Datasets:
File size: 10,542 Bytes
1b62586 | 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 | #!/usr/bin/env python3
"""Join document generations with SEGALE/COMET scores and optional groups."""
from __future__ import annotations
import argparse
import json
import statistics
from collections import Counter, defaultdict
from pathlib import Path
STANDARD_METADATA_FIELDS = (
"schema_version",
"benchmark_version",
"track_id",
"split",
"corpus_id",
"work_id",
"official_directory_group",
"length_band",
"target_source_tokens",
"window_family_id",
)
def read_jsonl(path: Path) -> list[dict]:
with path.open(encoding="utf-8") as stream:
return [json.loads(line) for line in stream if line.strip()]
def mean(values) -> float | None:
present = [value for value in values if isinstance(value, (int, float))]
return statistics.fmean(present) if present else None
def group_summary(rows: list[dict]) -> dict:
statuses = Counter(row["generation"]["status"] for row in rows)
scored_cases = sum(row["generation"]["status"] == "ok" for row in rows)
successful = [row for row in rows if row['generation']['status'] == 'ok']
caps = [row['generation']['cap_hit'] for row in successful
if isinstance(row['generation'].get('cap_hit'), bool)]
empties = [row['diagnostics']['empty_output'] for row in successful
if isinstance(row['diagnostics'].get('empty_output'), bool)]
finishes = Counter(row['generation']['finish_reason'] for row in successful
if isinstance(row['generation'].get('finish_reason'), str))
diagnostic_fields = ('null_source_char_ratio', 'null_hypothesis_char_ratio',
'exact_duplicate_sentence_ratio')
return {
"cases": len(rows),
"scored_cases": scored_cases,
"failed_cases": len(rows) - scored_cases,
"generation_status_counts": dict(sorted(statuses.items())),
"segale_comet": mean(row.get("segale_comet") for row in rows),
"aligned_only_comet": mean(row.get("aligned_only_comet") for row in rows),
"na_ratio": mean(row.get("na_ratio") for row in rows),
"hypothesis_reference_char_ratio": mean(
row.get("hypothesis_reference_char_ratio") for row in rows
),
"diagnostics": {
"case_mean": {key: mean(row['diagnostics'].get(key) for row in rows)
for key in diagnostic_fields},
"observed_cases": {key: sum(row['diagnostics'].get(key) is not None for row in rows)
for key in diagnostic_fields},
"under_translation_nulls": sum(row.get('under_translation_nulls') or 0 for row in rows),
"over_translation_nulls": sum(row.get('over_translation_nulls') or 0 for row in rows),
"null_count_observed_cases": sum(row.get('under_translation_nulls') is not None
and row.get('over_translation_nulls') is not None for row in rows),
"cap_observed_cases": len(caps),
"cap_hit_cases": sum(caps),
"cap_hit_ratio": mean(caps),
"empty_output_observed_cases": len(empties),
"empty_output_cases": sum(empties),
"empty_output_ratio": mean(empties),
"finish_reason_counts": dict(sorted(finishes.items())),
},
}
def case_metadata(case: dict) -> dict:
metadata = case.get("metadata", {})
if not isinstance(metadata, dict):
raise ValueError(f"Case {case.get('case_id')} metadata must be an object")
result = dict(metadata)
for key in STANDARD_METADATA_FIELDS:
if key in case:
result[key] = case[key]
return result
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--cases", type=Path, required=True)
parser.add_argument("--generations", type=Path, required=True)
parser.add_argument("--comet-summary", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
parser.add_argument("--suite-id", required=True)
parser.add_argument("--system-key", required=True)
parser.add_argument("--group-by", action="append", default=[])
parser.add_argument("--chrf-dir", type=Path)
args = parser.parse_args()
cases = read_jsonl(args.cases)
generations = read_jsonl(args.generations)
comet = json.loads(args.comet_summary.read_text(encoding="utf-8"))
cases_by_id = {row["case_id"]: row for row in cases}
generations_by_id = {row["case_id"]: row for row in generations}
scores_by_id = {row["case_id"]: row for row in comet["cases"]}
if len(cases_by_id) != len(cases) or len(generations_by_id) != len(generations):
raise ValueError("Duplicate case_id")
expected = set(cases_by_id)
unexpected_generations = set(generations_by_id) - expected
if unexpected_generations:
raise ValueError(f"Generation contains unknown cases: {sorted(unexpected_generations)}")
expected_scores = {
case_id
for case_id, generation in generations_by_id.items()
if generation.get("status", "ok") == "ok"
}
if set(scores_by_id) != expected_scores:
raise ValueError("Score coverage differs from successful generations")
rows = []
statuses = Counter()
for case in cases:
case_id = case["case_id"]
generation = generations_by_id.get(case_id)
status = "missing" if generation is None else generation.get("status", "ok")
if not isinstance(status, str) or not status:
raise ValueError(f"Generation {case_id} has an invalid status")
score = scores_by_id.get(case_id, {})
statuses[status] += 1
rows.append(
{
"case_id": case_id,
"metadata": case_metadata(case),
"generation": {
key: (generation or {}).get(key)
for key in (
"status",
"finish_reason",
"input_tokens",
"output_tokens",
"cap_hit",
"model_revision",
"generation_config_sha256",
"output_sha256",
)
} | {"status": status},
"segale_comet": score.get("comet"),
"aligned_only_comet": score.get("comet_aligned_only"),
"na_ratio": score.get("na_ratio"),
"hypothesis_reference_char_ratio": score.get(
"hypothesis_reference_char_ratio"
),
"under_translation_nulls": score.get("under_translation_nulls"),
"over_translation_nulls": score.get("over_translation_nulls"),
"position_buckets": score.get("position_buckets"),
"diagnostics": {
**score.get('diagnostics', {}),
"exact_duplicate_sentence_ratio": score.get('exact_duplicate_sentence_ratio'),
"empty_output": (not generation['mt'].strip())
if generation and isinstance(generation.get('mt'), str) else None,
},
}
)
grouped = {}
for field in args.group_by:
buckets = defaultdict(list)
for row in rows:
value = row["metadata"].get(field)
if isinstance(value, (dict, list)):
raise ValueError(f"Grouping field {field} must be a scalar")
label = "__missing__" if value is None else str(value)
buckets[label].append(row)
grouped[field] = {
label: group_summary(bucket) for label, bucket in sorted(buckets.items())
}
scored_case_count = statuses["ok"]
result = {
"schema_version": "document-segale-summary-v1",
"suite_id": args.suite_id,
"system_key": args.system_key,
"case_count": len(rows),
"scored_case_count": scored_case_count,
"failed_case_count": len(rows) - scored_case_count,
"generation_status_counts": dict(sorted(statuses.items())),
"scored": group_summary(rows),
"group_by": args.group_by,
"groups": grouped,
"cases": rows,
}
if args.chrf_dir:
from score_document_chrf import aggregate, sha_file
completed = json.loads((args.chrf_dir / "COMPLETED.json").read_text())
artifact_path = args.chrf_dir / "artifact-manifest.json"
if completed["artifact_manifest_sha256"] != sha_file(artifact_path):
raise ValueError("chrF2 artifact manifest hash mismatch")
artifacts = json.loads(artifact_path.read_text())
for name in ("summary.json", "cases.jsonl"):
if artifacts[name] != sha_file(args.chrf_dir / name):
raise ValueError("chrF2 artifact hash mismatch")
auxiliary = json.loads((args.chrf_dir / "summary.json").read_text())
chrf_rows = read_jsonl(args.chrf_dir / "cases.jsonl")
if auxiliary["suite_id"] != args.suite_id or auxiliary["system_key"] != args.system_key:
raise ValueError("chrF2 run identity mismatch")
for name, path in (("cases", args.cases), ("generations", args.generations)):
if auxiliary["inputs"][name]["sha256"] != sha_file(path):
raise ValueError("chrF2 input hash mismatch")
if [r["case_id"] for r in chrf_rows] != [r["case_id"] for r in rows]:
raise ValueError("chrF2 case coverage/order mismatch")
for row, extra in zip(rows, chrf_rows, strict=True):
if row["generation"]["status"] != extra["generation_status"]:
raise ValueError("chrF2 generation status mismatch")
row["auxiliary_metrics"] = {"chrf2": extra}
result["auxiliary_metrics"] = {"chrf2": {
"metric": auxiliary["metric"], "aggregate": aggregate(chrf_rows),
"artifact": "chrf2/summary.json",
}}
for field, buckets in grouped.items():
for label, bucket in buckets.items():
subset = [r for r in chrf_rows
if ("__missing__" if r["metadata"].get(field) is None
else str(r["metadata"][field])) == label]
bucket["auxiliary_metrics"] = {"chrf2": aggregate(subset)}
args.output.write_text(
json.dumps(result, ensure_ascii=False, indent=2) + "\n", encoding="utf-8"
)
print(f"Summarized {len(rows)} document cases")
if __name__ == "__main__":
main()
|