Datasets:

---

hf-internal-testing

/

tokenizers-test-data

like 0

Follow

Hugging Face Internal Testing Organization 174

The encode benches in huggingface/tokenizers load their model configs from this dataset. Some entries were slimmed-down fixtures or diverged from the source models, which made bench numbers unrepresentative (reported as "vocab is smaller than the real model"). This PR makes every bench config byte-identical to tokenizer.json in the original model repo:

• gpt-oss.json (new): from openai/gpt-oss-20b (byte-identical in openai/gpt-oss-120b). 199,998 vocab / 446,189 merges vs 3,060 / 4,000 in gpt-oss-slim.json.
• glm-5.2.json (new): from zai-org/GLM-5.2. 154,820 vocab / 321,649 merges vs 2,951 / 4,000 in glm-5.2-slim.json.
• gemma-4.json (new): from google/gemma-4-31B-it (the variant gemma-4-slim.json was derived from — vocab/merges are identical across gemma-4 sizes, only the base models' post_processor differs). 262,144 vocab / 514,906 merges vs 2,787 / 4,000 in the slim file.
• mistral-small-4.json (new): from mistralai/Mistral-Small-4-119B-2603 — the current tekken chain (Sequence[Split, ByteLevel] pre-tokenizer, 131,072 vocab, 1,000 added special tokens), a workload archetype the benches didn't cover.
• llama-2.json (refresh): synced byte-identical with meta-llama/Llama-2-7b-hf. The previous copy had the same vocab and merge set but 119 trailing whitespace merges in a different order — merge order is BPE priority, so whitespace runs tokenized differently.
• llama-3-tokenizer.json (refresh): synced byte-identical with meta-llama/Llama-3.1-8B (the file was already semantically this tokenizer, but re-serialized in the tuple-merges format, 17.2MB vs 9.1MB). Vocab, merges, and added tokens are unchanged — encode ids are identical.
• .gitattributes: track the three new configs via LFS/Xet, same as the other full configs.

The -slim fixtures are left untouched. The bench manifest in huggingface/tokenizers points at the full configs as of the fix/bpe-benches branch.