Datasets:
File size: 10,320 Bytes
0041cfd | 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 | #!/usr/bin/env python3
"""Score hypotheses from run_benchmark.py against the benchmark annotations.
python score_benchmark.py --hyps runs/deepgram_nova3.jsonl
Reports span-level entity accuracy (the headline metric), a per-type and
per-subset breakdown, and WER. Writes a per-span hits CSV so two runs can be
diffed clip by clip instead of argued about as a single percentage.
"""
from __future__ import annotations
import argparse
import csv
import json
import os
import sys
from collections import defaultdict
from datetime import datetime, timezone
import common
import matcher_adapter
def parse_args(argv=None):
p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
p.add_argument("--hyps", required=True, help="JSONL produced by run_benchmark.py")
p.add_argument("--out-prefix", default=None, help="default: alongside --hyps")
p.add_argument("--dataset", default=common.DEFAULT_DATASET)
p.add_argument("--split", default=common.DEFAULT_SPLIT)
p.add_argument("--revision", default=None)
p.add_argument("--local-path", default=None)
p.add_argument("--tier", default="A",
help="entity spans of this tier are scored ('A' is the published protocol; "
"'all' scores every span). WER always uses every scored clip.")
p.add_argument("--matcher", default=None, help="module:callable override")
p.add_argument("--matcher-path", default=None, help="path to matcher_v5.py")
p.add_argument("--wer-normalizer", default="basic", choices=["basic", "whisper_english", "none"])
p.add_argument("--no-wer", action="store_true")
p.add_argument("--compare", default=None, help="another hits CSV to diff against")
for field in common.CANDIDATES:
p.add_argument(f"--{field.replace('_', '-')}-column", dest=f"{field}_column", default=None)
return p.parse_args(argv)
def read_manifest(hyps_path: str) -> dict:
path = os.path.splitext(hyps_path)[0] + ".manifest.json"
if os.path.exists(path):
with open(path) as fh:
return json.load(fh)
return {}
def read_hyps(path: str) -> tuple[dict[str, str], int]:
table: dict[str, str] = {}
failed = 0
with open(path) as fh:
for line in fh:
line = line.strip()
if not line:
continue
record = json.loads(line)
if record.get("hypothesis") is None:
failed += 1
continue
table[record["clip_id"]] = record["hypothesis"]
return table, failed
def percent(hits: int, total: int) -> str:
return f"{100.0 * hits / total:.1f}%" if total else "n/a"
def main(argv=None) -> int:
args = parse_args(argv)
manifest = read_manifest(args.hyps)
tier = args.tier or "A"
revision = args.revision or manifest.get("revision")
dataset = args.dataset
local_path = args.local_path
if not local_path and str(manifest.get("dataset", "")).startswith("local:"):
local_path = manifest["dataset"].split("local:", 1)[1]
elif manifest.get("dataset") and args.dataset == common.DEFAULT_DATASET:
dataset = manifest["dataset"]
matcher, matcher_desc = matcher_adapter.load_matcher(args.matcher, args.matcher_path)
common.eprint(f"Matcher: {matcher_desc}")
passed, total_tests = matcher_adapter.self_test(args.matcher_path)
if total_tests:
common.eprint(f"Matcher self-test: {passed}/{total_tests}")
if passed != total_tests:
common.eprint("! self-test failing -- numbers from this run are not comparable")
return 2
hyps, failed = read_hyps(args.hyps)
common.eprint(f"Hypotheses: {len(hyps)} usable, {failed} failed")
ds = common.load_bench(dataset, args.split, revision, local_path)
overrides = {f: getattr(args, f"{f}_column") for f in common.CANDIDATES}
schema = common.resolve_schema(ds.column_names, overrides)
ds = common.undecode_audio(ds, schema) # ids only; never decode audio to score
want_tier = None if not tier or tier.lower() == "all" else tier.strip().upper()
rows: list[dict] = []
by_type: dict[str, list[int]] = defaultdict(lambda: [0, 0])
by_subset: dict[str, list[int]] = defaultdict(lambda: [0, 0])
hits = spans = 0
missing_clips = 0
refs: list[str] = []
preds: list[str] = []
for index, row in enumerate(ds):
clip = common.row_id(row, schema, index)
if clip not in hyps:
missing_clips += 1
continue
hypothesis = hyps[clip]
entities = common.parse_list_field(row[schema.entities])
types = common.parse_list_field(row[schema.entity_types]) if schema.entity_types else []
subset = str(row[schema.subset]) if schema.subset else ""
if schema.reference and not args.no_wer:
refs.append(str(row[schema.reference]))
preds.append(hypothesis)
tiers = common.span_tiers(row, schema, len(entities))
for position, entity in enumerate(entities):
if want_tier is not None and tiers[position] != want_tier:
continue
etype = types[position] if position < len(types) else "UNKNOWN"
hit = bool(matcher(entity, hypothesis))
spans += 1
hits += hit
by_type[etype][0] += hit
by_type[etype][1] += 1
if subset:
by_subset[subset][0] += hit
by_subset[subset][1] += 1
rows.append({
"clip_id": clip,
"entity": entity,
"entity_type": etype,
"subset": subset,
"tier": tiers[position],
"hit": int(hit),
"hypothesis": hypothesis,
})
if missing_clips:
common.eprint(f"! {missing_clips} clips in the tier had no hypothesis -- run is incomplete")
wer = None
if refs and not args.no_wer:
import jiwer
normalize = common.get_normalizer(args.wer_normalizer)
pairs = [(normalize(r), normalize(p)) for r, p in zip(refs, preds)]
pairs = [(r, p) for r, p in pairs if r]
if pairs:
wer = jiwer.wer([r for r, _ in pairs], [p for _, p in pairs])
prefix = args.out_prefix or os.path.splitext(args.hyps)[0]
hits_csv = f"{prefix}.hits.csv"
with open(hits_csv, "w", newline="") as fh:
writer = csv.DictWriter(
fh, fieldnames=["clip_id", "entity", "entity_type", "subset", "tier", "hit", "hypothesis"])
writer.writeheader()
writer.writerows(rows)
summary = {
"scored_utc": datetime.now(timezone.utc).isoformat(timespec="seconds"),
"dataset": dataset if not local_path else f"local:{local_path}",
"revision": revision,
"tier": tier,
"wer_scope": "all clips with a hypothesis (tier filter applies to entity spans only)",
"system_name": manifest.get("system_name") or manifest.get("provider"),
"provider": manifest.get("provider"),
"provider_opts": manifest.get("provider_opts"),
"matcher": matcher_desc,
"matcher_self_test": f"{passed}/{total_tests}" if total_tests else "not exposed",
"wer_normalizer": None if args.no_wer else args.wer_normalizer,
"clips_scored": len({r["clip_id"] for r in rows}),
"clips_with_hypothesis": len(hyps),
"clips_missing": missing_clips,
"clips_failed": failed,
"entity_spans": spans,
"entity_hits": hits,
"entity_accuracy": round(hits / spans, 4) if spans else None,
"wer": round(wer, 4) if wer is not None else None,
"by_type": {k: {"hits": v[0], "spans": v[1], "accuracy": round(v[0] / v[1], 4)} for k, v in sorted(by_type.items())},
"by_subset": {k: {"hits": v[0], "spans": v[1], "accuracy": round(v[0] / v[1], 4)} for k, v in sorted(by_subset.items())},
}
summary_path = f"{prefix}.summary.json"
with open(summary_path, "w") as fh:
json.dump(summary, fh, indent=2)
print()
print(f"Entity Transcription Benchmark -- entity spans: tier {tier}")
print(f" system {summary['system_name'] or '?'}")
print(f" transport {manifest.get('provider', '?')} {manifest.get('provider_opts', '')}")
print(f" clips w/ spans {summary['clips_scored']} of {len(hyps)} transcribed")
print(f" entity accuracy {percent(hits, spans)} ({hits}/{spans} spans)")
if wer is not None:
print(f" WER {100 * wer:.2f}% (normalizer: {args.wer_normalizer})")
print()
print(f" {'type':<16}{'acc':>8}{'spans':>9}")
for etype, (h, t) in sorted(by_type.items(), key=lambda kv: -kv[1][1]):
print(f" {etype:<16}{percent(h, t):>8}{t:>9}")
if by_subset:
print()
print(f" {'subset':<16}{'acc':>8}{'spans':>9}")
for name, (h, t) in sorted(by_subset.items(), key=lambda kv: -kv[1][1]):
print(f" {name:<16}{percent(h, t):>8}{t:>9}")
if args.compare:
diff_path = f"{prefix}.diff.csv"
compare(rows, args.compare, diff_path)
print()
print(f"Per-span hits: {hits_csv}")
print(f"Summary: {summary_path}")
return 0
def compare(rows: list[dict], other_path: str, out_path: str) -> None:
key = lambda r: (r["clip_id"], r["entity"]) # noqa: E731
mine = {key(r): int(r["hit"]) for r in rows}
with open(other_path) as fh:
theirs = {key(r): int(r["hit"]) for r in csv.DictReader(fh)}
shared = sorted(set(mine) & set(theirs))
disagree = [k for k in shared if mine[k] != theirs[k]]
with open(out_path, "w", newline="") as fh:
writer = csv.writer(fh)
writer.writerow(["clip_id", "entity", "this_run", "reference_run"])
for k in disagree:
writer.writerow([k[0], k[1], mine[k], theirs[k]])
print()
print(f" compared against {os.path.basename(other_path)}")
print(f" shared spans {len(shared)}")
print(f" disagreements {len(disagree)} ({percent(len(disagree), len(shared))})")
print(f" only here {len(set(mine) - set(theirs))}")
print(f" only reference {len(set(theirs) - set(mine))}")
print(f" diff: {out_path}")
if __name__ == "__main__":
sys.exit(main())
|