File size: 9,937 Bytes
b68f9f3 5ee4f7e b68f9f3 5ee4f7e b68f9f3 5ee4f7e b68f9f3 5ee4f7e b68f9f3 5ee4f7e b68f9f3 5ee4f7e b68f9f3 5ee4f7e b68f9f3 5ee4f7e b68f9f3 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 | from __future__ import annotations
import argparse
import hashlib
import json
import sys
from collections.abc import Sequence
from datetime import UTC, datetime
from pathlib import Path
from typing import Any, TextIO, cast
from .ajimee import (
AJIMEE_LICENSE,
AJIMEE_REPOSITORY,
AJIMEE_REVISION,
AJIMEE_SHA256,
)
from .evaluation import (
EvaluationItem,
Prediction,
PredictionProvenance,
evaluate_predictions,
)
def _parser() -> argparse.ArgumentParser:
parser = argparse.ArgumentParser(prog="deberta-ime-evaluate")
parser.add_argument("--items", type=Path, required=True)
parser.add_argument("--predictions", type=Path, required=True)
parser.add_argument("--format", choices=("generic", "ajimee"), default="generic")
parser.add_argument("--dataset-name", default="unspecified")
parser.add_argument("--dataset-revision", default="unspecified")
parser.add_argument("--dataset-license", default="unspecified")
parser.add_argument("--candidate-limit", type=int, default=8)
parser.add_argument("--output-dir", type=Path, default=Path("outputs"))
parser.add_argument("--stem", default="finite_candidate_evaluation")
return parser
def run(
argv: Sequence[str] | None = None,
*,
stdout: TextIO | None = None,
) -> int:
args = _parser().parse_args(argv)
items = _load_items(args.items, item_format=args.format)
predictions = _load_predictions(args.predictions)
evaluation = evaluate_predictions(
items,
predictions,
candidate_limit=args.candidate_limit,
)
items_artifact = _artifact(args.items)
predictions_artifact = _artifact(args.predictions)
pinned_ajimee = args.format == "ajimee" and items_artifact["sha256"] == AJIMEE_SHA256
report: dict[str, Any] = {
"schema_version": 2,
"generated_at": datetime.now(UTC).isoformat(),
"status": "LOCAL_FINITE_CANDIDATE_EVALUATION",
"format": args.format,
"dataset": (
{
"name": AJIMEE_REPOSITORY,
"revision": AJIMEE_REVISION,
"license": AJIMEE_LICENSE,
}
if pinned_ajimee
else {
"name": "AJIMEE-compatible input",
"revision": "unverified-by-sha256",
"license": "unspecified",
}
if args.format == "ajimee"
else {
"name": args.dataset_name,
"revision": args.dataset_revision,
"license": args.dataset_license,
}
),
"artifacts": {
"items": items_artifact,
"predictions": predictions_artifact,
},
"evaluation": evaluation,
"claim_boundaries": [
"An empty or missing prediction abstains and preserves the input text.",
"Overcorrection is reported only for rows explicitly labelled clean.",
"Candidates are scored as supplied; this command performs no model inference.",
"AJIMEE provenance is pinned only when the item artifact SHA-256 matches.",
"Prediction provenance, reason, and margin are declarations from the "
"supplied artifact.",
],
}
args.output_dir.mkdir(parents=True, exist_ok=True)
json_path = args.output_dir / f"{args.stem}.json"
markdown_path = args.output_dir / f"{args.stem}.md"
json_path.write_text(
json.dumps(report, ensure_ascii=False, indent=2) + "\n",
encoding="utf-8",
)
markdown_path.write_text(_render_markdown(report), encoding="utf-8")
(stdout or sys.stdout).write(
json.dumps(
{"ok": True, "json": str(json_path), "markdown": str(markdown_path)},
ensure_ascii=False,
indent=2,
)
+ "\n"
)
return 0
def _load_items(path: Path, *, item_format: str) -> tuple[EvaluationItem, ...]:
payload = _load_json_array(path)
items: list[EvaluationItem] = []
for row_number, raw_item in enumerate(payload, start=1):
if not isinstance(raw_item, dict):
raise ValueError(f"item row {row_number}: expected an object")
if item_format == "ajimee":
item_id = raw_item.get("index")
input_text = raw_item.get("input")
references = raw_item.get("expected_output")
label: object = "unspecified"
else:
item_id = raw_item.get("id")
input_text = raw_item.get("input")
references = raw_item.get("references")
label = raw_item.get("label", "unspecified")
if not isinstance(item_id, str) or not isinstance(input_text, str):
raise ValueError(f"item row {row_number}: id and input must be strings")
if (
not isinstance(references, list)
or not references
or not all(isinstance(reference, str) for reference in references)
):
raise ValueError(
f"item row {row_number}: references must be a non-empty string list"
)
if label not in ("clean", "typo", "unspecified"):
raise ValueError(f"item row {row_number}: invalid label")
items.append(
EvaluationItem(
item_id=item_id,
input_text=input_text,
references=tuple(cast(list[str], references)),
label=label,
)
)
return tuple(items)
def _load_predictions(path: Path) -> tuple[Prediction, ...]:
payload = _load_json_array(path)
predictions: list[Prediction] = []
for row_number, raw_prediction in enumerate(payload, start=1):
if not isinstance(raw_prediction, dict):
raise ValueError(f"prediction row {row_number}: expected an object")
item_id = raw_prediction.get("id", raw_prediction.get("index"))
candidates = raw_prediction.get("candidates")
if not isinstance(item_id, str):
raise ValueError(f"prediction row {row_number}: id must be a string")
if not isinstance(candidates, list) or not all(
isinstance(candidate, str) and bool(candidate) for candidate in candidates
):
raise ValueError(f"prediction row {row_number}: candidates must be a string list")
provenance = raw_prediction.get("provenance", "unspecified")
if provenance not in ("provider", "deberta", "lfm", "rule", "unspecified"):
raise ValueError(f"prediction row {row_number}: invalid provenance")
reason = raw_prediction.get("reason")
if reason is not None and not isinstance(reason, str):
raise ValueError(f"prediction row {row_number}: reason must be a string")
margin = raw_prediction.get("margin")
if margin is not None and (
isinstance(margin, bool) or not isinstance(margin, (int, float))
):
raise ValueError(f"prediction row {row_number}: margin must be a number")
predictions.append(
Prediction(
item_id=item_id,
candidates=tuple(cast(list[str], candidates)),
provenance=cast(PredictionProvenance, provenance),
reason=reason,
margin=None if margin is None else float(margin),
)
)
return tuple(predictions)
def _load_json_array(path: Path) -> list[object]:
payload: object = json.loads(path.read_text(encoding="utf-8"))
if not isinstance(payload, list):
raise ValueError(f"{path}: expected a JSON array")
return cast(list[object], payload)
def _artifact(path: Path) -> dict[str, str | int]:
data = path.read_bytes()
return {
"path": str(path),
"sha256": hashlib.sha256(data).hexdigest(),
"size_bytes": len(data),
}
def _render_markdown(report: dict[str, Any]) -> str:
dataset = report["dataset"]
evaluation = report["evaluation"]
metrics = evaluation["metrics"]
overcorrection = metrics["overcorrection_rate"]
rendered_overcorrection = (
"not available (no explicit clean rows)"
if overcorrection is None
else f"{overcorrection:.2%}"
)
reported_margin = metrics["mean_reported_margin"]
rendered_margin = (
"not reported"
if reported_margin is None
else f"{reported_margin:.4f} ({metrics['reported_margin_rows']} rows)"
)
provenance = (
", ".join(
f"{name}={count}" for name, count in metrics["declared_provenance_counts"].items()
)
or "not reported"
)
return "\n".join(
[
"# Finite-candidate evaluation",
"",
f"Generated: {report['generated_at']}",
f"Dataset: {dataset['name']} @ {dataset['revision']}",
f"Dataset license/boundary: {dataset['license']}",
f"Rows: {evaluation['rows']}",
f"Candidate limit: {evaluation['candidate_limit']}",
"",
f"- Effective Acc@1: {metrics['effective_acc_at_1']:.2%}",
f"- Candidate Recall@k: {metrics['candidate_recall_at_k']:.2%}",
f"- Mean MinCER: {metrics['mean_min_cer']:.2%}",
f"- Abstention: {metrics['abstention_rate']:.2%}",
f"- Overcorrection: {rendered_overcorrection}",
f"- Accepted candidate misses: {metrics['accepted_candidate_miss_rows']}",
f"- Selection errors: {metrics['selection_error_rows']}",
f"- Declared provenance: {provenance}",
f"- Mean declared margin: {rendered_margin}",
"",
"An abstention preserves the input. No model inference was performed. "
"Provenance/reason/margin are artifact declarations, not independently "
"verified facts.",
"",
]
)
def main() -> None:
raise SystemExit(run())
if __name__ == "__main__":
main()
|