| """Scores predictions on the benchmark: accuracy and expected calibration error (ECE) per task. |
| |
| predictions/<task>.jsonl: one line per item, {"source_id": ..., "probabilities": [p_option_1, p_option_2, ...]} |
| in the item's option order. python evaluate.py predictions/ |
| """ |
| import json |
| import sys |
| from pathlib import Path |
|
|
| CALIBRATION_BIN_EDGES = [0.0, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0] |
|
|
|
|
| def load(path): |
| with path.open(encoding="utf-8") as lines: |
| return [json.loads(line) for line in lines if line.strip()] |
|
|
|
|
| def expected_calibration_error(confidences, correct): |
| error = 0.0 |
| for low, high in zip(CALIBRATION_BIN_EDGES, CALIBRATION_BIN_EDGES[1:]): |
| in_bin = [index for index, confidence in enumerate(confidences) if low <= confidence < high or confidence == high == 1.0] |
| if in_bin: |
| accuracy = sum(correct[index] for index in in_bin) / len(in_bin) |
| confidence = sum(confidences[index] for index in in_bin) / len(in_bin) |
| error += len(in_bin) / len(confidences) * abs(accuracy - confidence) |
| return error |
|
|
|
|
| def main(): |
| predictions_directory = Path(sys.argv[1]) |
| print(f"{'task':12s} {'items':>6s} {'accuracy':>9s} {'ECE':>6s}") |
| for data_path in sorted((Path(__file__).parent / "data").glob("*.jsonl")): |
| predictions_path = predictions_directory / data_path.name |
| if not predictions_path.exists(): |
| continue |
| items = load(data_path) |
| predictions = {prediction["source_id"]: prediction["probabilities"] for prediction in load(predictions_path)} |
| confidences, correct = [], [] |
| for item in items: |
| if item["source_id"] not in predictions: |
| raise ValueError(f"{predictions_path}: no prediction for {item['source_id']}") |
| probabilities = predictions[item["source_id"]] |
| if len(probabilities) != len(item["options"]): |
| raise ValueError(f"{item['source_id']}: {len(probabilities)} probabilities for {len(item['options'])} options") |
| choice = max(range(len(probabilities)), key=probabilities.__getitem__) |
| confidences.append(probabilities[choice]) |
| correct.append(choice == item["answer_index"]) |
| print(f"{data_path.stem:12s} {len(items):6d} {sum(correct) / len(items):9.3f} {expected_calibration_error(confidences, correct):6.3f}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
