scripts/score_hf_dataset.py

#!/usr/bin/env python3
"""Score new predictions with the paper's exact answer/abstention rules, offline."""

from __future__ import annotations

import argparse
import json
from collections import defaultdict
from pathlib import Path

from analyze_crossdomain_visual import strict_answer_match
from analyze_visual_judges import number, valid_parsed
from hf_release_common import STATES, read_rows, write_json


def score(labels, predictions):
    gold = {row["item_id"]: row for row in labels}
    records = {row["item_id"]: row for row in predictions}
    if not gold or len(gold) != len(labels):
        raise ValueError("Empty or duplicate gold IDs")
    if len(records) != len(predictions):
        raise ValueError("Duplicate prediction IDs")
    if records.keys() != gold.keys():
        raise ValueError(
            f"Prediction coverage differs: {len(gold.keys() - records.keys())} missing, {len(records.keys() - gold.keys())} unexpected"
        )
    output = {}
    for source in sorted({row["source"] for row in labels}):
        selected = [row for row in labels if row["source"] == source]
        groups = defaultdict(list)
        state_counts = {
            state: {"failures": 0, "invalid": 0, "valid_decision_errors": 0, "views": 0}
            for state in STATES[source]
        }
        counts = {
            "failures": 0,
            "invalid": 0,
            "valid_decision_errors": 0,
            "valid": 0,
            "groups_correct": 0,
            "joint_groups_correct": 0,
            "strict_answers_correct": 0,
            "answerable_views": 0,
            "unanswerable_views": 0,
        }
        for label in selected:
            record = dict(records[label["item_id"]])
            if "status" not in record:
                record = {
                    "status": "completed",
                    "parsed": {key: record.get(key) for key in ("answerable", "answer", "reason")},
                }
            elif record["status"] not in {"completed", "invalid_schema"}:
                raise ValueError(
                    "Unfinished prediction status; submit a terminal output for every item"
                )
            if record["status"] == "invalid_schema" and record.get("terminal_invalid") is not True:
                raise ValueError("invalid_schema must explicitly set terminal_invalid=true")
            parsed = valid_parsed(record)
            invalid = parsed is None
            wrong = not invalid and parsed["answerable"] != label["answerable"]
            correct = False
            if parsed and parsed["answerable"] and label["answerable"]:
                if source == "plotqa":
                    answer, target = number(parsed["answer"]), number(label["target"])
                    correct = answer is not None and target is not None and answer == target
                else:
                    correct = strict_answer_match(parsed["answer"], label["target"])
            fail = invalid or wrong
            state = state_counts[label["state"]]
            for values in (state, counts):
                values["invalid"] += int(invalid)
                values["valid_decision_errors"] += int(wrong)
                values["failures"] += int(fail)
            state["views"] += 1
            counts["valid"] += int(not invalid)
            counts["strict_answers_correct"] += int(correct)
            counts["answerable_views"] += int(label["answerable"])
            counts["unanswerable_views"] += int(not label["answerable"])
            groups[label["group_id"]].append(
                (label["state"], fail, correct or not label["answerable"])
            )
        for rows in groups.values():
            if len(rows) != len(STATES[source]) or {row[0] for row in rows} != set(STATES[source]):
                raise ValueError("Missing or repeated sibling state")
            success = all(not row[1] for row in rows)
            counts["groups_correct"] += int(success)
            counts["joint_groups_correct"] += int(success and all(row[2] for row in rows))
        answerable_failures = sum(
            row["failures"] for state, row in state_counts.items() if not state.startswith("U_")
        )
        unanswerable_failures = counts["failures"] - answerable_failures
        output[source] = {
            "counts": counts,
            "states": state_counts,
            "groups": len(groups),
            "views": len(selected),
            "metrics": {
                "decision_accuracy": 1 - counts["failures"] / len(selected),
                "group_decision_success": counts["groups_correct"] / len(groups),
                "joint_success": counts["joint_groups_correct"] / len(groups),
                "supported_answer_accuracy": counts["strict_answers_correct"]
                / counts["answerable_views"],
                "balanced_failure": (
                    answerable_failures / counts["answerable_views"]
                    + unanswerable_failures / counts["unanswerable_views"]
                )
                / 2,
            },
        }
    return {
        "status": "passed",
        "prediction_rows": len(predictions),
        "by_source": output,
        "scoring": "exact chart rational equality; strict normalized scene answers; invalid outputs fail; all siblings required",
    }


def main():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--dataset", type=Path, default=Path(__file__).resolve().parents[1])
    parser.add_argument("--predictions", type=Path, required=True)
    parser.add_argument("--sources", nargs="+", choices=list(STATES), default=list(STATES))
    parser.add_argument("--output", type=Path, required=True)
    args = parser.parse_args()
    labels = [
        row
        for row in read_rows(args.dataset / "metadata/views.jsonl.gz")
        if row["source"] in args.sources
    ]
    result = score(labels, read_rows(args.predictions))
    write_json(args.output, result)
    print(json.dumps({"status": result["status"], "rows": result["prediction_rows"]}))


if __name__ == "__main__":
    main()