Datasets:

---

hf-internal-testing

/

tokenizers-test-data

like 0

Follow

Hugging Face Internal Testing Organization 176

tokbench's make models downloads models/<name>/tokenizer.json for its five benchmark models from this dataset. None of those paths exist yet, so a fresh clone fails at make models. This PR adds them.

Each file is the unmodified tokenizer.json from the official model repo (full vocabularies — all five pass tokbench's truncated-vocab check, and the full benchmark matrix was run locally against these exact files before opening this PR):

Notes:

• deepseek-v4: DeepSeek-V4-Pro and DeepSeek-V4-Flash serve a byte-identical tokenizer.json (same LFS object), so one file covers the family.
• llama-3: taken from the gated meta-llama/Meta-Llama-3-8B repo rather than reusing the root-level llama-3-tokenizer.json already in this dataset, because that file is a Llama-3.1 config: same 128000-entry vocab and identical merges, but ids 128004/128008/128010 hold <|finetune_right_pad_id|> / <|eom_id|> / <|python_tag|> where Llama-3 has <|reserved_special_token_*|>. tokbench names this model llama-3 and renders its chat fixtures from the Llama-3 (not 3.1) template, so the Llama-3 config is the consistent choice. With the current fixture set the two are interchangeable anyway — no fixture contains any of the differing special tokens.

Closing: this was the wrong fix, and the files were never missing.

All five configs are already in this dataset as flat root-level files (gpt2.json, deepseek-v4.json, bert-base-uncased.json, t5-base.json, llama-3-tokenizer.json). The first four are byte-identical to the official model repos' tokenizer.json. llama-3-tokenizer.json carries the Llama-3.1 special-token names (<|eom_id|>, <|python_tag|>, <|finetune_right_pad_id|> at ids 128008/128010/128004), but its model, pre_tokenizer, normalizer, post_processor and decoder are identical to Llama-3's, and it derives a byte-identical ranks.tiktoken and pattern.txt.

What actually broke was tokbench's Makefile: it fetched models/<name>/tokenizer.json, a layout this dataset does not use. That is fixed in tokbench instead, so no data is duplicated here.