An English pretraining corpus for the GoLLeM-v5 tiny-LM efficiency track (Glint Tiny-ML Leaderboard). ARC-MIX is a reasoning/knowledge-enriched expansion of SlayerLab/minimal-en-corpus-5b: the base multi-source EN mixture with ARC-relevant data upweighted (gold ~3×, related ~2×) to strengthen the ARC-Easy axis — the binding efficiency-constraint at this scale.

Key facts

Property	Value
Tokens (packaged, uint16)	9,417,035,832 (~9.42B)
Binary size	18,834,071,664 bytes (`train.bin`, little-endian `uint16`)
Tokenizer	BPE, vocab 12,288 (shared with minimal-en-corpus-5b)
Base	`SlayerLab/minimal-en-corpus-5b` (5.40B) + ARC-enrichment
Decontamination	13-gram shingles vs WikiText-2 / BLiMP / ARC test sets
`train.bin` SHA-256	`e87f594f7866fae7… (full in manifest)` (full checksum in manifest)
Language	English

What ARC-MIX adds over the base

minimal-en-corpus-5b is a 15-source EN blend (FineWeb-Edu, DCLM, StackExchange, open-web-math, FineMath, scientific papers, books/Gutenberg, code, CC-News, Wikipedia, …). ARC-MIX re-weights the mixture toward science/reasoning/QA-relevant content:

gold ARC-relevant sources ~3× (science QA, structured reasoning),
related sources ~2× (open-web-math, FineMath, scientific papers, educational web),
general web retained as the fluency/coverage backbone.

The goal is the ARC-Easy lever (capacity-gated per the GoLLeM-v5 study, finding W11): the same enrichment is null at 16M but lifts ARC at ≥32M capacity.

Provenance & preparation

Inherits the base corpus preparation: source-specific extraction, English filtering, exact + MinHash/LSH near-dedup, benchmark decontamination, whole-document train/val split, deterministic packaging. The ARC enrichment is applied as a deterministic re-weighted selection over the decontaminated base pool (no benchmark leakage introduced by the upweighting).

Usage (nanoGPT binaries)

import numpy as np
train = np.memmap("train.bin", dtype="<u2", mode="r")   # little-endian uint16
vocab_size = 12288
block_size = 1024

Configure the GoLLeM-v5 trainer (train_gpt_ref.py) with --vocab 12288 --dtype uint16.

Used by

GoLLeM-v5 32M (Path-B v1) — board #16 (eff 75.51), 16B tokens on ARC-MIX.
GoLLeM-v5 64M A/B v1 — clean Muon-vs-AdamW study, 13.1B tokens (~1.39 epochs) on this 9.42B corpus.

Licensing & provenance

Aggregate corpus; no new unified license on the underlying documents. Each document retains its source_id; use/redistribution remain subject to the upstream dataset terms (review upstream cards, especially for commercial use). Base provenance: SlayerLab/minimal-en-corpus-5b.

Base corpus: SlayerLab/minimal-en-corpus-5b
Models: SlayerLab/gollem-v5-ckpts
Board: Glint Tiny-ML Leaderboard

Downloads last month: 43

Total file size:

6.12 kB

GoLLeM-v5 ARC-MIX 9.4B (EN)