eval.py

"""Score a prompt-safety detector against the promptgate-eval benchmark.

Standard library only. Works offline.

    # against the published rule pack (downloads rules.json once)
    python eval.py --hub NagaYu/promptgate-rules

    # against a local checkout of the rule pack
    python eval.py --rules ../promptgate-rules

    # against the PromptGate Space engine in this repo (needs gradio + pandas)
    python eval.py --engine app --app-dir ..

    # machine readable
    python eval.py --rules ../promptgate-rules --json results.json

Exit code is 0 when the run completes, 1 on a setup error.
"""

from __future__ import annotations

import argparse
import json
import os
import sys
from typing import Any, Callable, Dict, List

LABELS = ("block", "sanitize", "allow")
HERE = os.path.dirname(os.path.abspath(__file__))
DEFAULT_DATA = os.path.join(HERE, "data", "dev.jsonl")


def load_cases(path: str) -> List[Dict[str, Any]]:
    """Guarantees: returns the benchmark rows, raising only if the file is unreadable."""
    rows: List[Dict[str, Any]] = []
    with open(path, "r", encoding="utf-8") as handle:
        for line in handle:
            line = line.strip()
            if line:
                rows.append(json.loads(line))
    return rows


def build_predictor(args: argparse.Namespace) -> Callable[[str], Dict[str, Any]]:
    """Guarantees: returns a text -> assessment callable for the selected engine."""
    if args.engine == "app":
        sys.path.insert(0, os.path.abspath(args.app_dir))
        import app  # type: ignore

        def predict_app(text: str) -> Dict[str, Any]:
            """Guarantees: returns the Space engine's assessment for one prompt."""
            return app.SafetyEngine.assess(text)

        return predict_app

    rules_dir = os.path.abspath(args.rules) if args.rules else os.path.join(
        os.path.dirname(HERE), "promptgate-rules")
    sys.path.insert(0, rules_dir)
    from promptgate_rules import RuleEngine  # type: ignore

    engine = RuleEngine.from_hub(args.hub) if args.hub else RuleEngine.load_default(
        os.path.join(rules_dir, "rules.json"))

    def predict_rules(text: str) -> Dict[str, Any]:
        """Guarantees: returns the standalone rule engine's assessment for one prompt."""
        return engine.assess(text)

    return predict_rules


def score(cases: List[Dict[str, Any]], predict: Callable[[str], Dict[str, Any]]) -> Dict[str, Any]:
    """Guarantees: returns accuracy, per-label PRF, confusion matrix and every miss."""
    confusion = {truth: {pred: 0 for pred in LABELS} for truth in LABELS}
    misses: List[Dict[str, Any]] = []
    by_family: Dict[str, Dict[str, int]] = {}
    by_difficulty: Dict[str, Dict[str, int]] = {}
    category_hits: Dict[str, Dict[str, int]] = {}

    for case in cases:
        report = predict(case["text"])
        predicted = str(report.get("verdict", "allow"))
        truth = str(case["label"])
        if predicted not in LABELS:
            predicted = "allow"
        confusion[truth][predicted] += 1

        fam = by_family.setdefault(case.get("family", "?"), {"n": 0, "correct": 0})
        fam["n"] += 1
        dif = by_difficulty.setdefault(case.get("difficulty", "?"), {"n": 0, "correct": 0})
        dif["n"] += 1
        if predicted == truth:
            fam["correct"] += 1
            dif["correct"] += 1
        else:
            misses.append({
                "id": case["id"], "expected": truth, "predicted": predicted,
                "family": case.get("family"), "difficulty": case.get("difficulty"),
                "lang": case.get("lang"), "notes": case.get("notes", ""),
                "text": case["text"][:110],
                "detected_types": report.get("risk_types", []),
            })

        detected = set(report.get("risk_types", []))
        for expected_type in case.get("categories", []):
            row = category_hits.setdefault(expected_type, {"expected": 0, "detected": 0})
            row["expected"] += 1
            if expected_type in detected:
                row["detected"] += 1

    total = len(cases)
    correct = sum(confusion[label][label] for label in LABELS)
    per_label: Dict[str, Dict[str, float]] = {}
    for label in LABELS:
        tp = confusion[label][label]
        fp = sum(confusion[other][label] for other in LABELS if other != label)
        fn = sum(confusion[label][other] for other in LABELS if other != label)
        precision = tp / (tp + fp) if (tp + fp) else 0.0
        recall = tp / (tp + fn) if (tp + fn) else 0.0
        f1 = (2 * precision * recall / (precision + recall)) if (precision + recall) else 0.0
        per_label[label] = {"precision": round(precision, 4), "recall": round(recall, 4),
                            "f1": round(f1, 4), "support": tp + fn}

    attacks = [c for c in cases if c["family"] == "injection"]
    attacks_flagged = sum(1 for c in attacks
                          if predict(c["text"]).get("verdict") in ("block", "sanitize"))
    benign = [c for c in cases if c["label"] == "allow"]
    benign_flagged = sum(1 for c in benign
                         if predict(c["text"]).get("verdict") != "allow")
    sensitive = [c for c in cases if c["family"] in ("pii", "secret")]
    sensitive_caught = sum(1 for c in sensitive
                           if predict(c["text"]).get("verdict") in ("sanitize", "block"))

    return {
        "cases": total,
        "exact_verdict_accuracy": round(correct / total, 4) if total else 0.0,
        "per_label": per_label,
        "confusion_matrix": confusion,
        "attack_flag_rate": round(attacks_flagged / len(attacks), 4) if attacks else 0.0,
        "benign_false_positive_rate": round(benign_flagged / len(benign), 4) if benign else 0.0,
        "sensitive_redaction_rate": round(sensitive_caught / len(sensitive), 4) if sensitive else 0.0,
        "by_family": {k: {"n": v["n"], "accuracy": round(v["correct"] / v["n"], 4)}
                      for k, v in sorted(by_family.items())},
        "by_difficulty": {k: {"n": v["n"], "accuracy": round(v["correct"] / v["n"], 4)}
                          for k, v in sorted(by_difficulty.items())},
        "category_recall": {k: {"expected": v["expected"], "detected": v["detected"],
                                "recall": round(v["detected"] / v["expected"], 4)}
                            for k, v in sorted(category_hits.items())},
        "misses": misses,
    }


def to_markdown(results: Dict[str, Any], title: str) -> str:
    """Guarantees: renders the results dict as a Markdown report string."""
    lines = ["## {}".format(title), ""]
    lines.append("- cases: **{}**".format(results["cases"]))
    lines.append("- exact verdict accuracy: **{:.1%}**".format(results["exact_verdict_accuracy"]))
    lines.append("- attacks flagged (block or sanitize): **{:.1%}**".format(results["attack_flag_rate"]))
    lines.append("- PII/secret prompts redacted: **{:.1%}**".format(results["sensitive_redaction_rate"]))
    lines.append("- benign false-positive rate: **{:.1%}**".format(results["benign_false_positive_rate"]))
    lines.append("")
    lines.append("| verdict | precision | recall | F1 | support |")
    lines.append("| --- | --- | --- | --- | --- |")
    for label in LABELS:
        row = results["per_label"][label]
        lines.append("| {} | {:.3f} | {:.3f} | {:.3f} | {} |".format(
            label, row["precision"], row["recall"], row["f1"], row["support"]))
    lines.append("")
    lines.append("| family | n | accuracy |")
    lines.append("| --- | --- | --- |")
    for name, row in results["by_family"].items():
        lines.append("| {} | {} | {:.1%} |".format(name, row["n"], row["accuracy"]))
    lines.append("")
    if results["misses"]:
        lines.append("### Misses ({})".format(len(results["misses"])))
        lines.append("")
        lines.append("| id | expected | predicted | why it is hard |")
        lines.append("| --- | --- | --- | --- |")
        for miss in results["misses"]:
            lines.append("| `{}` | {} | {} | {} |".format(
                miss["id"], miss["expected"], miss["predicted"],
                miss["notes"] or miss["text"][:60]))
    else:
        lines.append("No misses.")
    lines.append("")
    return "\n".join(lines)


def main() -> int:
    """Guarantees: runs the benchmark, prints a report, and returns an exit code."""
    parser = argparse.ArgumentParser(description="Evaluate a detector on promptgate-eval.")
    parser.add_argument("--data", default=DEFAULT_DATA, help="path to eval.jsonl")
    parser.add_argument("--engine", choices=["rules", "app"], default="rules",
                        help="'rules' = standalone rule pack, 'app' = PromptGate Space engine")
    parser.add_argument("--rules", default="", help="directory holding rules.json + promptgate_rules.py")
    parser.add_argument("--hub", default="", help="load the rule pack from this Hub repo id")
    parser.add_argument("--app-dir", default="..", help="directory containing app.py (for --engine app)")
    parser.add_argument("--json", default="", help="write the full results JSON here")
    parser.add_argument("--markdown", default="", help="write a Markdown report here")
    args = parser.parse_args()

    try:
        cases = load_cases(args.data)
        predict = build_predictor(args)
    except Exception as exc:
        print("setup failed: {}: {}".format(type(exc).__name__, exc), file=sys.stderr)
        return 1

    results = score(cases, predict)
    title = "promptgate-eval - {} engine".format(args.engine)
    report = to_markdown(results, title)
    print(report)

    if args.json:
        with open(args.json, "w", encoding="utf-8") as handle:
            json.dump(results, handle, ensure_ascii=False, indent=2)
            handle.write("\n")
        print("wrote {}".format(args.json))
    if args.markdown:
        with open(args.markdown, "w", encoding="utf-8") as handle:
            handle.write(report)
        print("wrote {}".format(args.markdown))
    return 0


if __name__ == "__main__":
    raise SystemExit(main())