Replace the prior V2 artifacts with the fully audited corrected V2 corpus: 1,525,213 documents and 1,986,440,772 canonical tokens in disjoint train, validation, and anneal splits. Rename the repository to the approximate 2B identity.
Add 32k and 49k CustomBPE and SuperBPE tokenizers
Publish exact source artifacts and Hugging Face-compatible tokenizers for CustomBPE and SuperBPE at 32,000 and 49,152 vocabulary sizes. Existing nanoGPT binaries remain tied to the canonical 12,288 tokenizer.
