Datasets:
parent string | tokenizer string | vocab_size int64 | dtype string | n_tokens int64 | n_docs int64 | seq_len int64 | n_sequences int64 | v3_prefix_tokens int64 | v3_prefix_seqs int64 | new_perm_seed int64 | sources list | docs_per_source dict | sha256 dict | elapsed_s float64 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
data/v3 | NousResearch/Llama-2-7b-hf | 32,000 | uint16 | 64,600,000,000 | 53,738,564 | 2,048 | 31,542,968 | 8,200,000,000 | 4,003,906 | 2 | [
"sample-10BT",
"fineweb-edu-full"
] | {
"sample-10BT": 9672101,
"fineweb-edu-full": 44066463
} | {
"tokens_u16.bin": "cdc92f6fde31115c9125631f1a9970712593f007264fcf81fab48175b66fe6fa",
"tokens_u16_doc_offsets.npy": "8169c3469df8217dd3195ae511009ca0bf47298bb84f085be5db791432c1db4f",
"seq_permutation.npy": "ac41a04512e8f76c66306d5eea4da538083977a5aa1c70482284b7dfb7888736",
"main_heldout.bin": "23887c3297bff9... | 21,423.5 |
Pre-tokenised language-model training corpus, v4
The v4 build of the v3 corpus: a
byte-identical superset of v3 for longer runs. 64,600,000,000 tokens of FineWeb-Edu, tokenised with the Llama-2
tokenizer (vocabulary 32,000, uint16, no BOS/EOS, documents concatenated), plus the same three held-out evaluation
slices as v3. Sequences are 2,048 tokens; seq_permutation.npy fixes the training order so that a run on the first N
sequences is a byte-identical prefix of any longer run.
Relation to v3. The first 8,200,000,000 tokens of tokens_u16.bin are v3's tokens_u16.bin byte for byte, the
first 6,954,441 document offsets are v3's, and the first 4,003,906 entries of seq_permutation.npy are v3's
permutation. A run on the first N <= 4,003,906 sequences of v4 is therefore identical to the same run on v3 (and, through
v3, the first 512,695 sequences are the 1.05e9-token v2 block). The 27,539,062 new sequences follow in a seeded random
order (seed 2). New tokens come from the remainder of FineWeb-Edu sample-10BT after the point where v3 stopped (same
stream order), then from the full FineWeb-Edu dump (default config, file order); 53,738,564 documents in total
(9,672,101 from sample-10BT, 44,066,463 from the full dump). The full dump was not de-duplicated against
sample-10BT; since the sample is a small subset of the dump, a small fraction of the new tokens repeat earlier
documents.
| file | content |
|---|---|
tokens_u16.bin |
64.6e9 training tokens (uint16); first 8.2e9 equal v3 |
tokens_u16.sha256 |
SHA-256 of tokens_u16.bin |
tokens_u16_doc_offsets.npy |
document boundaries |
seq_permutation.npy |
fixed order of the 31,542,968 sequences; first 4,003,906 equal v3 |
main_heldout.bin |
in-domain held-out slice (identical to v3) |
xd_lang.bin (+offsets) |
10M tokens of German web text (FineWeb-2 deu_Latn), cross-domain probe (identical to v3) |
xd_code.bin (+offsets) |
10M tokens of Python (the-stack-smol-xs, codeparrot-clean-valid), cross-domain probe (identical to v3) |
artifact_meta_v4.json |
manifest with the SHA-256 of every file |
Sources and licences: FineWeb-Edu and FineWeb-2 (ODC-By 1.0); the-stack-smol-xs and codeparrot-clean-valid (permissively licensed source files; see the upstream cards). This repository redistributes token ids only.
- Downloads last month
- 33