fineweb-edu-bpe32k-10BT

FineWeb-Edu sample-10BT, pre-tokenized with a 32,768-vocab ByteLevel BPE tokenizer trained on a 2GB sample of the same corpus. This is the BPE twin of ShadeEngine/fineweb-edu-bytes-10BT: both encode the exact same documents, so models trained on the two can be compared head-to-head in bits-per-byte.

Format

nanoGPT-style flat binary shards of uint16 token ids, ~500MB each (train_*.bin, val_*.bin), little-endian. Token id 0 is <|doc|>, prepended to every document as a separator. No other structure — just np.memmap(path, dtype=np.uint16) and sample windows.

• Total: 9,991,958,085 tokens (~10B), 9.66M documents
• Average compression: 4.602 bytes/token on this corpus (recorded in meta.json as avg_bytes_per_token)
• Every 1000th document went to the validation split (same split as the byte dataset)

Tokenizer

tokenizer_32k.json is included — a HuggingFace tokenizers ByteLevelBPETokenizer file. Load with:

from tokenizers import Tokenizer
tok = Tokenizer.from_file("tokenizer_32k.json")

Provenance

Built as part of a byte-level vs BPE vs H-Net comparison replicating the experiments in Cartesia's "From Tokens to Bytes" posts at ~100M scale. License follows the source dataset (ODC-By).