| |
| """Strict offline scoring of versioned answer manifests and saved choices.""" |
|
|
| import argparse |
| import json |
| from pathlib import Path |
|
|
|
|
| DOMAINS = {"mc4": {"A", "B", "C", "D"}, "binary": {"是", "否"}} |
|
|
|
|
| def normalize(value): |
| if not isinstance(value, str): |
| raise ValueError("Answers must be strings") |
| return value.strip().upper() |
|
|
|
|
| def read_rows(path): |
| with Path(path).open(encoding="utf-8") as stream: |
| rows = [json.loads(line) for line in stream if line.strip()] |
| if not rows: |
| raise ValueError("Input must not be empty") |
| indexed = {} |
| for row in rows: |
| item_id = row["item_id"] |
| if not isinstance(item_id, str) or not item_id: |
| raise ValueError("item_id must be a nonempty string") |
| if item_id in indexed: |
| raise ValueError("Duplicate item_id: " + item_id) |
| indexed[item_id] = row |
| return indexed |
|
|
|
|
| def score(manifest_path, predictions_path): |
| manifest = read_rows(manifest_path) |
| predictions = read_rows(predictions_path) |
| if manifest.keys() != predictions.keys(): |
| raise ValueError( |
| "ID coverage mismatch: missing=%d unexpected=%d" |
| % (len(manifest.keys() - predictions.keys()), len(predictions.keys() - manifest.keys())) |
| ) |
| by_task = {} |
| invalid = 0 |
| for item_id, expected in manifest.items(): |
| prediction = predictions[item_id] |
| if expected["task_name"] != prediction["task_name"]: |
| raise ValueError("Task mismatch: " + item_id) |
| domain = DOMAINS[expected["answer_type"]] |
| truth = normalize(expected["ground_truth"]) |
| answer = normalize(prediction["model_answer"]) |
| if truth not in domain: |
| raise ValueError("Ground truth outside answer domain: " + item_id) |
| if answer and answer not in domain: |
| raise ValueError("Prediction outside answer domain: " + item_id) |
| invalid += not answer |
| task = by_task.setdefault(expected["task_name"], {"correct": 0, "total": 0}) |
| task["total"] += 1 |
| task["correct"] += int(answer == truth) |
| for counts in by_task.values(): |
| counts["accuracy_percent"] = 100 * counts["correct"] / counts["total"] |
| correct = sum(counts["correct"] for counts in by_task.values()) |
| return { |
| "correct": correct, |
| "total": len(manifest), |
| "accuracy_percent": 100 * correct / len(manifest), |
| "empty_incorrect_predictions": invalid, |
| "by_task": dict(sorted(by_task.items())), |
| } |
|
|
|
|
| def main(): |
| parser = argparse.ArgumentParser(description=__doc__) |
| parser.add_argument("--manifest", required=True, type=Path) |
| parser.add_argument("--predictions", required=True, type=Path) |
| arguments = parser.parse_args() |
| print(json.dumps(score(arguments.manifest, arguments.predictions), ensure_ascii=False, indent=2)) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
