first commit

Codes/4_model_evaluation/eval_sort.py

@@ -0,0 +1,46 @@
+import json
+from collections import defaultdict
+
+prefix = "head_Llama-3.1-8B-Instruct_head_"
+score = "prc"
+target = "inst"
+topks = "0 3.125p 6.25p 9.375p 12.5p 15.625p 18.75p 21.875p 25p 28.125p 31.25p 50p 62.5p 75p 100p".split()
+
+attacks = ["none", "naive", "ignore", "escape_separation"]
+
+# [metric][attack] -> list aligned with topks
+attack_asr = defaultdict(list)
+attack_vr  = defaultdict(list)
+
+for t in topks:
+    with open(f"{prefix}{score}_{target}_{t}.json", "r") as f:
+        data = json.load(f)
+
+    # 每個檔案內的 summary 可能順序不固定，先做成 dict 方便取
+    by_attack = {s["attack"]: s for s in data["summary"]}
+
+    for a in attacks:
+        s = by_attack[a]
+        attack_asr[a].append(s["attack_success_rate"])
+        attack_vr[a].append(s["valid_rate"])
+
+def fmt_topk(t):
+    return t[:-1] if t.endswith("p") else t
+
+def fmt_num(x):
+    # 你範例是小數點後 9 位左右；這裡固定 9 位再去掉尾端 0
+    s = f"{x:.9f}".rstrip("0").rstrip(".")
+    return s
+
+def print_table(title, metric_dict):
+    print(title)
+    print("\t".join(["topk"] + attacks))
+    for i, t in enumerate(topks):
+        row = [fmt_topk(t)]
+        for a in attacks:
+            row.append(fmt_num(metric_dict[a][i]))
+        print("\t".join(row))
+
+print_table("ASR", attack_asr)
+print()
+print_table("Valid", attack_vr)