document tails (--float-tails: the last partial chunk of a split document cannot see its own beginning in any layout, so it is placed like a short document; strict alignment left 14.3% of window space empty for this length profile).

Using the aligned mixes

The layout assumes the loader reads window k as train[k*T : (k+1)*T] (+1 token for the shifted target), starting at offset 0, sequentially - which is what BluScriptCP's DataLoaderLite does with one data-parallel rank. A loader that reads at random offsets, or with a different T, loses the alignment (the data is still valid, just no longer aligned). Windows are shuffled at build time. The last token of each window predicts the first token of an unrelated window (0.003% of targets).

Intended for training with intra-document masking (CP_DOC_MASK=1, EOT 50256): a window that begins with a continuation chunk has no leading EOT and forms its own document id 0.

Source data

Selected from the fullchess pretraining shards (commoncrawl, code, math: navingv corpora; Lichess chess_v2 1-epoch; Stockfish adamkarvonen/chess_games stockfish_dataset). These documents were already seen once during fullchess pretraining.