mesh-1t-prod-mix: the epsilon production mix

The exact token stream of the epsilon run of parallax-8b (Chutes / Mesh, launched 2026-09-20): a 973.1B-token curriculum served from nine flat llama3 token files. NOT a text dataset: every artifact is a flat uint32 token file (NousResearch/Meta-Llama-3-8B tokenizer, EOS 128001 between documents). Files larger than 32 GiB are stored as <name>.part-NNN chunks: cat name.part-* > name reassembles them; per-part sha256 in CHUNKS.sha256, whole-file sha256 in SHA256SUMS.txt.

Curriculum

Two phases, no restart between them. The fleet trains at seq_len 4096, 19,005,440 tokens per fleet step (grad-accum 20 until 100B tokens, 40 until 600B, 80 after). LR peaks at 2e-4 after 1024 warmup steps, stays flat, and decays with a 1-sqrt schedule from 800B served tokens to the end of the curriculum (floor 2e-5).

Stationary phase: 899.8B tokens served

member	tokens in file	use_fraction	served	where	recipe (build weights)
p1full	594.0B	0.839	498.6B	phase1/	ClimbMix + FinePDFs-Edu premix 0.558 / UltraFineWeb 0.135 / Stack code 0.158 / CC-Math 0.046 / raw-tech 0.053 / diversity 0.050 (two 300B halves built 2026-08, see phase1/*/mix_manifest.json)
p2fresh	258.3B	0.996	257.3B	phase2/train.u32.bin	fresh, globally deduplicated against phase 1: ClimbMix-fresh 0.388 / FinePDFs-Edu 0.294 / Stack code 0.212 / raw-tech 0.059 / CC-Math 0.028 / diversity 0.017 (see phase2/mix_manifest.json)
web_v1	128.7B	0.310	39.9B	members/web_v1/train.u32.bin	FineWeb-Edu (350BT sample) 0.72 / UltraFineWeb l1 en hq 0.27 / Project Gutenberg (en) 0.01
km_nemomath	39.6B	0.997	39.5B	members/km_nemomath/train.u32.bin (train parts pending; val + manifests present)	Nemotron-CC-Math-v1 4plus (first 75% slice)
km_know	25.9B	0.996	25.8B	members/km_know/train.u32.bin	Cosmopedia v2 (smollm-corpus) 0.73 / dolmino pes2o 0.18 / dolmino wiki 0.09
km_fmath	28.7B	0.997	28.6B	members/km_fmath/train.u32.bin	FineMath 3+ 0.76 / proof-pile-2 algebraic-stack 0.24
books_v1	10.1B	1.000	10.1B	members/books_v1/train.u32.bin	LoC public-domain books 0.69 / gutenberg_english 0.29 / writingprompts 0.01

Anneal phase: 73.4B tokens served (starts at 899.8B)

member	tokens in file	use_fraction	served	where	recipe (build weights)
anneal_a	45.8B	1.000	45.8B	members/anneal_a/train.u32.bin	dolmino dclm (high quality) 0.60 / dolmino flan 0.24 / dolmino math 0.16 (gsm8k, tinyGSM excluded)
anneal_b	27.6B	1.000	27.6B	members/anneal_b/train.u32.bin	Nemotron-CC-Math-v1 4plus reserve (last 25%) 0.24 / FinePDFs-Edu top 0.20 / StarCoder data 0.36 (python .12, js .05, java .04, cpp .04, ts .03, go .02, rust .02, c .02, sql .01, shell .01) / OpenMathInstruct-2 0.10 / Cosmopedia v2 reserve 0.06 / dolmino wiki reserve 0.04

How a phase is read (epsilon/corpus_members_*_v9.json, sample law in the Mesh repo's docs/ops/CORPUS_MEMBERS.md): each member holds (n_tokens - 1) // 4096 samples of 4097 tokens; use_fraction keeps floor(f * shards) whole shards of the member, chosen by a seeded per-member permutation (seed 20260917), so nothing is ever repeated; the phase is the concatenation of the members' kept shards in table order, and the fleet's shard authority hands those shards out in a seeded permutation (seed 20260918). Every member was globally deduplicated (33-bit document hashes) against everything served before it and decontaminated (13-gram, stride 4) against the eval suite; the per-source counts, duplicate and decontamination hits are in each member's mix_manifest.json.

Layout

• phase1/ : p1full. The 2026-08 phase-1 build, two 300B halves of four shards each; the served file is the concatenation bb35_half/train_shard0..3 then bb8_half/train_shard0..3 (2,376,165,637,104 bytes, sha256 9ba7eb539419e10924ab4e37e3168a0e852563a39004c0d924ab237eacbaf0b7).
• phase2/ : p2fresh (train.u32.bin.part-NNN), the 2026-09-12 fresh dedup rebuild.
• members/<name>/ : the seven newer members (train.u32.bin.part-NNN, val.u32.bin, mix_manifest.json, manifest_<name>.json build recipe, shuffle record, SHA256SUMS.txt, CHUNKS.sha256). members/km_nemomath/ currently holds the val split and manifests only; its train parts are published after the run finishes (the file lives on the live corpus streamer).
• val/val_mix_v2.u32.bin (+ .json layout) : the 2048-window x 4097-token held-out mix the run's headline eval uses (windows drawn per member in proportion to the served mix; counts in the json).
• epsilon/EPSILON_MIX.json : machine-readable version of this page (per-member tokens, fractions, sha256, parts).
• MANIFEST.json : the same plus the 2026-08 phase-1 and 2026-09-12 phase-2 provenance blocks.

The earlier phase-3 / phase-4 plan stubs were never served and have been removed.