File size: 6,940 Bytes
f91d9a0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
"""Finetune-style monitor-next delta task."""

from __future__ import annotations

from collections import Counter
from pathlib import Path
from typing import Any

from ..io import BENCHMARK_ROOT, DATA_DIR, manifest_examples, select_shard
from ..metrics import bootstrap_sem, classification_scores
from ..parsing import normalize_step_id, parse_next_step_delta_response
from .base import BenchmarkTask, TaskSpec


class MonitorNextStepTask(BenchmarkTask):
    spec = TaskSpec(
        name="monitor_next_step",
        display_name="Protocol Monitoring Advance Prediction",
        default_manifest=DATA_DIR / "monitor_next_step.json",
    )
    primary_metric = "advance_step_balanced_accuracy"

    def load_examples(
        self,
        *,
        benchmark_root: Path = BENCHMARK_ROOT,
        manifest_path: Path | None = None,
        video_root: Path | None = None,
        num_shards: int = 1,
        shard_index: int = 0,
        limit: int | None = None,
    ) -> list[dict[str, Any]]:
        rows = manifest_examples(manifest_path or self.default_manifest)
        root = video_root or benchmark_root
        for row in rows:
            row["task"] = self.name
            row["_video_abs"] = str(root / str(row["video_path"]))
        rows = sorted(rows, key=lambda row: str(row.get("eval_id")))
        if limit is not None:
            rows = rows[:limit]
        return select_shard(rows, num_shards, shard_index)

    def parse_record(self, row: dict[str, Any]) -> dict[str, Any]:
        if row.get("error"):
            return row
        parsed = parse_next_step_delta_response(str(row.get("raw_response") or ""))
        pred_current_norm = normalize_step_id(parsed.get("pred_current_step_id"))
        target_current_norm = normalize_step_id(row.get("target_current_step_id"))
        target_start_norm = normalize_step_id(row.get("target_start_step_id"))
        pred_advances = None
        if pred_current_norm is not None and target_start_norm is not None:
            pred_advances = pred_current_norm != target_start_norm
        row.update(
            {
                **parsed,
                "pred_current_step_id_normalized": pred_current_norm,
                "target_current_step_id_normalized": target_current_norm,
                "pred_advances_step": pred_advances,
                "current_step_correct": pred_current_norm is not None and pred_current_norm == target_current_norm,
            }
        )
        return row

    def _summary(self, rows: list[dict[str, Any]]) -> dict[str, Any]:
        parsed = self.parse_rows(rows)
        scored = [row for row in parsed if not row.get("error")]
        current_correct = 0
        current_parsed = 0
        advance_pairs: list[tuple[str, str | None]] = []
        skipped_pairs: list[tuple[str, str | None]] = []
        skipped_step_ref_correct = 0
        skipped_step_ref_total = 0
        for row in scored:
            pred_current = normalize_step_id(row.get("pred_current_step_id"))
            target_current = normalize_step_id(row.get("target_current_step_id"))
            target_start = normalize_step_id(row.get("target_start_step_id"))
            if pred_current is not None:
                current_parsed += 1
            current_correct += int(pred_current is not None and pred_current == target_current)
            pred_advances = None
            if pred_current is not None and target_start is not None:
                pred_advances = pred_current != target_start
            advance_pairs.append((str(bool(row.get("target_advances_step"))), None if pred_advances is None else str(bool(pred_advances))))
            skipped_pred = row.get("pred_has_step_skipped")
            skipped_pairs.append((str(bool(row.get("target_has_error"))), None if skipped_pred is None else str(bool(skipped_pred))))
            if row.get("target_has_error"):
                skipped_step_ref_total += 1
                pred_ref = normalize_step_id(row.get("pred_error_step_id"))
                target_ref = normalize_step_id(row.get("target_error_step_id"))
                skipped_step_ref_correct += int(pred_ref is not None and pred_ref == target_ref)
        advance = classification_scores(advance_pairs, labels=["False", "True"])
        skipped = classification_scores(skipped_pairs, labels=["False", "True"])
        n = len(scored)
        parse_errors = sum(1 for row in scored if not row.get("pred_parse_ok"))
        return {
            "task": self.name,
            "display_name": self.display_name,
            "rows": len(rows),
            "scored": n,
            "errors": sum(1 for row in rows if row.get("error")),
            "parse_errors": parse_errors,
            "parse_success_rate": (n - parse_errors) / n if n else None,
            "current_step_accuracy": current_correct / n if n else None,
            "current_step_accuracy_parsed_only": current_correct / current_parsed if current_parsed else None,
            "advance_step_accuracy": advance["accuracy"],
            "advance_step_balanced_accuracy": advance["balanced_accuracy"],
            "advance_step_macro_f1": advance["macro_f1"],
            "advance_step_macro_precision": advance["macro_precision"],
            "advance_step_macro_recall": advance["macro_recall"],
            "advance_step_precision": advance["precision"],
            "advance_step_recall": advance["recall"],
            "advance_step_f1": advance["f1"],
            "advance_step_confusion": advance["confusion"],
            "skipped_step_accuracy": skipped["accuracy"],
            "skipped_step_balanced_accuracy": skipped["balanced_accuracy"],
            "skipped_step_macro_f1": skipped["macro_f1"],
            "skipped_step_macro_precision": skipped["macro_precision"],
            "skipped_step_macro_recall": skipped["macro_recall"],
            "skipped_step_confusion": skipped["confusion"],
            "skipped_step_ref_accuracy": skipped_step_ref_correct / skipped_step_ref_total if skipped_step_ref_total else None,
            "task_counts": dict(Counter(str(row.get("task_type")) for row in scored)),
            "history_variant_counts": dict(Counter(str(row.get("history_variant")) for row in scored)),
            "window_kind_counts": dict(Counter(str(row.get("window_kind")) for row in scored)),
        }

    def score(self, rows: list[dict[str, Any]]) -> dict[str, Any]:
        parsed = self.parse_rows(rows)
        summary = self._summary(parsed)
        for metric in (
            "advance_step_accuracy",
            "advance_step_balanced_accuracy",
            "advance_step_macro_f1",
            "advance_step_macro_precision",
            "advance_step_macro_recall",
            "skipped_step_accuracy",
            "skipped_step_balanced_accuracy",
            "skipped_step_macro_f1",
        ):
            summary[f"{metric}_sem"] = bootstrap_sem(parsed, metric, self._summary, iterations=200)
        return summary