Contributing a model
MILIM-Bench scores any G2P system, including LLMs. Run the model yourself with
whatever tooling you like, then add a column to data/predictions.tsv.
Only the words named in each row's Label are scored, so the model is given
the sentence plus the words to convert, and returns phonemes for just those.
1. Run the model
Use one of the prompts in PROMPTS.md as the system prompt. If you run one unchanged, reuse its id.
You are welcome to write your own instead. Read the existing ones as a reference, improve on them, and include yours in the PR as a new section with the next free id. A better prompt is a legitimate contribution, so there is no need to stick to what is already there.
Your column is named model@id, e.g. gemini-3.1-pro@p1, so readers can tell
which rows ran the same prompt and are directly comparable. The same model can
appear more than once with different prompts.
Send one row per line, tab-separated: the Text from data/gold.tsv, then the
target words. Batch around 250 rows per call; the benchmark has 1,653 rows.
תגידי, הכל בסדר איתך? איתך
אני נותן לך את הספר, תקרא אותו. לך
אני נותן לך את הספר, תקראי אותו. לך
שלחתי אליך מכתב, את קיבלת אותו? אליך קיבלת
The response is one line per input line, in order, with the phonemes space-separated:
ʔitˈaχ
leχˈa
lˈaχ
ʔelˈajiχ kibˈalt
Check the line count matches what you sent. A dropped line shifts every row after it; split the batch and retry rather than realigning by hand.
2. Write the predictions
Add a column to data/predictions.tsv named after your model, same rows in the
same order as data/gold.tsv. Each cell pairs the indexes from Label with the
phonemes you got back:
3=ʔitˈaχ
2=leχˈa
2=lˈaχ
1=ʔelˈajiχ 4=kibˈalt
Normalization is automatic: x → χ, r → ʁ, g → ɡ, and anything
outside the phoneme set is stripped, punctuation included.
3. Score
uv run src/evaluate.py
All columns are picked up automatically and reported per category. Open a PR with the new column.