File size: 4,294 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 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 | """Paired comparison of ablation runs (vbe/ablate.py) against the baseline.
python -m vbe.ablate_report results/ablate results/phase3_qbase_full_hybrid_lr1e-5 [out.json]
For every (variant, budget, benchmark): accuracy, the difference to the baseline on the
same questions with a 95% interval from resampling videos (2,000 draws), the same on the
questions the variant actually changed (``reused: false``) and on the long-video subset
(Video-MME long, LongVideoBench >= 3 min, all of MLVU), plus MLVU M-Avg.
"""
import glob
import json
import os
import re
import sys
from collections import defaultdict
import numpy as np
LONG = {"videomme": lambda m: m["duration"] == "long",
"longvideobench": lambda m: str(m["duration_group"]) in ("600", "3600"),
"mlvu": lambda m: True}
def load(run_dir, bench):
recs = {}
for f in sorted(glob.glob(os.path.join(run_dir, bench, "*.jsonl"))):
for line in open(f):
if line.strip():
r = json.loads(line)
recs[r["qid"]] = r
return recs
def paired(var, base, qids, draws=2000, seed=0):
"""(variant acc, baseline acc, delta, lo, hi, n) on ``qids``; interval over videos."""
if not qids:
return None
by_vid = defaultdict(lambda: [0.0, 0])
a = b = 0
for q in qids:
ca, cb = var[q]["pred"] == var[q]["answer"], base[q]["pred"] == base[q]["answer"]
a += ca; b += cb
v = by_vid[base[q]["video"]]
v[0] += ca - cb; v[1] += 1
d = np.array([x[0] for x in by_vid.values()]); n = np.array([x[1] for x in by_vid.values()])
rng = np.random.default_rng(seed)
idx = rng.integers(0, len(d), (draws, len(d)))
boot = 100 * d[idx].sum(1) / n[idx].sum(1)
k = len(qids)
return dict(acc=100 * a / k, base=100 * b / k, delta=100 * (a - b) / k,
lo=float(np.percentile(boot, 2.5)), hi=float(np.percentile(boot, 97.5)), n=k)
def m_avg(recs, qids):
by = defaultdict(list)
for q in qids:
by[recs[q]["meta"]["task_type"]].append(recs[q]["pred"] == recs[q]["answer"])
return 100 * float(np.mean([np.mean(v) for v in by.values()]))
def main():
root, base_prefix = sys.argv[1], sys.argv[2]
out = {}
for run in sorted(glob.glob(os.path.join(root, "*_b*"))):
m = re.match(r"(.+)_b(\d+)$", os.path.basename(run))
if not m:
continue
variant, budget = m.group(1), int(m.group(2))
for bench in ("videomme", "mlvu", "longvideobench"):
var = load(run, bench)
if not var:
continue
base = load(f"{base_prefix}_b{budget}", bench)
qids = sorted(q for q in var if q in base)
row = dict(n_total=len(base), complete=len(qids) == len(base),
all=paired(var, base, qids),
changed=paired(var, base, [q for q in qids if not var[q].get("reused")]),
long=paired(var, base, [q for q in qids if LONG[bench](base[q]["meta"])]))
if bench == "mlvu" and row["complete"]:
row["m_avg"], row["m_avg_base"] = m_avg(var, qids), m_avg(base, qids)
modes = defaultdict(int)
for q in qids:
if not var[q].get("reused"):
modes[var[q].get("mode", "error")] += 1
row["changed_modes"] = dict(modes)
out.setdefault(variant, {}).setdefault(str(budget), {})[bench] = row
def fmt(p):
if p is None:
return "—"
star = "*" if p["lo"] > 0 or p["hi"] < 0 else " "
return f"{p['acc']:5.1f} ({p['delta']:+5.1f}{star} n={p['n']})"
for variant in sorted(out):
for budget in sorted(out[variant], key=int):
for bench, r in out[variant][budget].items():
flag = "" if r["complete"] else " [INCOMPLETE]"
extra = f" M-Avg {r['m_avg']:.1f} ({r['m_avg'] - r['m_avg_base']:+.1f})" if "m_avg" in r else ""
print(f"{variant:11s} {budget:>5} {bench:15s} all {fmt(r['all'])} changed {fmt(r['changed'])} "
f"long {fmt(r['long'])}{extra}{flag}")
if len(sys.argv) > 3:
with open(sys.argv[3], "w") as f:
json.dump(out, f, indent=1)
if __name__ == "__main__":
main()
|