Datasets:

procmarco
/

ita-en-code-tokens-48k

license: odc-by
language:
  - it
  - en
task_categories:
  - text-generation
tags:
  - italian
  - english
  - code
  - fineweb-2
  - fineweb-edu
  - pretraining-tokens

ita-en-code-tokens-48k

Pretokenized 40% Italian / 40% English / 20% code pretraining tokens (~40B), tokenized with procmarco/ita-en-code-bpe-48k (vocab 49152). Documents packed <|bos|> … <|eos|>.

Sources: FineWeb-2 ita_Latn (filtered), FineWeb-Edu sample/100BT (int_score≥3), codeparrot/github-code-clean. The three streams are interleaved to the 40/40/20 token ratio.

Format

tokens_000000.bin, … — fixed 268,435,456 tokens each (uint16 LE, 512 MiB).
tokens_eval.bin — disjoint held-out eval (16,777,216 tokens).
meta.json — manifest (shard size, per-source token counts, positions, special ids).

import numpy as np
a = np.fromfile("tokens_000000.bin", dtype="<u2")   # ids in [0, 49152)

Trainer consumes via --token-shard-dir <dir> --token-shard-tokens 268435456.