Upload code/set4_goldfish.py with huggingface_hub

code/set4_goldfish.py

@@ -137,6 +137,9 @@ for spec in A.pairs.split(","):
     BODY = [k for k in SD_E if not (k.endswith("wte.weight") or k.endswith("lm_head.weight"))]
 
     # ------------- alignments (fitted BEFORE merging)
+    R_emb, n_anch = G2.emb_procrustes(SD_E, SD_X, tok_e, tok_x)
+    sd_emb = G2.apply_resid(SD_X_V, D, R=R_emb)
+    sd_emb2, _ = G2.align_full(SD_E, sd_emb, D, NH, None, None, "permutation", body_keys=BODY)
     sdp, ip = G2.align_full(SD_E, SD_X_V, D, NH, acts_e, acts_x, "permutation", body_keys=BODY)
     sdo, io = G2.align_full(SD_E, SD_X_V, D, NH, acts_e, acts_x, "orthogonal", body_keys=BODY)
     sdpf, ipf = G2.align_full(SD_E, SD_X_V, D, NH, acts_e, acts_x, "permutation", body_keys=BODY, accept_each=False)
@@ -146,7 +149,7 @@ for spec in A.pairs.split(","):
     # ------------- predictors (pre-merge)
     KEYS = shared_keys(SD_E, SD_X)
     fa, fb = flat(SD_E, KEYS), flat(SD_X, KEYS)
-    p = {"weight_cosine": float(fa @ fb / (np.linalg.norm(fa) * np.linalg.norm(fb))),
+    p = {"n_emb_anchors": n_anch, "weight_cosine": float(fa @ fb / (np.linalg.norm(fa) * np.linalg.norm(fb))),
          "vocab_overlap": len(anchors) / V}
     p["weight_cosine_body"] = float(np.mean([
         float(np.asarray(SD_E[k], float).ravel() @ np.asarray(SD_X[k], float).ravel() /
@@ -177,6 +180,8 @@ for spec in A.pairs.split(","):
              "M1c_vocab_orth_avg": MG.average([SD_E, sdo]),
              "M1d_vocab_perm_forced": MG.average([SD_E, sdpf]),
              "M1e_vocab_orth_forced": MG.average([SD_E, sdof]),
+             "M1g_emb_procrustes": MG.average([SD_E, sd_emb]),
+             "M1h_emb_proc_units": MG.average([SD_E, sd_emb2]),
              "M1f_perm_novocab": MG.average([SD_E, G2.align_full(SD_E, SD_X, D, NH, acts_e, acts_x, "permutation", body_keys=BODY)[0]])}
     res = {}
     for name, sd in rungs.items():
@@ -218,6 +223,6 @@ for spec in A.pairs.split(","):
         f"M1a={res['M1a_vocab_avg']['delta_floor_mean']:+.4f}  "
         f"M1b_perm={res['M1b_vocab_perm_avg']['delta_floor_mean']:+.4f}  "
         f"M1c_orth={res['M1c_vocab_orth_avg']['delta_floor_mean']:+.4f}  ({r['secs']:.0f}s)")
-    del rungs, sdp, sdo, sdpf, sdof, SD_X, SD_X_V; gc.collect()
+    del rungs, sdp, sdo, sdpf, sdof, sd_emb, sd_emb2, SD_X, SD_X_V; gc.collect()
 fh.close()
 log("DONE set4")