Skip to content

fix: Makefile fixtures download - #2

Open
SBrandeis wants to merge 1 commit into
huggingface:tokbench-initfrom
SBrandeis:fix/makefile-model-paths
Open

fix: Makefile fixtures download#2
SBrandeis wants to merge 1 commit into
huggingface:tokbench-initfrom
SBrandeis:fix/makefile-model-paths

Conversation

@SBrandeis

Copy link
Copy Markdown
Contributor

Fix the Makefile

`make models` asked hf-internal-testing/tokenizers-test-data for
`models/<name>/tokenizer.json`. That layout does not exist in the dataset, so
every fetch 404'd and a fresh clone could not get past `make models`. The CI
bench job runs the same target and fails the same way.

The dataset stores each config as one flat file at its root, and the names do
not all follow one pattern (`gpt2.json`, but `llama-3-tokenizer.json`), so
BENCH_MODELS now carries a `name:file` pair per model and the fetch loop splits
it. No data had to be added to the dataset: all five configs were already
there.

Four of the five are byte-identical to the official model repos'
tokenizer.json. `llama-3-tokenizer.json` spells ids 128004/128008/128010 as the
Llama-3.1 names (`<|finetune_right_pad_id|>`, `<|eom_id|>`, `<|python_tag|>`)
where Llama-3 has reserved tokens, 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, so the benchmark reads the same
vocabulary either way.

Also moves the comment describing this target back above it; it had drifted
above `bigfixtures`.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant