miguelcsx/babylm-2026-mosaic-bpe-corpus

MOSAIC BBPE16K Encoded Corpus

Reproducible encoded training artifact shared by the MOSAIC D384 model family. It contains the fixed 10M-word Strict-Small corpus representation used for a controlled 100M-word exposure schedule.

Contents

• —packed token and word arrays for training and validation;
• —document offsets and token frequencies;
• —document-complexity metadata;
• —lexical recombination neighbors used by the variation-set arms;
• —manifest.json with source and artifact provenance.

This repository contains encoded research artifacts, not an additional source of training text. The model repositories bundle the exact tokenizer needed for loading and evaluation.

Associated models

• —MOSAIC Sleep — TOLM D384 BBPE16K
• —MOSAIC Memory — TOLM D384 BBPE16K
• —MOSAIC Joint — TOLM D384 BBPE16K