A byte-level BPE tokenizer trained exclusively on SEC filings, drawn from a 4,356-document, 1.3 GB cleaned EDGAR corpus (2010-2024, ~4,250 companies; exact training-set size below). The fair-fight variant: identical pre-tokenization rules to GPT-2/GPT-4/GPT-4o/Qwen3, so the comparison isolates vocabulary.

Every number below is measured on 216 held-out documents no variant ever trained on; a 450-document cross-form probe (10-Q, 8-K, DEF 14A) confirmed the margins generalize past the training form. Full methodology, benchmark tables, and the corpus pipeline: write-up · repository.

or with the tokenizers library directly:

Vocab size 65,536, min merge frequency 2, pre-tokenization mode gpt2, 4,140 training documents, one <|endoftext|> special token. Training corpus built from primary filing documents (iXBRL/HTML cleaned to text with numbers and real Unicode preserved); 5% held-out split with content-hash deduplication before any training.

Trained with standard GPT-2-style byte-level pre-tokenization; no special caveats. A tokenizer is bound to the model trained with it, so this is a drop-in for new models and tokenizer research, not for existing checkpoints. Compression is the measured claim; see the write-up for the full limitations section.