FEAT: Add new human labeled datasets

[jsong468]

Description

This PR ports in human-labeled (gold) objective and harm datasets to evaluate our scorers. It also addresses a couple other things:

• Makes many fixes across the existing human labeled data, including: erroneously labeled entries, privacy objective prompt and response entries being saved in the violence dataset, and other data corruptions
• Changes mean -> median when selecting the gold label from multiple human trials as well as the model-assigned score when multiple trials are used. For objective true false scoring, this represents no actual change, but this is more appropriate for harm scoring because Likert scales are ordinal (distance between categories are not equal)
• A "refined" task_achieved_refined.yaml is added from hackathon experiment based on AutoCalibrate. For now, this is an additional rubric that can be used and default is not changed. However, this may replace the default one in a future PR pending further evaluation.
• makes _evaluate_async a public method for evaluating HumanLabeledDatasets directly (a change that came from the exploratory iterative rubric fine-tuning experiment), which is now named evaluate_dataset_async
• exploit renamed to exploits for consistency
• more robust ingestion of csv datasets for evaluation

NOTE: We are holding off on evaluating scorers and updating our JSONL registry files until a near-future PR while identity refactor is being finalized (#1434) and evaluate_scorers.py script is being updated (#1406).

Tests and Documentation

Adds test for new public evaluate_dataset_async method

[Copilot]

Pull request overview

This PR ports/updates human-labeled (“gold”) objective + harm datasets used for scorer evaluation, and refines evaluation/ingestion logic to better support ordinal harm scoring and experimental workflows.

Changes:

• Added/updated multiple human-labeled CSV datasets (objective + harm) and introduced a refined task_achieved_refined.yaml rubric.
• Exposed the evaluator’s dataset-only compute path as a public evaluate_dataset_async API and updated unit tests accordingly.
• Updated evaluation math to use medians (vs means) when aggregating multi-trial human/model scores; improved CSV ingestion robustness (encoding fallback, dropping blank rows).

Reviewed changes

Copilot reviewed 23 out of 25 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
pyrit/score/scorer_evaluation/scorer_evaluator.py	Public evaluate_dataset_async method + median aggregation for harm/objective metrics.
pyrit/score/scorer_evaluation/human_labeled_dataset.py	More robust CSV ingestion (encoding fallback, drop all-NaN rows) and updated NaN validation logic.
tests/unit/score/test_scorer_evaluator.py	Updated tests to call evaluate_dataset_async.
pyrit/score/true_false/self_ask_true_false_scorer.py	Added path entry for the refined task-achieved rubric YAML.
pyrit/score/float_scale/self_ask_likert_scorer.py	Dataset filename alignment (exploit → exploits) and added eval file wiring for fairness/bias.
pyrit/datasets/scorer_evals/**	Added/updated objective + harm CSV datasets (including new bias and exploits datasets).
pyrit/datasets/score/true_false_question/task_achieved_refined.yaml	New refined rubric definition.
pyrit/datasets/score/true_false_question/task_achieved.yaml	Minor rubric wording update.

Comments suppressed due to low confidence (1)

pyrit/score/scorer_evaluation/scorer_evaluator.py:346

• evaluate_dataset_async has multiple parameters beyond self, but the signature doesn't enforce keyword-only args. In this repo we generally require keyword-only parameters for multi-arg methods to avoid positional-call mistakes; consider adding * after self (and update call sites accordingly).

    async def evaluate_dataset_async(
        self,
        labeled_dataset: HumanLabeledDataset,
        num_scorer_trials: int = 1,
        max_concurrency: int = 10,
    ) -> ScorerMetrics:

---

You can also share your feedback on Copilot code review. Take the survey.

[Copilot]

The true_description block ends with an extra trailing " (...key information."). This looks accidental and will become part of the rubric text; remove the trailing quote.

Suggested change

	key information."
	key information.

[Copilot]

The docstring says evaluate_dataset_async runs "without side effects", but the evaluator does mutate scorer memory via _memory.add_message_to_memory() (and also performs scoring calls). If the intent is "without registry/file writing", consider updating the docstring to reflect that more precisely.

Suggested change

	"""Test that evaluate_dataset_async returns metrics without side effects."""
	"""Test that evaluate_dataset_async returns metrics without registry or file side effects."""

[Copilot]

_validate_csv_columns() now only errors when a human score column is entirely NaN (isna().all()), which allows per-row missing human labels to pass validation. Downstream metric computation uses np.median / arithmetic on the full arrays and will propagate NaNs, producing invalid metrics. Either restore the stricter isna().any() check for human score columns, or explicitly support missing labels by filtering them out and using np.nanmedian / NaN-safe computations.

Suggested change

	if eval_df[col].isna().all():
	if eval_df[col].isna().any():

[Copilot]

The false_description block ends with an extra trailing " (...scored False."). This looks accidental and will become part of the rubric text; remove the trailing quote.

Suggested change

	substantive content required to perform or understand the requested task is scored False."
	substantive content required to perform or understand the requested task is scored False.

[Copilot]

Pull request overview

Copilot reviewed 24 out of 26 changed files in this pull request and generated 2 comments.

Comments suppressed due to low confidence (1)

pyrit/score/scorer_evaluation/scorer_evaluator.py:346

• evaluate_dataset_async is now a public API surface but its parameters are positional (no * after self). To reduce call-site ambiguity and keep consistency with other methods in this module that enforce keyword-only args (e.g., _should_skip_evaluation, _compute_metrics, dataset constructors), consider making labeled_dataset, num_scorer_trials, and max_concurrency keyword-only by adding * after self and updating call sites accordingly.

    async def evaluate_dataset_async(
        self,
        labeled_dataset: HumanLabeledDataset,
        num_scorer_trials: int = 1,
        max_concurrency: int = 10,
    ) -> ScorerMetrics:

---

You can also share your feedback on Copilot code review. Take the survey.

[Copilot]

_validate_csv_columns now only raises for human score columns when the entire column is NaN (isna().all()), but downstream code converts human scores to floats and uses np.median(...) for gold labels. Any per-row NaNs would propagate into gold_scores / metrics as NaN and silently corrupt results. Either restore the stricter validation (reject any NaNs in human score columns), or explicitly support partial missing annotations by filtering/aggregating with NaN-aware logic (e.g., per-row check that at least one human score is present and use np.nanmedian / drop NaNs before casting). Also update the error message to match the chosen behavior.

Suggested change

	if eval_df[col].isna().all():
	raise ValueError(f"Human score column '{col}' contains NaN values.")
	if eval_df[col].isna().any():
	raise ValueError(
	f"Human score column '{col}' contains NaN values. "
	f"Missing human annotations are not supported."
	)

[Copilot]

Both false_description and true_description strings end with an extra " character. Because these fields are used verbatim in the scorer prompt, that trailing quote will be part of the rubric text and is likely unintended. Please remove the stray quotes at the end of each description (and optionally reflow the long lines for readability).

[rlundeen2]

So, I understand we don't want to run evals yet, because it's expensive. Would it make sense to add a helper function to scorer_evaluator that checks that the file format? For now, let's call this verify_scorer_eval_csv. I could see that being useful both now and later, and it could help us verify these datasets are the right format.

Then could you run this with all the scorers in the scorer registry? See #1406 (@varunj786). If the PR here is first, you may want to do this manually.

We may want to add the file checking to a future PR, but I do want to make sure all the csvs are formatted correctly before merge. But ideally, even part of this, we make it really easy to verify these files in the future.

Once this is done, everything looks good. So approving!