Supra2-100M-tokens

# Supra2-100M — pretraining token bins **DO NOT USE.** Internal artifact of the Supra2-100M training run, published only because public storage is free. These are raw `uint16` token bins for one specific tokenizer (`supra2-tokenizer.json`) and carry the licenses of their upstream sources (FineWeb-Edu, DCLM, The Stack v3, FineMath, Cosmopedia, Wikipedia). No redistribution rights are granted or implied. Files may be deleted at any time. Layout: `tokens_XXX.bin` (4B tok each), `val_000.bin`, `recipe.json`, `SHA256SUMS.txt`, `supra2-tokenizer.json`.