phase7: fix label handling - per-record labels are strings, map to DEC indices

phase7/postprocess_p7.py

@@ -54,12 +54,15 @@ def vecs(entry):
     n = entry["n"]
     ids, labels, rp = entry["ids"], entry["labels"], entry["raw_probs"]
     assert len(ids) == len(labels) == len(rp) == n, (len(ids), len(labels), len(rp), n)
-    return ids, np.array(labels), np.array(rp, dtype=np.float64)
+    # labels are stored as strings ("benign"/"suspicious"/"malicious") - map to indices
+    lab_idx = [DEC.index(str(l).lower()) for l in labels]
+    return ids, np.array(lab_idx), np.array(rp, dtype=np.float64)
 
 align_report = {}
 for s in sets_common:
     ids0 = pr_m[s]["sharky-0.5B"]["ids"]
-    lab0 = np.array(pr_m[s]["sharky-0.5B"]["labels"])
+    lab0 = vecs(pr_m[s]["sharky-0.5B"])[1]
+    vc = {d: int((lab0 == i).sum()) for i, d in enumerate(DEC)}
     for arm in arms_m:
         e = pr_m[s][arm]
         ids, lab, rp = vecs(e)
@@ -71,13 +74,15 @@ for s in sets_common:
         assert ids == ids0 and np.array_equal(lab, lab0), f"{s}/{arm}: cross-file misalignment vs minicpm"
         assert rp.shape[1] == 3 and np.abs(rp.sum(1) - 1.0).max() < 1e-2, f"{s}/{arm}: unnormalized"
     align_report[s] = {"n": len(ids0), "arms_checked": sorted(arms_m + arms_k),
-                       "ids_and_labels_aligned": True}
-    print(f"ALIGNED {s}: n={len(ids0)} across {sorted(arms_m + arms_k)}", flush=True)
+                       "label_counts": vc, "ids_and_labels_aligned": True}
+    print(f"ALIGNED {s}: n={len(ids0)} labels={vc}", flush=True)
+    assert vc["benign"] > 0 and vc["suspicious"] > 0 and vc["malicious"] > 0, \
+        f"{s}: degenerate label distribution {vc}"
 
 # ---- 2. kev probability sanity -----------------------------------------------
 kev_sanity = {}
 for s in sets_common:
-    lab = np.array(pr_m[s]["sharky-0.5B"]["labels"])
+    lab = vecs(pr_m[s]["sharky-0.5B"])[1]
     for arm in arms_k:
         rp = np.array(pr_k[s][arm]["raw_probs"], dtype=np.float64)
         am = rp.argmax(1)
@@ -113,8 +118,13 @@ def correct_vec(rp, lab, t=None):
 
 bootstrap = {}
 for s in sets_common:
-    lab = np.array(pr_m[s]["sharky-0.5B"]["labels"])
+    lab = vecs(pr_m[s]["sharky-0.5B"])[1]
     corr_sharky = correct_vec(np.array(pr_m[s]["sharky-0.5B"]["raw_probs"]), lab, T_CARD)
+    sharky_acc = float(corr_sharky.mean())
+    target = 0.9865 if s == "locked" else 0.9848
+    assert abs(sharky_acc - target) < 0.01, \
+        f"sanity: sharky acc {sharky_acc} on {s} deviates from aggregate {target}"
+    print(f"SHARKYACC {s}: {sharky_acc:.4f}", flush=True)
     for arm, src in [("minicpm4-0.5B-head-only-control", pr_m)] + [(a, pr_k) for a in arms_k]:
         rp = np.array(src[s][arm]["raw_probs"], dtype=np.float64)
         tfit = 1.0
@@ -131,7 +141,7 @@ gate_curves = {}
 for s in sets_common:
     rp = np.array(pr_m[s]["sharky-0.5B"]["raw_probs"], dtype=np.float64)
     p = rp ** (1.0 / T_CARD); p = p / p.sum(1, keepdims=True)
-    lab = np.array(pr_m[s]["sharky-0.5B"]["labels"])
+    lab = vecs(pr_m[s]["sharky-0.5B"])[1]
     pred = p.argmax(1)
     n = len(lab)
     rows = []