Asmachta — Hebrew Attributed QA

131 Hebrew question-answer records for testing whether a model's generated answer is actually grounded in its source document. Every claim in reference_answer carries a character-exact quoted span from source_text — checkable with string equality, no judge model needed. A third of the questions are deliberately unanswerable, so you can measure hallucination-vs-abstention directly instead of inferring it.

Quick start

import json

with open("asmachta.json", encoding="utf-8") as f:
    records = json.load(f)

record = records[0]
print(record["question"])
print(record["reference_answer"])
for claim in record["claims"]:
    print(" claim:", claim["text"])
    for span in claim["attribution"]:
        excerpt = record["source_text"][span["start_char"]:span["end_char"]]
        assert excerpt == span["source_excerpt"]  # always true, that's the point
        print("  supported by:", excerpt)

Fields

Field	Meaning
id	Record ID
genre	dialogue (Knesset), encyclopedic (Wikipedia), or journalistic (news)
difficulty	0 = unanswerable by design, 1 = simple, 2 = complex/multi-span
question / source_text	The question and the document to answer it from
reference_answer	Gold answer, decomposed into claims
claims[].attribution[]	start_char/end_char/source_excerpt/verified — the span in source_text that backs this claim, checked character-for-character
human_ratings	Four annotators' 0–2 scores on 5 quality dimensions — indicative, not precise ground truth

Unanswerable (difficulty=0) records have empty claims — there's nothing to attribute when the correct answer is "not in the document."

Running and scoring your own model

score.py (included in this repo, no dependencies beyond the standard library) checks whether a model's quoted evidence for a claim actually occurs in source_text — exact match, then whitespace-normalized, then a fuzzy match for copying artifacts. It expects model output in the format:

1. <claim text> [<verbatim quote from source_text>]
2. <claim text> [<verbatim quote from source_text>]

A minimal end-to-end example — call your model on every record, then score each response:

import json
from score import score_model_output

with open("asmachta.json", encoding="utf-8") as f:
    records = json.load(f)

PROMPT = """ענה על השאלה אך ורק על סמך המסמך הבא. פרק את תשובתך למשפטים
נפרדים ("טענות"). אחרי כל טענה, בסוגריים מרובעים, צטט קטע מדויק ומילולי
מהמסמך שתומך בה. פורמט: מספור עוקב החל מ-1, טענה אחת בכל שורה.

מסמך:
{source_text}

שאלה:
{question}

תשובה:"""

def call_your_model(prompt: str) -> str:
    # Plug in your own inference call here.
    raise NotImplementedError

results = []
for record in records:
    if record["difficulty"] == 0:
        continue  # unanswerable items need a different check -- see below
    prompt = PROMPT.format(source_text=record["source_text"], question=record["question"])
    output = call_your_model(prompt)
    scored = score_model_output(output, record["source_text"])
    results.append({"id": record["id"], **scored})

overall = sum(r["precision"] for r in results if r["precision"] is not None) / len(results)
print(f"verified-quote rate: {overall:.3f} over {len(results)} records")

For difficulty=0 (unanswerable) records, score whether your model correctly declines instead — for example, by checking whether its raw output signals "not in the document" rather than attempting a citation. There's no single fixed phrase to check for: judge it however fits your model's expected refusal style.

A note on format compliance

score.py only scores output that matches the format above. A model that has the right answer but doesn't follow this exact structure — for instance, a reasoning model that leaves extra text before or after its numbered claims — will score as ungrounded on precision, not because the content was wrong, but because nothing could be parsed. That's a different failure mode from actually getting the answer wrong, and it's worth keeping the two separate in your own evaluation rather than letting a low score stand in for both. If you're comparing models with different formatting reliability, consider checking format compliance and grounding as two distinct things.

License

MIT. See LICENSE.md.

Credits

Dataset design: Tal Geva. Annotation project: Eyal Rosenstein. Maintenance and consulting: Noam Ordan