[Lanre Shittu (Shizoqua)]

Summary

Fixes #398.

When a SearchQA item's answers field was a single string instead of a list, evaluate() and the EM, F1 and sub_em helpers iterated its characters. The correct answer Paris scored EM 0, while P or s scored EM 1 and F1 1. A single string gold answer is now treated as a one-item list, both when scoring and in the gold answers saved on rollout rows.

User impact: hard is EM and soft is F1, so such items gave an inverted training reward (correct answers penalized, one-letter answers rewarded) and wrong reported scores. Saved rows also recorded the bare string, so anything that later iterated them got characters again.

Proof: the regression tests fail on upstream 45f65ca and pass with this change.

Root cause

evaluate() in skillopt/envs/searchqa/evaluator.py passed gold_answers straight to exact_match(), f1_score() and sub_em(), which loop over it. Iterating a str yields characters. In rollout.py, process_one() stored the raw item["answers"] as gold_answers, and run_batch()'s timeout and error rows stored it as gold_answer. materialize_searchqa.py writes lists, but custom data or a hand-written split can store "answers": "Paris", and nothing normalized it.

Changes

• rollout.py: a new _gold_answers(item) helper returns the gold answers as a list. process_one() uses it for scoring and the saved row, and the timeout and error rows use it too.
• evaluator.py: evaluate() also wraps a string gold_answers in a list, for direct callers. List gold answers are unchanged.
• Tests: tests/test_searchqa_evaluator.py covers scoring; new tests/test_searchqa_rollout_gold_answers.py covers the real process_one() row (fake model response) and the run_batch() error row.
• CHANGELOG entry under Unreleased / Fixed.

Updated after review feedback (thanks Muhammad Abdullah Butt (@1AbdullahButt)) to normalize once in the rollout as well.

Reproduction

from skillopt.envs.searchqa.evaluator import evaluate

for response in ["<answer>P</answer>", "<answer>s</answer>", "<answer>Paris</answer>", "<answer>London</answer>"]:
    result = evaluate(response, "Paris")
    print(response, result["em"], result["f1"], result["sub_em"])

Before:

<answer>P</answer> 1.0 1.0 1.0
<answer>s</answer> 1.0 1.0 1.0
<answer>Paris</answer> 0.0 0.0 1.0
<answer>London</answer> 0.0 0.0 0.0

After: Paris scores 1.0 1.0 1.0 and the other three score 0.0 0.0 0.0. Rollout rows record ["Paris"].

Testing

• python -m pytest tests/test_searchqa_evaluator.py tests/test_searchqa_rollout_gold_answers.py -q: on upstream 45f65ca, 5 failed and 6 passed; with the fix, 11 passed.
• All SearchQA-related tests plus tests/test_benchmark_task_output_paths.py: 61 passed.
• python -m pytest -q: 2221 passed, 12 skipped, 4 xfailed.
• python -m ruff check and python -m ruff format --check on the changed files: clean.
• python -m mkdocs build --strict: passed.

Risk and compatibility

Only items whose answers is a single string change, and those were scored and recorded incorrectly before. List inputs behave exactly as before. No backend, config key, file format or CLI change.

[Muhammad Abdullah Butt (1AbdullahButt)]

The fix in evaluate() is right. isinstance(gold_answers, str) wraps it before exact_match, f1_score and sub_em loop over it. Both callers in rollout.py (:243, :321) go through evaluate(), so the reward and the reported scores are both covered.

The raw string still travels on next to the fix, though. rollout.py:187 reads gold_answers = item.get("answers", []) and keeps that value. It's written to the trajectory at :198 ("gold_answers": gold_answers) and at :418 ("gold_answer": item.get("answers", [])). So for these items the score says gold_answers == ["Paris"] (the new test asserts that), while the trajectory records "Paris". Anything that later iterates the trajectory's gold answers, such as a report or a re-scoring script, gets characters again.

Normalizing once at the source, right after :187, would keep the two in step:

gold_answers = item.get("answers", [])
if isinstance(gold_answers, str):
    gold_answers = [gold_answers]

Then use gold_answers at :418 as well. You can keep the evaluate() guard too, for direct callers.

Merge note: #401, #402 and #403 each add an entry at the same spot in CHANGELOG.md. git merge-tree of any two of the heads conflicts there, though each merges cleanly against main on its own. Whichever lands second needs a rebase.

I read the diff, evaluator.py and rollout.py at the PR head. I didn't run the tests.

Reviewed with help from Claude Code.

[Lanre Shittu (Shizoqua)]

Thanks, good catch. Pushed c875c29: gold answers are now normalized once in rollout.py, and both the process_one row and the timeout/error rows use the normalized list, so saved rows record ["Paris"]. The evaluate() guard stays for direct callers. Added tests for both paths.