File size: 7,314 Bytes
8640a2e | 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 | #!/usr/bin/env python3
"""Compute per-model metrics + attributes + real statistics, dump one JSON for the report page."""
import json, math, os
EVAL = "/mnt/data/cvhci_video_understanding/eval"
ATTR = {
"Qwen2.5-VL-7B": dict(org="Alibaba", total=7, active=7, moe=False, thinking=False, year=2025),
"Qwen3-VL-8B": dict(org="Alibaba", total=8, active=8, moe=False, thinking=False, year=2026),
"Qwen3.5-9B": dict(org="Alibaba", total=9, active=9, moe=False, thinking=False, year=2026),
"Kimi-VL-A3B-Thinking-2506": dict(org="Moonshot", total=16, active=3, moe=True, thinking=True, year=2025),
"GLM-4.1V-9B-Thinking": dict(org="Z.ai", total=9, active=9, moe=False, thinking=True, year=2025),
"InternVL3.5-8B": dict(org="OpenGVLab",total=8, active=8, moe=False, thinking=False, year=2025),
"ERNIE-4.5-VL-28B-A3B": dict(org="Baidu", total=28, active=3, moe=True, thinking=False, year=2025),
"Qwen3-VL-30B-A3B-Thinking": dict(org="Alibaba", total=30, active=3, moe=True, thinking=True, year=2026),
"Pixtral-12B-2409": dict(org="Mistral", total=12, active=12, moe=False, thinking=False, year=2024),
"InternVL3.5-14B": dict(org="OpenGVLab",total=14, active=14, moe=False, thinking=False, year=2025),
"InternVL3.5-20B-A4B": dict(org="OpenGVLab",total=20, active=4, moe=True, thinking=False, year=2026),
"Qwen3.5-27B": dict(org="Alibaba", total=27, active=27, moe=False, thinking=False, year=2026),
"InternVL3.5-30B-A3B": dict(org="OpenGVLab",total=30, active=3, moe=True, thinking=False, year=2025),
"Qwen3.5-35B-A3B": dict(org="Alibaba", total=35, active=3, moe=True, thinking=False, year=2026),
"InternVL3.5-38B": dict(org="OpenGVLab",total=38, active=38, moe=False, thinking=False, year=2025),
"Qwen3.5-122B-A10B": dict(org="Alibaba", total=122, active=10, moe=True, thinking=False, year=2026),
"InternVL3.5-241B-A28B": dict(org="OpenGVLab",total=241, active=28, moe=True, thinking=False, year=2025),
"Qwen3-VL-235B-A22B-Instruct": dict(org="Alibaba", total=235, active=22, moe=True, thinking=False, year=2026),
"Qwen3-VL-235B-A22B-Thinking": dict(org="Alibaba", total=235, active=22, moe=True, thinking=True, year=2026),
"GLM-4.6V": dict(org="Z.ai", total=106, active=None,moe=True, thinking=False, year=2026),
"Qwen2.5-VL-72B-Instruct": dict(org="Alibaba", total=72, active=72, moe=False, thinking=False, year=2025),
"Gemma-3n-E4B-it": dict(org="Google", total=8, active=4, moe=False, thinking=False, year=2025),
"Gemma-4-31B-it": dict(org="Google", total=31, active=31, moe=False, thinking=False, year=2026),
"Gemma-4-26B-A4B-it": dict(org="Google", total=26, active=4, moe=True, thinking=False, year=2026),
}
MIN_N = 700 # 排除样本太少的(Molmo2-8B n=26, Pixtral 部分分片, gemini 部分跑)
def load(tag):
rows = []
for l in open(f"{EVAL}/results/{tag}.jsonl"):
try:
r = json.loads(l)
except Exception:
continue
if not r.get("error"):
rows.append(r)
return list({r["review_id"]: r for r in rows}.values())
def main():
models = []
for tag in sorted(ATTR):
path = f"{EVAL}/results/{tag}.jsonl"
if not os.path.exists(path):
continue
rows = load(tag)
n = len(rows)
if n < MIN_N:
continue
exact = sum(r["exact"] for r in rows)
overlap = sum(r["overlap"] for r in rows) / n
wellformed = sum(1 for r in rows if r["n_pred"] == r["n_select"]) / n
p = exact / n
se = (p * (1 - p) / n) ** 0.5
by_unit = {}
for r in rows:
by_unit.setdefault(r["unit"], []).append(r["exact"])
by_unit = {u: sum(v) / len(v) for u, v in by_unit.items()}
a = ATTR[tag]
models.append({
"model": tag, "org": a["org"], "total_b": a["total"], "active_b": a["active"],
"moe": a["moe"], "thinking": a["thinking"], "year": a["year"],
"n": n, "exact": round(100 * p, 2), "exact_lo": round(100 * max(0, p - 1.96 * se), 2),
"exact_hi": round(100 * min(1, p + 1.96 * se), 2),
"overlap": round(100 * overlap, 2), "wellformed": round(100 * wellformed, 2),
"by_unit": {u: round(100 * v, 1) for u, v in by_unit.items()},
})
# ---- 真实统计：log10(active_b) 与 exact 的 Pearson 相关，仅用 active_b 已知的模型 ----
pts = [(math.log10(m["active_b"]), m["exact"]) for m in models if m["active_b"]]
def pearson(pts):
n = len(pts); mx = sum(x for x, _ in pts) / n; my = sum(y for _, y in pts) / n
cov = sum((x - mx) * (y - my) for x, y in pts)
sx = (sum((x - mx) ** 2 for x, _ in pts)) ** 0.5
sy = (sum((y - my) ** 2 for _, y in pts)) ** 0.5
return cov / (sx * sy) if sx and sy else 0.0
r_active = pearson(pts)
pts_total = [(math.log10(m["total_b"]), m["exact"]) for m in models]
r_total = pearson(pts_total)
moe_pts = [m["exact"] for m in models if m["moe"]]
dense_pts = [m["exact"] for m in models if not m["moe"]]
think_pts = [m["exact"] for m in models if m["thinking"]]
nonthink_pts = [m["exact"] for m in models if not m["thinking"]]
org_avg = {}
for m in models:
org_avg.setdefault(m["org"], []).append(m["exact"])
org_avg = {o: round(sum(v) / len(v), 1) for o, v in sorted(org_avg.items(), key=lambda kv: -sum(kv[1]) / len(kv[1]))}
# 235B 同底座 Instruct vs Thinking 直接对照(唯一一对严格控制其它变量的样本)
pair = {m["model"]: m["exact"] for m in models if "235B-A22B" in m["model"]}
out = {
"models": sorted(models, key=lambda m: -m["exact"]),
"stats": {
"n_models": len(models),
"r_active_params": round(r_active, 3),
"r_total_params": round(r_total, 3),
"moe_mean": round(sum(moe_pts) / len(moe_pts), 1), "moe_n": len(moe_pts),
"dense_mean": round(sum(dense_pts) / len(dense_pts), 1), "dense_n": len(dense_pts),
"thinking_mean": round(sum(think_pts) / len(think_pts), 1), "thinking_n": len(think_pts),
"nonthinking_mean": round(sum(nonthink_pts) / len(nonthink_pts), 1), "nonthinking_n": len(nonthink_pts),
"org_avg": org_avg,
"instruct_vs_thinking_235B": pair,
},
}
json.dump(out, open(f"{EVAL}/analysis.json", "w"), ensure_ascii=False, indent=1)
print(f"models={len(models)} r(log active_params, exact)={r_active:.3f} r(log total_params, exact)={r_total:.3f}")
print(f"MoE mean={out['stats']['moe_mean']}% (n={out['stats']['moe_n']}) Dense mean={out['stats']['dense_mean']}% (n={out['stats']['dense_n']})")
print(f"Thinking mean={out['stats']['thinking_mean']}% (n={out['stats']['thinking_n']}) Non-thinking mean={out['stats']['nonthinking_mean']}% (n={out['stats']['nonthinking_n']})")
print("按厂商均分:", out['stats']['org_avg'])
print("235B Instruct vs Thinking:", pair)
if __name__ == "__main__":
main()
|