Spaces:
Running
Running
File size: 12,832 Bytes
76b5743 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 | #!/usr/bin/env python3
"""
大规模测试消融归因:测试几十个不同场景的 case,观察是否出现"全部为 0"的情况。
"""
import json
import sys
import os
import urllib.request
import urllib.error
# ---------------------------------------------------------------------------
# Configuration
# ---------------------------------------------------------------------------
API_BASE = "http://127.0.0.1:5001"
# ---------------------------------------------------------------------------
# Test cases: diverse scenarios
# ---------------------------------------------------------------------------
test_cases = [
# --- 基础事实 ---
("中国的首都是", "北京"),
("法国的首都是", "巴黎"),
("地球绕着什么转", "太阳"),
("1+1=", "2"),
("水的化学式是", "H2O"),
("太阳从什么方向升起", "东"),
# --- 单 token 目标 ---
("The capital of France is", " Paris"),
("The color of the sky is", " blue"),
("Apple's founder was", " Steve"),
("Python is a programming", " language"),
# --- 否定 / 对比 ---
("猫不属于", "鱼类"),
("企鹅虽然不会飞,但属于", "鸟类"),
("鲸鱼生活在水中,但它不是鱼,而是", "哺乳动物"),
# --- 逻辑推理 ---
("如果今天是周一,那么明天是", "周二"),
("一年有十二个月,一个月大约有30", "天"),
("人喝水是为了", "解渴"),
("汽车没有油就", "无法"),
# --- 中文长句子 ---
("人工智能是计算机科学的一个分支,它企图了解智能的实质,并生产出一种新的能以人类智能相似的方式做出反应的智能机器。该领域的研究包括", "机器人"),
("机器学习是人工智能的子领域,它专注于让计算机通过经验自动改善性能。监督学习是最常见的机器学习范式之一,它使用", "标注"),
("深度学习是机器学习的一个分支,它使用多层神经网络来学习数据的表示。卷积神经网络特别适合处理", "图像"),
# --- 英文长句子 ---
("Artificial intelligence is a branch of computer science that aims to create machines that can perform tasks that typically require human intelligence. One of the most popular applications is", " machine"),
("Natural language processing is a subfield of AI that focuses on the interaction between computers and humans through natural language. A key challenge is", " understanding"),
("The transformer architecture introduced in 'Attention is All You Need' has become the foundation of most modern language models. It relies primarily on", " attention"),
# --- 重复结构 / 模板 ---
("I think therefore I", " am"),
("To be or not to", " be"),
("Hello world is the first program every", " programmer"),
# --- 数字 / 数学 ---
("The square root of 9 is", " 3"),
("2 multiplied by 6 equals", " 12"),
("The year after 2023 is", " 2024"),
# --- 知识问答 ---
("爱因斯坦提出了", "相对论"),
("牛顿发现了", "万有引力"),
("莎士比亚是著名的", "戏剧"),
("莫扎特是一位", "作曲"),
# --- 短文本 / 单字目标 ---
("天对地,雨对风,大陆对长", "空"),
("床前明月", "光"),
# --- ambiguous / 开放 ---
("明天会不会下雨取决于", "天气"),
("要成功就必须付出", "努力"),
# --- top-1(不指定 target)的自动模式 ---
("我今天中午吃了", None), # top-1
("The best thing about summer is", None), # top-1
("在我看来,最重要的品质是", None), # top-1
]
def test_via_api(context, target_prediction):
"""通过 HTTP API 测试"""
payload = {
"context": context,
"model": "base",
"source_page": "attribution",
}
if target_prediction is not None:
payload["target_prediction"] = target_prediction
req = urllib.request.Request(
f"{API_BASE}/api/ablation-attribute",
data=json.dumps(payload).encode("utf-8"),
headers={"Content-Type": "application/json"},
method="POST",
)
try:
with urllib.request.urlopen(req, timeout=120) as resp:
data = json.loads(resp.read().decode("utf-8"))
return data
except urllib.error.HTTPError as e:
body = e.read().decode("utf-8")
return {"error": f"HTTP {e.code}: {body}", "success": False}
except Exception as e:
return {"error": str(e), "success": False}
def analyze_results(results):
"""分析测试结果,找出全为 0 的情况"""
zero_score_cases = []
non_zero_cases = [] # has at least one non-zero score
error_cases = []
for i, (case, r) in enumerate(results):
context, target = case
if "error" in r:
error_cases.append((i, context, target, r["error"]))
continue
if "token_attribution" not in r:
error_cases.append((i, context, target, "No token_attribution in response: " + json.dumps(r)[:200]))
continue
scores = [e["score"] for e in r["token_attribution"]]
non_zero_scores = [s for s in scores if abs(s) > 1e-10]
info = {
"case_idx": i,
"context": repr((context[:50] + "...") if len(context) > 50 else context),
"target": target,
"target_token": r.get("target_token", "?"),
"target_prob": r.get("target_prob"),
"n_tokens": len(r["token_attribution"]),
"n_non_zero": len(non_zero_scores),
"max_score": max(scores) if scores else 0,
"min_score": min(scores) if scores else 0,
"all_scores": scores,
"all_deltas": [e.get("delta_logit", "N/A") for e in r["token_attribution"]],
}
if len(non_zero_scores) == 0:
zero_score_cases.append(info)
else:
non_zero_cases.append(info)
return {
"zero_score_cases": zero_score_cases,
"non_zero_cases": non_zero_cases,
"error_cases": error_cases,
"total": len(results),
"zero_count": len(zero_score_cases),
"non_zero_count": len(non_zero_cases),
"error_count": len(error_cases),
}
def print_analysis(stats):
print("\n" + "=" * 80)
print("📊 消融归因大规模测试报告")
print("=" * 80)
print(f"总测试用例: {stats['total']}")
print(f" ✅ 有非零分: {stats['non_zero_count']}")
print(f" ❌ 全为 0: {stats['zero_count']}")
print(f" ⚠️ 错误: {stats['error_count']}")
print()
if stats['error_cases']:
print("--- 错误用例 ---")
for idx, ctx, tgt, err in stats['error_cases']:
print(f" [{idx}] context={repr(str(ctx)[:40])}, target={tgt}")
print(f" error: {err}")
print()
if stats['zero_score_cases']:
print("--- 🚨 全为 0 的用例 ---")
for info in stats['zero_score_cases']:
print(f" [{info['case_idx']}] context={info['context']}")
print(f" target={info['target']}, target_token={info['target_token']!r}")
print(f" target_prob={info['target_prob']}, n_tokens={info['n_tokens']}")
print(f" scores: {info['all_scores']}")
# Check delta_logits
deltas = [d for d in info['all_deltas'] if d != "N/A"]
if deltas:
non_zero_deltas = [d for d in deltas if abs(d) > 1e-10]
print(f" delta_logits: {info['all_deltas'][:10]}...")
if non_zero_deltas:
print(f" ⚠️ scores 全 0,但 delta_logits 有非零值! count={len(non_zero_deltas)}")
else:
print(f" delta_logits 也全 0")
print()
if stats['non_zero_cases']:
print("--- ✅ 有非零分的用例(前 8 个)---")
for info in stats['non_zero_cases'][:8]:
non_zero_ratio = f"{info['n_non_zero']}/{info['n_tokens']}"
print(f" [{info['case_idx']}] context={info['context']}")
print(f" target={info['target']}, non_zero_ratio={non_zero_ratio}")
print(f" max_score={info['max_score']:.6f}, min_score={info['min_score']:.4f}, target_prob={info['target_prob']}")
print(f" scores={info['all_scores']}")
print()
def main():
print(f"🚀 启动消融归因大规模测试 ({len(test_cases)} 个用例)")
print(f"目标服务器: {API_BASE}")
print()
# 先检查服务器是否可达
try:
with urllib.request.urlopen(f"{API_BASE}/", timeout=5) as _:
print("✅ 服务器可达")
except Exception as e:
print(f"⚠️ 服务器不可达 ({e}),继续尝试...")
print()
results = []
for i, (context, target) in enumerate(test_cases):
short_ctx = (context[:40] + "...") if len(context) > 40 else context
print(f"[{i+1}/{len(test_cases)}] testing: context={repr(short_ctx)}, target={target}")
try:
r = test_via_api(context, target)
results.append(((context, target), r))
# 检查是否有 score
if "token_attribution" in r:
scores = [e["score"] for e in r["token_attribution"]]
non_zero = sum(1 for s in scores if abs(s) > 1e-10)
total = len(scores)
if non_zero == 0:
print(f" ❌ 全为 0! target_prob={r.get('target_prob')}, target_token={r.get('target_token')!r}")
else:
max_s = max(scores)
print(f" ✅ {non_zero}/{total} 非零, max_score={max_s:.6f}")
elif "error" in r:
print(f" ⚠️ error: {r['error'][:80]}")
except Exception as e:
print(f" ❌ 异常: {e}")
results.append(((context, target), {"error": str(e)}))
stats = analyze_results(results)
print_analysis(stats)
# 深入分析全零根因
if stats['zero_count'] > 0:
first_zero = stats['zero_score_cases'][0]
ctx = test_cases[first_zero['case_idx']][0]
tgt = test_cases[first_zero['case_idx']][1]
print("=" * 80)
print("🔍 对第一个全零 case 做深度诊断")
print("=" * 80)
diagnose_ablation(ctx, tgt)
def diagnose_ablation(context, target_prediction=None):
"""对单个 case 做详细诊断"""
payload = {
"context": context,
"model": "base",
"source_page": "attribution",
}
if target_prediction is not None:
payload["target_prediction"] = target_prediction
req = urllib.request.Request(
f"{API_BASE}/api/ablation-attribute",
data=json.dumps(payload).encode("utf-8"),
headers={"Content-Type": "application/json"},
method="POST",
)
try:
with urllib.request.urlopen(req, timeout=120) as resp:
data = json.loads(resp.read().decode("utf-8"))
except Exception as e:
print(f" API 调用失败: {e}")
return
if "token_attribution" in data:
print(f" token_attribution 条目数: {len(data['token_attribution'])}")
print(f" target_prob: {data['target_prob']}")
print(f" target_token: {data['target_token']!r}")
print(f" debug_info: {data.get('debug_info')}")
scores = [e["score"] for e in data["token_attribution"]]
deltas = [e.get("delta_logit", "N/A") for e in data["token_attribution"]]
raws = [e["raw"] for e in data["token_attribution"]]
print(f" scores: {scores}")
print(f" delta_logits: {deltas}")
print(f" raws: {raws}")
# 如果 scores 全 0 但 delta_logit 非全 0,说明 prob 变化在 round_to_sig_figs 之下
if all(abs(s) < 1e-10 for s in scores):
numeric_deltas = [d for d in deltas if isinstance(d, (int, float))]
non_zero_deltas = [d for d in numeric_deltas if abs(d) > 1e-10]
if non_zero_deltas:
print(f"\n 🔑 关键发现: scores 全部 round 到 0,但 delta_logits 有非零值!")
print(f" 说明模型对遮挡确实有反应,但概率变化太小 (< 5e-8)")
print(f" 非零 deltas 示例: {non_zero_deltas[:10]}")
else:
print(f"\n 🔑 delta_logits 也是全 0 (或 {len(numeric_deltas)}/{len(deltas)} 为数值型),说明遮挡完全不影响预测")
print(f"\n 📝 debug_info topk_tokens: {data.get('debug_info', {}).get('topk_tokens', [])}")
print(f" 📝 debug_info topk_probs: {data.get('debug_info', {}).get('topk_probs', [])}")
if __name__ == "__main__":
main() |