source/tests/test_ruler_data.py

from reproduction.prepare_ruler_data import (
    LLAMA3_ASSISTANT_SUFFIX,
    LLAMA3_USER_PREFIX,
    make_exact_llama3_chat_prompt,
    normalize_row,
    prefix_pad_text,
)


def test_normalize_ruler_row_matches_official_harness_schema():
    row = {"index": 3, "input": "prompt", "outputs": ["answer"], "length": 32701}
    assert normalize_row(row) == {
        "index": 3,
        "input": "prompt",
        "outputs": ["answer"],
        "length": 32701,
        "length_w_model_temp": 32701,
        "answer_prefix": "",
    }


def test_prefix_padding_is_inserted_before_complete_original_prompt():
    padded = prefix_pad_text("Question: preserved?", pad_tokens=3)
    assert padded == " filler filler filler\nQuestion: preserved?"
    assert padded.endswith("Question: preserved?")


class _WhitespaceTokenizer:
    def __call__(self, text, add_special_tokens=False):
        assert add_special_tokens is False
        return {"input_ids": text.split()}


def test_exact_llama3_chat_prompt_preserves_task_and_assistant_generation_suffix():
    raw = "Document: the answer-token is preserved. Question: what is it?"
    target = 40
    prompt = make_exact_llama3_chat_prompt(
        raw, tokenizer=_WhitespaceTokenizer(), target_tokens=target
    )
    token_ids = _WhitespaceTokenizer()(prompt, add_special_tokens=False)["input_ids"]
    truncated = token_ids[-target:]
    assert len(token_ids) == target
    assert truncated == token_ids
    assert prompt.startswith(LLAMA3_USER_PREFIX)
    assert prompt.endswith(LLAMA3_ASSISTANT_SUFFIX)
    assert raw in prompt
    assert "answer-token" in prompt