arm string | boundary_policy string | elapsed_sec float64 | shard_tokens int64 | shards int64 | source_tokens dict | special_token_ids dict | tokenizer string | tokens int64 |
|---|---|---|---|---|---|---|---|---|
fpabl1-b-fp | source streams are already BOS/document/EOS packed; arm scheduler interleaves token chunks | 37.857518 | 100,000,000 | 50 | {
"fp_en": 1000000000,
"fp_ita": 1000000000,
"fw2_ita": 1000000000,
"fwedu": 1000000000,
"replay": 1000000000
} | {
"bos": 49119,
"eos": 49120,
"pad": 49121
} | runs/mixed-tokenizer-48k/tokenizer | 5,000,000,000 |
fpabl1-arm-b-fp-tokens-48k
Pre-tokenized bins for the FinePhrase vs FineWeb ablation (fpabl1), arm fpabl1-b-fp.
- Tokenizer:
runs/mixed-tokenizer-48k/tokenizer(byte-BPE, 48k vocab; special ids bos=49119, eos=49120, pad=49121). - Format:
uint16little-endian, 50 shards x 100,000,000 tokens = 5,000,000,000 tokens total. - Boundary policy: source streams are already BOS/document/EOS packed; arm scheduler interleaves token chunks
Source token composition:
- fp_en: 1,000,000,000
- fp_ita: 1,000,000,000
- fw2_ita: 1,000,000,000
- fwedu: 1,000,000,000
- replay: 1,000,000,000
See meta.json for the full manifest. Companion arm: the other fpabl1-arm-*-tokens-48k repo.
- Downloads last month
- 316