File size: 3,164 Bytes
765ab6e | 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 | """Accuracy for one results directory (results/<run>/<benchmark>/*.jsonl).
python -m vbe.score results/phase1_qbase_internvid
Writes <run>/scores.json and prints a summary. Unanswered questions (unreadable video,
OOM, no letter in the response) count as wrong and are reported separately.
"""
import glob
import json
import os
import sys
from collections import defaultdict
from vbe.benchmarks import DEFAULT_ROOT, load_benchmark
# per-benchmark breakdown fields; the headline number is in the first entry
BREAKDOWN = {
"videomme": ["duration", "task_type", "domain"],
"mlvu": ["task_type"],
"longvideobench": ["duration_group", "question_category", "level"],
}
def _acc(recs):
return round(100.0 * sum(r["pred"] == r["answer"] for r in recs) / max(len(recs), 1), 2)
def score_benchmark(bench, files, data_root):
recs = {}
for f in files:
for line in open(f):
if line.strip():
r = json.loads(line)
recs[r["qid"]] = r # a resumed run may have re-written a qid; keep the last
expected = {s["qid"] for s in load_benchmark(bench, data_root)}
recs = [r for q, r in recs.items() if q in expected]
res = dict(
n=len(recs),
expected=len(expected),
complete=len(recs) == len(expected),
accuracy=_acc(recs),
no_answer=sum(r["pred"] == "" for r in recs),
errors=sum("error" in r for r in recs),
)
for field in BREAKDOWN[bench]:
by = defaultdict(list)
for r in recs:
by[str(r["meta"][field])].append(r)
res[f"by_{field}"] = {k: dict(n=len(v), acc=_acc(v)) for k, v in sorted(by.items())}
# budget layout the video ended up in (raw / hybrid / forced8), see vbe/budget.py
by = defaultdict(list)
for r in recs:
by[r.get("mode", "error")].append(r)
res["by_mode"] = {k: dict(n=len(v), acc=_acc(v)) for k, v in sorted(by.items())}
if bench == "mlvu":
# MLVU's reported M-Avg: mean of per-task accuracies
tasks = res["by_task_type"].values()
res["m_avg"] = round(sum(t["acc"] for t in tasks) / max(len(tasks), 1), 2)
return res
def main():
run_dir = sys.argv[1]
data_root = sys.argv[2] if len(sys.argv) > 2 else DEFAULT_ROOT
scores = {}
for bench in BREAKDOWN:
files = sorted(glob.glob(os.path.join(run_dir, bench, "*.jsonl")))
if files:
scores[bench] = score_benchmark(bench, files, data_root)
with open(os.path.join(run_dir, "scores.json"), "w") as f:
json.dump(scores, f, indent=2, ensure_ascii=False)
for bench, r in scores.items():
flag = "" if r["complete"] else f" [INCOMPLETE {r['n']}/{r['expected']}]"
head = f"M-Avg {r['m_avg']:.2f} acc {r['accuracy']:.2f}" if bench == "mlvu" else f"acc {r['accuracy']:.2f}"
print(f"== {bench}: {head} (n={r['n']}, no answer {r['no_answer']}, errors {r['errors']}){flag}")
field = BREAKDOWN[bench][0]
print(" " + " ".join(f"{k}: {v['acc']:.2f}" for k, v in r[f"by_{field}"].items()))
print(f"-> {os.path.join(run_dir, 'scores.json')}")
if __name__ == "__main__":
main()
|