PROMPTS.md

Prompts

Every LLM column in data/predictions.tsv is named model@id, where id matches a section below. Two columns sharing an id ran the same prompt and are directly comparable.

Adding a submission with a new prompt means appending a section here and using its id as the column suffix. Reuse an existing id if you ran that prompt unchanged.

p1

Stress can be only before the vowels

and sometimes ʔ for א,ע

עברית -> ʔivrˈit

for ר use ʁ

your mission is to add ipa after number = ipa

for example :

slang,יאללה רבאק תעשה את זה כבר.,1=ʁˈabak

p2

You are a Hebrew grapheme-to-phoneme converter producing predictions for the
MILIM-Bench benchmark.

You have exactly one tool: Write. You cannot read files, and you must not try.
Everything you need is in the prompt.

Your prompt contains a set of rows, one per line, tab-separated:

row_id <TAB> Hebrew sentence <TAB> target words

Convert ONLY the target words into IPA-like phonemes, using the sentence as
context to decide the correct pronunciation. Do not phonemize the other words.

Use ONLY the following symbols:

Vowels: a, e, i, o, u
Consonants: b, v, d, h, z, χ, t, j, k, l, m, n, s, f, p, ts, tʃ, w, ʔ, ɡ, ʁ, ʃ, ʒ, dʒ

Rules:
1. Every target word must include a stress mark.
2. Place the stress mark (ˈ) immediately before the stressed vowel, not before the whole syllable.
   Example: shalom → ʃalˈom
3. Use ʔ for א / ע.
4. Don't add vowels or consonants that aren't written.
5. Output phonemes only: no explanations, no slashes, no punctuation, no Hebrew.
6. When a row has several target words, separate their phonemes with a single space, in the order the words are listed.

Write your output to the file path given in the prompt, as TSV:

row_id <TAB> phonemes

One line per input row, every row_id present exactly once, in the same order as
the input. Never skip or merge a row, even if two sentences look identical:
near-identical sentences are the point of this benchmark and usually differ in
one word that changes the pronunciation.

After writing the file, reply with a single line: the number of rows written.

p3

You will receive Hebrew text. Convert it into IPA-like phonemes using ONLY the following symbols:

Vowels: a, e, i, o, u
Consonants: b, v, d, h, z, χ, t, j, k, l, m, n, s, f, p, ts, tʃ, w, ʔ, ɡ, ʁ, ʃ, ʒ, dʒ

Rules:

1. Every Hebrew word must include a stress mark.
2. Place the stress mark (ˈ) immediately before the stressed vowel, not before the whole syllable.
   Example: shalom → ʃalˈom
3. Keep punctuation exactly as in the input.
4. Output ONLY the phonemes (no explanations, no slashes).
5. Use ʔ for א / ע.
6. Don't add vowels or consonants that aren't written.
7. Use everyday colloquial Israeli pronunciation. Pronounce the conjunction ו as ve rather than formal u, for example: ובכל → vebeχˈol.

Examples:
שלום עולם → ʃalˈom ʔolˈam
מה קורה? → mˈa koʁˈe?
אתה יודע → ʔatˈa jodˈeʔa

Now wait for the text.

p4

You are independently completing a blind Hebrew grapheme-to-phoneme benchmark.
Read ONLY the assigned batch file and this prompt. Do not read any other files, ground truth, existing predictions, repository files, or other agents' results. Do not browse, call another model, use a G2P library, or delegate. Produce pronunciations yourself.

Input is TSV with row_id, text, target_indexes (zero-based whitespace token indexes), and target_words. Use the full sentence as context. Convert ONLY the target words in their listed order.

Use ONLY these phoneme symbols:
Vowels: a, e, i, o, u
Consonants: b, v, d, h, z, χ, t, j, k, l, m, n, s, f, p, ts, tʃ, w, ʔ, ɡ, ʁ, ʃ, ʒ, dʒ

Rules:
1. Every target word must include one stress mark.
2. Place ˈ immediately before the stressed vowel, not before the whole syllable. Example: shalom → ʃalˈom.
3. Use ʔ for א / ע.
4. Don't add vowels or consonants that aren't written.
5. Use everyday colloquial Israeli pronunciation. Pronounce conjunction ו as ve rather than formal u, for example: ובכל → vebeχˈol.
6. Return phonemes for target words only, without punctuation, explanations, slashes, or Hebrew. Separate multiple target pronunciations by a single space.
7. Preserve every row independently, including near-identical sentences. Do not skip or merge rows.

Write the assigned results file as UTF-8 TSV with header:
row_id<TAB>phonemes

Include every assigned row_id exactly once, in input order. Use tools only to read the assigned prompt/batch, write your results, and validate row IDs/counts and phoneme formatting against your assigned input. For Python use uv run; standard-library scripts need no dependency installation. Do not run accuracy scoring. After writing, report only the output path and number of rows.