File size: 8,583 Bytes
f9bc0c1 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 | """Oracle analysis over benchmark reports."""
from __future__ import annotations
import json
from dataclasses import dataclass
from pathlib import Path
from typing import Any
from .benchmark import (
BenchmarkResult,
BenchmarkSummary,
_load_benchmark_rows,
summarize_benchmark_report,
)
from .errors import BenchmarkError
from .validation import metric_direction, sample_key
@dataclass(slots=True)
class OracleReport:
benchmark_report: str
oracle_name: str
metric: str
direction: str
systems: list[str]
sample_count: int
chosen_counts: dict[str, int]
oracle_report_path: str
oracle_summary_path: str
oracle_choices_path: str
def to_dict(self) -> dict[str, Any]:
return {
"benchmark_report": self.benchmark_report,
"oracle_name": self.oracle_name,
"metric": self.metric,
"direction": self.direction,
"systems": self.systems,
"sample_count": self.sample_count,
"chosen_counts": self.chosen_counts,
"oracle_report_path": self.oracle_report_path,
"oracle_summary_path": self.oracle_summary_path,
"oracle_choices_path": self.oracle_choices_path,
}
def build_benchmark_oracle(
report_path: str | Path,
output_dir: str | Path,
*,
metric: str = "cer",
systems: list[str] | None = None,
direction: str | None = None,
oracle_name: str = "oracle",
) -> tuple[OracleReport, BenchmarkSummary]:
rows = _load_benchmark_rows(report_path)
if not rows:
raise BenchmarkError("benchmark report has no rows")
compare_systems = systems or sorted({str(row.get("baseline") or "") for row in rows if str(row.get("baseline") or "")})
compare_systems = [name for name in compare_systems if name]
if len(compare_systems) < 2:
raise BenchmarkError("benchmark oracle requires at least two systems")
missing = [name for name in compare_systems if not any(row.get("baseline") == name for row in rows)]
if missing:
raise BenchmarkError(f"requested systems missing from benchmark report: {', '.join(missing)}")
target_direction = direction or metric_direction(metric)
oracle_rows, chosen_counts, choices = _select_oracle_rows(rows, compare_systems, metric, target_direction, oracle_name)
if not oracle_rows:
raise BenchmarkError(f"benchmark oracle found no shared comparable rows for metric {metric!r}")
out = Path(output_dir)
out.mkdir(parents=True, exist_ok=True)
combined_rows = [dict(row) for row in rows] + [result.to_dict() for result in oracle_rows]
oracle_report_path = out / "oracle-report.json"
oracle_report_path.write_text(json.dumps(combined_rows, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
oracle_choices_path = out / "oracle-choices.json"
oracle_choices_path.write_text(json.dumps(choices, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
summary = summarize_benchmark_report(
oracle_report_path,
out / "oracle-summary",
candidate=oracle_name,
baselines=compare_systems,
metrics=[metric],
)
report = OracleReport(
benchmark_report=str(Path(report_path)),
oracle_name=oracle_name,
metric=metric,
direction=target_direction,
systems=compare_systems,
sample_count=len(oracle_rows),
chosen_counts=chosen_counts,
oracle_report_path=str(oracle_report_path),
oracle_summary_path=str(out / "oracle-summary" / "summary.json"),
oracle_choices_path=str(oracle_choices_path),
)
(out / "oracle-analysis.json").write_text(json.dumps(report.to_dict(), ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
lines = [
"# Benchmark Oracle",
"",
f"Source report: `{report_path}`",
f"Oracle: `{oracle_name}`",
f"Metric: `{metric}`",
f"Direction: `{target_direction}`",
f"Samples: `{len(oracle_rows)}`",
"",
"## Chosen Counts",
"",
"| system | chosen |",
"| --- | ---: |",
]
for system, count in sorted(chosen_counts.items()):
lines.append(f"| {system} | {count} |")
lines.extend(
[
"",
f"Oracle report: `{oracle_report_path}`",
f"Oracle choices: `{oracle_choices_path}`",
f"Oracle summary: `{out / 'oracle-summary' / 'summary.json'}`",
]
)
(out / "oracle-analysis.md").write_text("\n".join(lines) + "\n", encoding="utf-8")
return report, summary
def _select_oracle_rows(
rows: list[dict[str, Any]],
systems: list[str],
metric: str,
direction: str,
oracle_name: str,
) -> tuple[list[BenchmarkResult], dict[str, int], list[dict[str, Any]]]:
rows_by_sample: dict[str, dict[str, dict[str, Any]]] = {}
for row in rows:
if row.get("status") != "ok":
continue
baseline = str(row.get("baseline") or "")
if baseline not in systems:
continue
metrics = row.get("metrics")
if not isinstance(metrics, dict) or metric not in metrics:
continue
sample = sample_key(row)
rows_by_sample.setdefault(sample, {})[baseline] = row
oracle_rows: list[BenchmarkResult] = []
chosen_counts = {system: 0 for system in systems}
choices: list[dict[str, Any]] = []
for sample, sample_rows in sorted(rows_by_sample.items()):
if any(system not in sample_rows for system in systems):
continue
chosen_system, chosen_row = _best_row(sample_rows, systems, metric, direction)
chosen_counts[chosen_system] += 1
oracle_row = dict(chosen_row)
oracle_row["baseline"] = oracle_name
oracle_row["error"] = None
metrics = dict(chosen_row.get("metrics") or {})
oracle_row["metrics"] = metrics
choices.append(
{
"sample_key": sample,
"sample_id": chosen_row.get("sample_id"),
"input_path": chosen_row.get("input_path"),
"chosen_system": chosen_system,
"metric": metric,
"metric_value": float(metrics[metric]),
"direction": direction,
}
)
oracle_result = BenchmarkResult(
baseline=str(oracle_row["baseline"]),
input_path=str(oracle_row.get("input_path") or ""),
status=str(oracle_row.get("status") or "ok"),
latency_seconds=float(oracle_row.get("latency_seconds") or 0.0),
sample_id=oracle_row.get("sample_id"),
slices=list(oracle_row.get("slices") or []),
metrics=metrics,
error=None,
)
oracle_result.slices = list(oracle_row.get("slices") or [])
oracle_rows.append(oracle_result)
return oracle_rows, {system: count for system, count in chosen_counts.items() if count}, choices
def _best_row(
sample_rows: dict[str, dict[str, Any]],
systems: list[str],
metric: str,
direction: str,
) -> tuple[str, dict[str, Any]]:
best_system = systems[0]
best_row = sample_rows[best_system]
best_value = float((best_row.get("metrics") or {})[metric])
for system in systems[1:]:
row = sample_rows[system]
value = float((row.get("metrics") or {})[metric])
if _is_better(value, best_value, direction):
best_system = system
best_row = row
best_value = value
continue
if value == best_value:
if _tie_break_quality(row) > _tie_break_quality(best_row):
best_system = system
best_row = row
best_value = value
elif _tie_break_quality(row) == _tie_break_quality(best_row):
row_latency = float(row.get("latency_seconds") or 0.0)
best_latency = float(best_row.get("latency_seconds") or 0.0)
if row_latency < best_latency:
best_system = system
best_row = row
best_value = value
return best_system, best_row
def _is_better(candidate: float, incumbent: float, direction: str) -> bool:
if direction == "lower":
return candidate < incumbent
return candidate > incumbent
def _tie_break_quality(row: dict[str, Any]) -> float:
metrics = row.get("metrics") or {}
value = metrics.get("quality_score")
if isinstance(value, int | float):
return float(value)
return float("-inf")
|