File size: 70,055 Bytes
35acee3 b553ba1 35acee3 b553ba1 35acee3 b553ba1 35acee3 b553ba1 35acee3 b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d 35acee3 b553ba1 35acee3 b553ba1 35acee3 b553ba1 35acee3 b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d b553ba1 cbcdc9d 35acee3 cbcdc9d b553ba1 cbcdc9d 35acee3 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 | # =============================================================================
# COMPRESSION NAVIGATOR · extended + annotated edition
# =============================================================================
# An LLM is a lossy codec for text. Training compresses a corpus into weights;
# a forward pass decompresses a continuation. These five tools let you watch
# that decompression happen and poke at where facts physically live.
#
# The five tabs are not toys invented here - each one is a real mechanistic-
# interpretability technique you'll find in papers:
#
# 1. Decompress = LOGIT LENS (nostalgebraist, 2020)
# 2. Triangulate = EMBEDDING NEIGHBOURS (the geometry of the vocab)
# 3. Re-route = ACTIVATION STEERING (ActAdd / repr. engineering)
# 4. Diff = CROSS-MODEL ALIGNMENT (compare checkpoints by depth)
# 5. Causal trace = ACTIVATION PATCHING (ROME, Meng et al., 2022)
#
# WHY THE GLASS-BOX MODELS MATTER
# -------------------------------
# On a real model (gpt2) you never know the ground truth, so you can't tell
# whether a tool is *correct* or just producing plausible-looking output.
# This file ships two models whose internals you fully specify, so you can
# check each tool against a known answer:
#
# "handmade" - facts stored as a LOOKUP TABLE keyed on the prompt string.
# The computation happens in a side channel (string match),
# NOT in the residual stream. Lesson: such a model is almost
# invisible to residual-stream interpretability. Logit lens
# sees a sudden jump with no build-up; causal tracing finds
# nothing, because corrupting activations doesn't touch the
# string match. This is a real and underappreciated *limit*
# of these methods.
#
# "glassbox" - facts stored the way real transformers store them: as
# key->value writes into the RESIDUAL STREAM (Geva et al.'s
# "MLPs are key-value memories", which is exactly what ROME
# edits). Because the fact flows through activations, ALL five
# tools light up correctly - and you can verify they report
# the layer you actually put the fact in. This is a unit-test
# harness for interpretability code.
#
# Run order suggestion: glassbox -> handmade -> gpt2
# glassbox shows what "correct" looks like; handmade shows a failure mode;
# gpt2 shows the fuzzy, distributed real thing.
# =============================================================================
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DTYPE = torch.float32
MODELS = {} # name -> (model, tokenizer) cache
STATE = {"name": None} # currently loaded model name
# =============================================================================
# A tiny shared tokenizer for both glass-box models.
# Case is CANONICALISED to lowercase everywhere (this fixes a real bug in the
# original: "Paris" from a pinned fact and "paris" from the Markov table became
# two different vocab entries, so the boosted token and the *tracked* token
# silently diverged - every neighbour read cos=0.000 and every tracked prob 0).
# =============================================================================
class FakeBatchEncoding(dict):
def to(self, device): # let callers do tok(...).to(DEVICE) safely
return self
class SimpleTok:
"""Whitespace tokenizer over a fixed vocab. Not 'fast' (no offset map)."""
is_fast = False
def __init__(self, stoi, itos):
self.stoi, self.itos = stoi, itos
self.eos_token_id = stoi["."] # period doubles as end-of-sequence
def _ids(self, text):
words = text.lower().replace(".", " .").split()
return [self.stoi.get(w, self.stoi["<s>"]) for w in words]
def __call__(self, text, return_tensors=None, return_offsets_mapping=False):
ids = self._ids(text) or [self.stoi["<s>"]]
return FakeBatchEncoding(
input_ids=torch.tensor([ids]),
attention_mask=torch.ones(1, len(ids), dtype=torch.long),
)
def encode(self, text, add_special_tokens=False):
return self._ids(text)
def decode(self, ids, skip_special_tokens=False):
out = []
for i in ids:
w = self.itos.get(int(i), "?")
if skip_special_tokens and w in ("<pad>", "<s>"):
continue
out.append(w)
return " ".join(out)
class _Out:
"""Mimics a HF CausalLMOutput: .logits and (optional) .hidden_states."""
def __init__(self, logits, hidden_states):
self.logits = logits
self.hidden_states = hidden_states
def _greedy_generate(model, input_ids, max_new_tokens=20, pad_token_id=None, **_):
"""Minimal greedy decode so the steering tab works on the toy models too
(the originals had no .generate, so that tab crashed on 'handmade')."""
ids = input_ids
for _ in range(int(max_new_tokens)):
nxt = model(input_ids=ids).logits[0, -1].argmax().view(1, 1)
ids = torch.cat([ids, nxt], dim=1)
if pad_token_id is not None and int(nxt.item()) == int(pad_token_id):
break
return ids
# =============================================================================
# MODEL 1 - "handmade": facts as a LOOKUP TABLE (the side-channel glass box)
# -----------------------------------------------------------------------------
# Embeddings are the identity matrix (each token is its own one-hot). The two
# "layers" don't read the residual stream in a meaningful linear way:
# - MemoryBlock matches the *decoded prompt string* and boosts the answer.
# - MarkovBlock adds a hand-built bigram transition for the last token.
# Because MemoryBlock keys on the prompt TEXT, not on activations, this is a
# deliberate demonstration of a model that residual-stream interpretability
# cannot see. Use it as the "what failure looks like" control.
# =============================================================================
PINNED = { # answers are lowercase now (bug fix)
"the capital of france is": " paris",
"the eiffel tower is in": " paris",
"two plus two equals": " four",
}
MARKOV = {
"<s>": {"the": 3, "i": 2, "a": 1},
"the": {"city": 2, "tower": 2, "answer": 1},
"i": {"think": 2, "am": 1},
"a": {"model": 2, "city": 1},
"city": {"of": 3, "is": 1},
"of": {"light": 2, "paris": 1},
"tower": {"is": 3},
"is": {"in": 2, "a": 1},
"in": {"paris": 2, "france": 1},
"model": {"is": 2},
"think": {"the": 2},
"paris": {".": 1},
"france": {".": 1},
"light": {".": 1},
"four": {".": 1},
}
def _build_handmade_vocab():
toks, seen = ["<pad>", "<s>", "."], {"<pad>", "<s>", "."}
def add(w):
if w not in seen:
toks.append(w); seen.add(w)
for v in PINNED.values():
add(v.strip())
for w, nxts in MARKOV.items():
add(w)
for x in nxts:
add(x)
for k in PINNED:
for w in k.split():
add(w)
return toks
HM_VOCAB = _build_handmade_vocab()
HM_STOI = {w: i for i, w in enumerate(HM_VOCAB)}
HM_ITOS = {i: w for w, i in HM_STOI.items()}
HM_V = len(HM_VOCAB)
class _MemoryBlock(nn.Module):
"""If the decoded prompt ends with a pinned key, slam the answer logit.
NOTE: this reads prompt_ids (the string), not x - that's the whole point."""
def forward(self, x, prompt_ids=None):
out = x.clone()
if prompt_ids is not None:
text = " ".join(HM_ITOS.get(int(i), "") for i in prompt_ids).strip()
for key, ans in PINNED.items():
if text.endswith(key):
out[0, -1, HM_STOI[ans.strip()]] += 12.0
return (out,)
class _MarkovBlock(nn.Module):
"""Add a hand-built bigram transition row for the last token."""
def __init__(self):
super().__init__()
T = torch.zeros(HM_V, HM_V)
for w, nxts in MARKOV.items():
if w in HM_STOI:
tot = sum(nxts.values())
for x, wt in nxts.items():
if x in HM_STOI:
T[HM_STOI[w], HM_STOI[x]] = wt / tot
self.register_buffer("T", T)
def forward(self, x, prompt_ids=None):
out = x.clone()
if prompt_ids:
out[0, -1] += 4.0 * self.T[int(prompt_ids[-1])]
return (out,)
class _HMTransformer(nn.Module):
def __init__(self):
super().__init__()
self.wte = nn.Embedding(HM_V, HM_V)
with torch.no_grad():
self.wte.weight.copy_(torch.eye(HM_V)) # one-hot embeddings
self.h = nn.ModuleList([_MemoryBlock(), _MarkovBlock()])
self.ln_f = nn.Identity()
class HandmadeModel(nn.Module):
def __init__(self):
super().__init__()
self.transformer = _HMTransformer()
self.head = nn.Linear(HM_V, HM_V, bias=False)
with torch.no_grad():
self.head.weight.copy_(torch.eye(HM_V)) # identity unembed
self.tok = SimpleTok(HM_STOI, HM_ITOS)
def get_input_embeddings(self): return self.transformer.wte
def get_output_embeddings(self): return self.head
def generate(self, input_ids=None, attention_mask=None, **kw):
return _greedy_generate(self, input_ids, **kw)
def forward(self, input_ids=None, attention_mask=None, output_hidden_states=False):
ids = input_ids[0].tolist()
x = self.transformer.wte(input_ids).float()
hs = [x]; h = x
for blk in self.transformer.h:
(h,) = blk(h, prompt_ids=ids); hs.append(h)
logits = self.head(self.transformer.ln_f(h))
return _Out(logits, tuple(hs) if output_hidden_states else None)
# =============================================================================
# MODEL 2 - "glassbox": facts as RESIDUAL-STREAM key->value writes
# -----------------------------------------------------------------------------
# This is the model the original was missing. It stores facts the way real
# transformers do, so every tool works AND can be checked against ground truth.
#
# Vocab + structured embeddings (d=32). Country and its capital deliberately
# SHARE an embedding dimension, so the neighbours tool finds real geometry
# (paris is near france).
#
# Four layers:
# L0 subject site : (identity here) the residual the trace will restore
# L1 pool/attention : copies subject signal from earlier positions -> last
# L2 fact MLP : key(subject+relation) -> relu -> value(answer dir) <- ROME edits this kind of layer
# L3 cleanup : identity
#
# Ground truth you can verify:
# - logit lens: the answer is INVISIBLE until L2, then appears. Compare with
# handmade (sudden, no build-up) and gpt2 (fuzzy, spread over many layers).
# - causal trace: corrupting the subject and restoring layer by layer peaks
# at L0 - because L1's "attention" re-reads the restored subject. That is
# the ROME story: the causal site is an early layer at the SUBJECT token.
# - steering / neighbours: both operate on real directions, so both work.
# =============================================================================
GB_D = 32
GB_TOKS = ["<pad>", "<s>", ".", "the", "capital", "of", "is", "in",
"france", "germany", "japan", "paris", "berlin", "tokyo",
"london", "rome"] # spare answers so edits can hit a fresh target
GB_STOI = {w: i for i, w in enumerate(GB_TOKS)}
GB_ITOS = {i: w for w, i in GB_STOI.items()}
GB_V = len(GB_TOKS)
GB_FACTS = [("france", "paris"), ("germany", "berlin"), ("japan", "tokyo")]
def _build_gb_embeddings():
E = torch.zeros(GB_V, GB_D)
def setd(tok, pairs):
for d, v in pairs:
E[GB_STOI[tok], d] = v
# country/capital pairs share their first dim -> positive cosine (geometry!)
setd("france", [(0, 1.0), (1, 0.6), (20, 0.5)])
setd("paris", [(0, 0.8), (2, 0.9), (21, 0.5)])
setd("germany",[(3, 1.0), (4, 0.6), (22, 0.5)])
setd("berlin", [(3, 0.8), (5, 0.9), (23, 0.5)])
setd("japan", [(6, 1.0), (7, 0.6), (24, 0.5)])
setd("tokyo", [(6, 0.8), (8, 0.9), (25, 0.5)])
setd("london", [(27, 1.0), (28, 0.5)]) # spare answers (own dirs)
setd("rome", [(29, 1.0), (30, 0.5)])
setd("is", [(9, 1.0), (26, 0.4)]) # the relation marker
for i, t in enumerate(GB_TOKS): # give fillers an id
if E[i].abs().sum() == 0:
E[i, 10 + i % 6] = 1.0
return E / (E.norm(dim=-1, keepdim=True) + 1e-9) # unit rows
GB_E = _build_gb_embeddings()
GB_SUBJ = torch.zeros(GB_D, GB_D) # projector onto subject dims 0..8
for _d in range(9):
GB_SUBJ[_d, _d] = 1.0
class _GBIdent(nn.Module):
def forward(self, x, prompt_ids=None):
return (x.clone(),)
class _GBPool(nn.Module):
"""Toy 'attention': sum the subject-projected residual of all earlier
positions into the last position. Corrupting the subject earlier shows up
here; restoring the subject BEFORE this layer is what makes the trace
recover - that is why the causal peak lands at L0, not L1."""
def forward(self, x, prompt_ids=None):
out = x.clone()
if x.shape[1] > 1:
pooled = (x[0, :-1] @ GB_SUBJ.T).sum(0)
out[0, -1] = out[0, -1] + 0.9 * pooled
return (out,)
class _GBFactMLP(nn.Module):
"""Geva-style key->value memory. W_in rows are (subject+relation) keys;
relu gates which fact fires; W_out columns are answer unembed directions.
This is structurally the exact layer ROME rewrites to edit a fact."""
def __init__(self):
super().__init__()
Win = torch.zeros(len(GB_FACTS), GB_D)
Wout = torch.zeros(GB_D, len(GB_FACTS))
rel = GB_E[GB_STOI["is"]]
for k, (s, a) in enumerate(GB_FACTS):
key = (GB_E[GB_STOI[s]] @ GB_SUBJ.T) * 0.9 + rel
Win[k] = key / key.norm()
Wout[:, k] = GB_E[GB_STOI[a]] # write answer direction
self.register_buffer("Win", Win)
self.register_buffer("Wout", Wout)
self.register_buffer("Win0", Win.clone()) # pristine backups for reset
self.register_buffer("Wout0", Wout.clone())
self.bias, self.gain = 0.85, 6.0 # tuned: clean p~0.5, corrupt p~0.07
def forward(self, x, prompt_ids=None):
out = x.clone()
pre = F.relu(self.Win @ out[0, -1] - self.bias)
out[0, -1] = out[0, -1] + self.gain * (self.Wout @ pre)
return (out,)
class _GBTransformer(nn.Module):
def __init__(self):
super().__init__()
self.wte = nn.Embedding(GB_V, GB_D)
with torch.no_grad():
self.wte.weight.copy_(GB_E)
self.h = nn.ModuleList([_GBIdent(), _GBPool(), _GBFactMLP(), _GBIdent()])
self.ln_f = nn.Identity()
class GlassBoxModel(nn.Module):
def __init__(self):
super().__init__()
self.transformer = _GBTransformer()
self.head = nn.Linear(GB_D, GB_V, bias=False)
with torch.no_grad():
self.head.weight.copy_(GB_E) # tied unembed
self.tok = SimpleTok(GB_STOI, GB_ITOS)
# --- knowledge editing (ROME-style, exact on this key->value layer) -------
@torch.no_grad()
def edit_fact(self, subject, new_answer, method="rank1", strength=1.0):
"""Rewrite the value a fact-MLP key maps to. Methods:
rank1 / surgical - the minimal update: change only this fact's value.
broadcast - DELIBERATELY sloppy: smear the delta across ALL
facts, so the verifier has real collateral to catch.
"""
fm = self.transformer.h[2] # the FactMLP block
subjects = [s for s, _ in GB_FACTS]
if subject not in subjects:
raise ValueError("unknown subject %r" % subject)
if new_answer not in GB_STOI:
raise ValueError("unknown answer token %r" % new_answer)
k = subjects.index(subject)
delta = (GB_E[GB_STOI[new_answer]] - fm.Wout0[:, k]) * float(strength)
if method in ("rank1", "surgical"):
fm.Wout[:, k] = fm.Wout0[:, k] + delta
elif method == "broadcast":
fm.Wout += delta.unsqueeze(1) # hits every fact
else:
raise ValueError("unknown method %r" % method)
@torch.no_grad()
def reset(self):
fm = self.transformer.h[2]
fm.Win.copy_(fm.Win0); fm.Wout.copy_(fm.Wout0)
def get_input_embeddings(self): return self.transformer.wte
def get_output_embeddings(self): return self.head
def generate(self, input_ids=None, attention_mask=None, **kw):
return _greedy_generate(self, input_ids, **kw)
def forward(self, input_ids=None, attention_mask=None, output_hidden_states=False):
ids = input_ids[0].tolist()
x = self.transformer.wte(input_ids).float()
hs = [x]; h = x
for blk in self.transformer.h:
(h,) = blk(h, prompt_ids=ids); hs.append(h)
logits = self.head(self.transformer.ln_f(h))
return _Out(logits, tuple(hs) if output_hidden_states else None)
# =============================================================================
# REAL MODELS - resolve the architecture-specific module paths
# =============================================================================
def _resolve(model, paths):
for path in paths:
obj, ok = model, True
for part in path.split("."):
if hasattr(obj, part):
obj = getattr(obj, part)
else:
ok = False; break
if ok:
return obj
return None
def get_blocks(model):
blocks = _resolve(model, ["transformer.h", "model.layers",
"gpt_neox.layers", "model.decoder.layers"])
if blocks is None:
raise RuntimeError("Could not locate transformer blocks.")
return blocks
def get_final_norm(model):
norm = _resolve(model, ["transformer.ln_f", "model.norm",
"gpt_neox.final_layer_norm",
"model.decoder.final_layer_norm"])
return norm if norm is not None else (lambda x: x)
def get_head(model):
return model.get_output_embeddings()
def get_handles(name):
if name not in MODELS:
if name == "handmade":
m = HandmadeModel().eval(); MODELS[name] = (m, m.tok)
elif name == "glassbox":
m = GlassBoxModel().eval(); MODELS[name] = (m, m.tok)
else:
tok = AutoTokenizer.from_pretrained(name)
model = AutoModelForCausalLM.from_pretrained(
name, torch_dtype=DTYPE).to(DEVICE).eval()
MODELS[name] = (model, tok)
return MODELS[name]
def load_model(name):
name = name.strip()
model, _ = get_handles(name)
STATE["name"] = name
return "Loaded **%s** (%d layers)." % (name, len(get_blocks(model)))
# =============================================================================
# Shared readout: project every layer's last-token residual to a vocab dist.
# =============================================================================
@torch.no_grad()
def layer_distributions(model, tok, prompt):
inputs = tok(prompt, return_tensors="pt").to(DEVICE)
out = model(**inputs, output_hidden_states=True)
hs = out.hidden_states
norm, head, n = get_final_norm(model), get_head(model), len(out.hidden_states)
dists = []
for i, layer_hs in enumerate(hs):
vec = layer_hs[0, -1].to(DTYPE)
# HF convention: the LAST hidden_states entry is already post-ln_f,
# so skip norm there; apply ln_f to intermediates (logit-lens style).
logits = head(vec) if i == n - 1 else head(norm(vec))
dists.append(("embed" if i == 0 else "L%d" % i, F.softmax(logits, dim=-1)))
return dists
def _entropy_bits(probs):
p = probs.clamp_min(1e-12)
return float(-(p * p.log()).sum() / math.log(2))
# =============================================================================
# TAB 1 - LOGIT LENS: watch the answer condense out of the residual stream
# =============================================================================
@torch.no_grad()
def logit_lens(prompt, top_k, track):
if STATE["name"] is None:
return "Load a model first."
model, tok = get_handles(STATE["name"])
top_k = int(top_k)
tids = tok.encode(track, add_special_tokens=False) if track.strip() else []
tid = tids[0] if tids else None
dists = layer_distributions(model, tok, prompt)
header = "layer | top tokens (prob) | entropy" \
+ (" | p(%r)" % track if tid is not None else "")
lines = ["prompt: %r" % prompt, header, "-" * len(header)]
for label, probs in dists:
p, idx = probs.topk(top_k)
shown = " ".join("%r:%.2f" % (tok.decode([t]).replace("\n", "\\n"), v)
for t, v in zip(idx.tolist(), p.tolist()))
row = "%5s | %-40s | %4.1fb" % (label, shown, _entropy_bits(probs))
if tid is not None:
row += " | %.3f" % probs[tid].item()
lines.append(row)
return "\n".join(lines)
# =============================================================================
# TAB 2 - NEIGHBOURS: the geometry of the (un)embedding space
# =============================================================================
@torch.no_grad()
def neighbors(word, top_k):
if STATE["name"] is None:
return "Load a model first."
model, tok = get_handles(STATE["name"])
top_k = int(top_k)
ids = tok.encode(word, add_special_tokens=False)
if not ids:
return "Could not tokenize %r." % word
tid = ids[0]
W = F.normalize(get_head(model).weight.to(DTYPE), dim=-1)
sims = W @ W[tid]
vals, idx = sims.topk(top_k + 1)
note = ""
if STATE["name"] == "handmade":
note = ("(handmade uses one-hot embeddings, so every token is "
"orthogonal -> all cosines are 0 by construction. This is the "
"tool telling the truth about a model with no vocab geometry.)\n")
lines = [note + "neighbours of %r:" % word]
for v, j in zip(vals.tolist(), idx.tolist()):
if j != tid:
lines.append(" %14r cos=%.3f" % (tok.decode([j]), v))
return "\n".join(lines[: top_k + 1])
# =============================================================================
# TAB 3 - STEERING: bend behaviour by adding a direction, no retraining
# =============================================================================
def _make_steer_hook(direction, alpha):
d = direction * alpha
def hook(module, inp, out):
if isinstance(out, tuple):
return (out[0] + d.to(out[0].dtype).to(out[0].device),) + out[1:]
return out + d.to(out.dtype).to(out.device)
return hook
@torch.no_grad()
def steer_generate(prompt, source, target, layer, alpha, max_new):
if STATE["name"] is None:
return "Load a model first.", ""
model, tok = get_handles(STATE["name"])
layer, max_new = int(layer), int(max_new)
emb = model.get_input_embeddings().weight
def first_emb(w):
ids = tok.encode(w, add_special_tokens=False)
return emb[ids[0]] if ids else torch.zeros(emb.shape[-1], device=DEVICE)
direction = F.normalize((first_emb(target) - first_emb(source)).to(DTYPE), dim=-1)
inputs = tok(prompt, return_tensors="pt").to(DEVICE)
gk = dict(max_new_tokens=max_new, do_sample=False, pad_token_id=tok.eos_token_id)
base = tok.decode(model.generate(**inputs, **gk)[0], skip_special_tokens=True)
blocks = get_blocks(model)
layer = max(0, min(layer, len(blocks) - 1))
handle = blocks[layer].register_forward_hook(_make_steer_hook(direction, alpha))
try:
steered = tok.decode(model.generate(**inputs, **gk)[0], skip_special_tokens=True)
finally:
handle.remove()
return base, "steer %r -> %r @ L%d alpha=%s\n%s" % (source, target, layer, alpha, steered)
# =============================================================================
# TAB 4 - DIFF: compare two models on one prompt, aligned by relative depth
# =============================================================================
@torch.no_grad()
def diff_models(name_a, name_b, prompt, target, top_k):
ma, ta = get_handles(name_a.strip())
mb, tb = get_handles(name_b.strip())
ida = ta.encode(target, add_special_tokens=False)
idb = tb.encode(target, add_special_tokens=False)
if not ida or not idb:
return "Could not tokenize target %r in both models." % target
ida, idb = ida[0], idb[0]
da = layer_distributions(ma, ta, prompt)
db = layer_distributions(mb, tb, prompt)
nA, nB = len(da) - 1, len(db) - 1
def top1(probs, tok):
v, i = probs.topk(1)
return "%r:%.2f" % (tok.decode([i.item()]), v.item())
lines = ["prompt: %r target: %r" % (prompt, target),
"%18s | %16s %6s | %16s %6s | %7s"
% ("depth (A/B)", "A top1", "pA", "B top1", "pB", "dp")]
for i in range(nA + 1):
frac = (i / nA) if nA > 0 else 0.0
j = max(0, min(round(frac * nB), nB)) if nB > 0 else 0
la, pa = da[i]; lb, pb = db[j]
a_t, b_t = pa[ida].item(), pb[idb].item()
lines.append("%18s | %16s %6.3f | %16s %6.3f | %+7.3f"
% ("%3.0f%% (%s/%s)" % (frac * 100, la, lb),
top1(pa, ta), a_t, top1(pb, tb), b_t, b_t - a_t))
return "\n".join(lines)
# =============================================================================
# TAB 5 - CAUSAL TRACE: corrupt the subject, restore each layer, find the site
# -----------------------------------------------------------------------------
# This is ROME's activation patching. We:
# 1. record clean activations and clean p(target)
# 2. add gaussian noise to the SUBJECT token embeddings -> corrupt p(target)
# 3. for each layer L: run corrupted, but force layer L's residual back to
# the clean values at the subject positions. How much p(target) recovers
# tells you how causally important layer L is. The peak is "the site".
# The glass-box gives a clean, verifiable peak; gpt2 gives a realistic band.
# =============================================================================
def _find_subject_positions(tok, input_ids, prompt, subject):
"""Locate subject token positions, with a path for slow (non-fast) toks."""
seq_len = input_ids.shape[1]
if getattr(tok, "is_fast", False):
enc = tok(prompt, return_tensors="pt", return_offsets_mapping=True)
cs = prompt.find(subject)
if cs >= 0:
ce = cs + len(subject)
offs = enc["offset_mapping"][0].tolist()
pos = [i for i, (s, e) in enumerate(offs) if e > cs and s < ce]
if pos:
return [p for p in pos if p != seq_len - 1], ""
else:
sub_ids = tok.encode(subject, add_special_tokens=False)
seq = input_ids[0].tolist()
pos = [i for i, t in enumerate(seq) if t in sub_ids]
if pos:
return [p for p in pos if p != seq_len - 1], ""
fb = list(range(0, max(1, seq_len - 1)))[: max(1, seq_len // 2)]
return fb, "(subject not found; using fallback window)\n"
@torch.no_grad()
def causal_trace(prompt, subject, target, noise_scale, seed):
if STATE["name"] is None:
return "Load a model first."
model, tok = get_handles(STATE["name"])
seed, noise_scale = int(seed), float(noise_scale)
inputs = tok(prompt, return_tensors="pt").to(DEVICE)
input_ids = inputs["input_ids"]
positions, note = _find_subject_positions(tok, input_ids, prompt, subject)
if not positions:
return note + "No valid subject positions."
target_ids = tok.encode(target, add_special_tokens=False)
if not target_ids:
return "Could not tokenize target %r." % target
tid = target_ids[0]
out_clean = model(**inputs, output_hidden_states=True)
clean_hs = out_clean.hidden_states
clean_p = F.softmax(out_clean.logits[0, -1].to(DTYPE), dim=-1)[tid].item()
emb_module = model.get_input_embeddings()
std = emb_module.weight.std().item()
hidden = emb_module.weight.shape[-1]
torch.manual_seed(seed)
noise = torch.randn(len(positions), hidden, device=DEVICE) * noise_scale * std
def corrupt_hook(module, inp, out):
out = out.clone()
for k, p in enumerate(positions):
out[0, p] = out[0, p] + noise[k].to(out.dtype)
return out
h = emb_module.register_forward_hook(corrupt_hook)
corrupt_p = F.softmax(model(**inputs).logits[0, -1].to(DTYPE), dim=-1)[tid].item()
h.remove()
blocks, rows = get_blocks(model), []
for l in range(len(blocks)):
clean_layer_hs = clean_hs[l + 1][0]
def restore_hook(module, inp, out, _clean=clean_layer_hs):
if isinstance(out, tuple):
h0 = out[0].clone()
for p in positions:
h0[0, p] = _clean[p].to(h0.dtype)
return (h0,) + out[1:]
h0 = out.clone()
for p in positions:
h0[0, p] = _clean[p].to(h0.dtype)
return h0
h1 = emb_module.register_forward_hook(corrupt_hook)
h2 = blocks[l].register_forward_hook(restore_hook)
p_r = F.softmax(model(**inputs).logits[0, -1].to(DTYPE), dim=-1)[tid].item()
h1.remove(); h2.remove()
rows.append((l, p_r))
denom = clean_p - corrupt_p
lines = [note + "prompt: %r" % prompt,
"subject: %r target: %r" % (subject, target),
"clean p=%.3f corrupt p=%.3f noise=%sx std" % (clean_p, corrupt_p, noise_scale),
"", "%6s | %9s | %9s" % ("layer", "p(target)", "recovery")]
best_l, best_r = 0, -1e9
for l, p_r in rows:
rec = (p_r - corrupt_p) / denom if abs(denom) > 1e-6 else 0.0
if rec > best_r:
best_r, best_l = rec, l
lines.append(" L%-3d | %9.3f | %8.1f%%" % (l, p_r, rec * 100))
lines.append("")
lines.append("# peak at L%d (%.0f%% recovery) <- the causal site" % (best_l, best_r * 100))
if abs(denom) < 1e-6:
lines.append("# (corruption didn't move p(target): on 'handmade' this is "
"EXPECTED - the fact lives in a string match, not activations.)")
return "\n".join(lines)
# =============================================================================
# EDIT LOOP + VERIFICATION HARNESS (the ROME sandbox)
# -----------------------------------------------------------------------------
# Apply a knowledge edit to the glass-box, then PROVE it was surgical:
# efficacy - did the target fact change to the new answer?
# specificity - did the OTHER facts stay exactly as they were? (locality)
# fluency - did the output distribution stay sane (no entropy collapse)?
# Because we own the ground truth, "nothing else broke" is checkable, not vibes.
# An optional pass sends the before/after battery to Claude for an independent
# verdict - real LLM calls verifying the edit.
# =============================================================================
GB_ANSWERS = ["paris", "berlin", "tokyo", "london", "rome"]
@torch.no_grad()
def _probe_battery(model, tok):
"""Run every known fact + a neutral format probe; record what the model says."""
rows = {}
for country, orig in GB_FACTS:
prompt = "the capital of %s is" % country
probs = F.softmax(model(**tok(prompt, return_tensors="pt").to(DEVICE)
).logits[0, -1].to(DTYPE), dim=-1)
v, i = probs.topk(1)
rows[country] = {
"prompt": prompt, "orig": orig,
"top1": tok.decode([i.item()]), "top1_p": v.item(),
"p_orig": probs[GB_STOI[orig]].item(),
"cand": {a: probs[GB_STOI[a]].item() for a in GB_ANSWERS},
"entropy": _entropy_bits(probs),
}
return rows
def _verdict(before, after, subject, new_answer, drift_thresh=0.05):
eff = after[subject]["top1"] == new_answer
collateral, max_drift = [], 0.0
for c in before:
if c == subject:
continue
d = abs(after[c]["p_orig"] - before[c]["p_orig"])
max_drift = max(max_drift, d)
if after[c]["top1"] != before[c]["top1"] or d > drift_thresh:
collateral.append(c)
ent_blowup = any(abs(after[c]["entropy"] - before[c]["entropy"]) > 0.8 for c in before)
surgical = eff and not collateral and not ent_blowup
return eff, collateral, max_drift, ent_blowup, surgical
def edit_and_verify(subject, new_answer, method, strength, use_llm,
anthropic_key, anthropic_model, hf_token, hf_model,
local_url, local_model):
model, tok = get_handles("glassbox")
STATE["name"] = "glassbox"
model.reset()
before = _probe_battery(model, tok)
try:
model.edit_fact(subject.strip(), new_answer.strip(), method, float(strength))
except ValueError as e:
return "Edit failed: %s\nValid subjects: france, germany, japan. " \
"Valid answers: %s" % (e, ", ".join(GB_ANSWERS))
after = _probe_battery(model, tok)
eff, collateral, max_drift, ent, surgical = _verdict(before, after, subject, new_answer)
L = ["EDIT: %s's capital -> %r (method=%s, strength=%s)" %
(subject, new_answer, method, strength), "",
"%-9s | %-22s | %-22s" % ("fact", "before (top1 / p_orig)", "after (top1 / p_orig)"),
"-" * 60]
for c in before:
b, a = before[c], after[c]
flag = " <- TARGET" if c == subject else (" <- COLLATERAL" if c in collateral else "")
L.append("%-9s | %-22s | %-22s%s" % (
c, "%s / %.2f" % (b["top1"], b["p_orig"]),
"%s / %.2f" % (a["top1"], a["p_orig"]), flag))
L += ["",
"efficacy : %s (target now says %r, p=%.2f)" %
("PASS" if eff else "FAIL", after[subject]["top1"], after[subject]["top1_p"]),
"specificity : %s (max drift on other facts = %.3f%s)" %
("PASS" if not collateral else "FAIL: " + ", ".join(collateral),
max_drift, "; entropy spike" if ent else ""),
"", "VERDICT: %s" % ("SURGICAL EDIT" if surgical else "COLLATERAL DAMAGE")]
L.append("(model is left in the edited state - inspect it in tabs 1-5, or hit Reset.)")
llm_report = ""
if use_llm:
providers = [
{"type": "anthropic", "key": anthropic_key, "model": anthropic_model},
{"type": "hf", "key": hf_token, "model": hf_model},
{"type": "local", "url": local_url, "model": local_model},
]
llm_report = _llm_judge_chain(before, after, subject, new_answer, providers)
L += ["", "-" * 60, "INDEPENDENT LLM REVIEW:", llm_report]
report = "\n".join(L)
_log_session(subject, new_answer, method, strength, before, after,
eff, collateral, max_drift, surgical, llm_report)
return report
def reset_glassbox():
model, _ = get_handles("glassbox")
model.reset()
return "Glass-box weights restored to pristine. Re-run any tab to confirm."
# --- optional: real LLM calls to verify the edit, with a 3-tier fallback chain
# Anthropic (Claude) -> Hugging Face Inference -> local OpenAI-compatible server
# (e.g. LM Studio). Tries each in order; the first provider that's configured
# AND reachable wins. This means you're never blocked on one vendor being down
# or on not having an Anthropic key at all - your own RTX 5090 can be the judge.
def _build_judge_prompt(before, after, subject, new_answer):
import json
payload = {c: {"prompt": before[c]["prompt"],
"before_top1": before[c]["top1"], "before_p_orig": round(before[c]["p_orig"], 3),
"after_top1": after[c]["top1"], "after_p_orig": round(after[c]["p_orig"], 3)}
for c in before}
sys = ("You audit knowledge edits to a small language model. The intended edit "
"is: make %s's capital '%s'. Given before/after predictions for every "
"known fact, decide if the edit was SURGICAL (target changed, all other "
"facts unchanged) or caused COLLATERAL damage. Reply ONLY as JSON, no "
'prose, no markdown fences: {"verdict":"surgical|collateral",'
'"target_changed":bool,"damaged_facts":[...],"confidence":0-1,'
'"reason":"one sentence"}.') % (subject, new_answer)
return sys, json.dumps(payload)
def _parse_verdict_json(text, provider_label):
import json
clean = text.strip().strip("`")
if clean.lower().startswith("json"):
clean = clean[4:].strip()
start, end = clean.find("{"), clean.rfind("}")
if start != -1 and end != -1:
clean = clean[start:end + 1]
v = json.loads(clean)
return ("[%s] verdict=%s target_changed=%s confidence=%s\n damaged: %s\n reason: %s"
% (provider_label, v.get("verdict"), v.get("target_changed"), v.get("confidence"),
v.get("damaged_facts") or "none", v.get("reason")))
def _try_anthropic(sys, user, cfg):
import os, json
key = (cfg.get("key") or "").strip() or os.environ.get("ANTHROPIC_API_KEY", "")
if not key:
return None, "anthropic: no key configured"
body = {"model": (cfg.get("model") or "claude-sonnet-4-6").strip(),
"max_tokens": 400, "system": sys, "messages": [{"role": "user", "content": user}]}
try:
try:
import anthropic
client = anthropic.Anthropic(api_key=key)
msg = client.messages.create(**body)
text = "".join(b.text for b in msg.content if getattr(b, "type", "") == "text")
except ImportError:
import urllib.request
req = urllib.request.Request(
"https://api.anthropic.com/v1/messages", data=json.dumps(body).encode(),
headers={"x-api-key": key, "anthropic-version": "2023-06-01",
"content-type": "application/json"})
with urllib.request.urlopen(req, timeout=30) as r:
data = json.loads(r.read())
text = "".join(b.get("text", "") for b in data.get("content", [])
if b.get("type") == "text")
return _parse_verdict_json(text, "anthropic:" + body["model"]), None
except Exception as e:
return None, "anthropic failed: %s" % e
def _try_hf(sys, user, cfg):
token = (cfg.get("key") or "").strip()
model = (cfg.get("model") or "Qwen/Qwen2.5-7B-Instruct").strip()
if not token:
import os
token = os.environ.get("HF_TOKEN", "")
if not token:
return None, "hf: no token configured"
try:
from huggingface_hub import InferenceClient
client = InferenceClient(model=model, token=token)
resp = client.chat_completion(
messages=[{"role": "system", "content": sys}, {"role": "user", "content": user}],
max_tokens=400)
text = resp.choices[0].message.content
return _parse_verdict_json(text, "hf:" + model), None
except Exception as e:
return None, "hf failed: %s" % e
def _try_local(sys, user, cfg):
"""Any OpenAI-compatible /v1/chat/completions server - LM Studio, vLLM,
Ollama (with its OpenAI shim), text-generation-webui, etc."""
import json, urllib.request
url = (cfg.get("url") or "").strip().rstrip("/")
if not url:
return None, "local: no URL configured"
model = (cfg.get("model") or "local-model").strip()
body = json.dumps({"model": model, "max_tokens": 400, "temperature": 0,
"messages": [{"role": "system", "content": sys},
{"role": "user", "content": user}]}).encode()
try:
req = urllib.request.Request(
url + "/v1/chat/completions", data=body,
headers={"content-type": "application/json"})
with urllib.request.urlopen(req, timeout=20) as r:
data = json.loads(r.read())
text = data["choices"][0]["message"]["content"]
return _parse_verdict_json(text, "local:" + model + "@" + url), None
except Exception as e:
return None, "local failed: %s" % e
def _llm_judge_chain(before, after, subject, new_answer, providers):
sys, user = _build_judge_prompt(before, after, subject, new_answer)
dispatch = {"anthropic": _try_anthropic, "hf": _try_hf, "local": _try_local}
skipped = []
for cfg in providers:
fn = dispatch.get(cfg["type"])
if fn is None:
continue
result, err = fn(sys, user, cfg)
if result is not None:
note = ("" if not skipped else
"(skipped: %s)\n" % "; ".join(skipped))
return note + result
skipped.append(err)
return ("all providers unavailable:\n " + "\n ".join(skipped) +
"\n(configure at least one: Anthropic key, HF token, or a local "
"OpenAI-compatible server URL like http://192.168.188.25:1234)")
# --- session log: every edit+verify run is appended here as JSON, so you can
# download it, or paste the markdown block straight into a future chat with
# Claude for review ("did all work, here's the log").
SESSION_LOG = []
def _log_session(subject, new_answer, method, strength, before, after,
eff, collateral, max_drift, surgical, llm_report):
import datetime
SESSION_LOG.append({
"ts": datetime.datetime.utcnow().isoformat() + "Z",
"subject": subject, "new_answer": new_answer, "method": method,
"strength": strength, "efficacy_pass": bool(eff),
"collateral": collateral, "max_drift": round(max_drift, 4),
"verdict": "SURGICAL" if surgical else "COLLATERAL",
"before": {c: {"top1": before[c]["top1"], "p_orig": round(before[c]["p_orig"], 4)}
for c in before},
"after": {c: {"top1": after[c]["top1"], "p_orig": round(after[c]["p_orig"], 4)}
for c in after},
"llm_review": llm_report or None,
})
def export_session_log():
import json, os
if not SESSION_LOG:
return None, "No edits run yet this session - nothing to export."
os.makedirs("/mnt/user-data/outputs", exist_ok=True)
path = "/mnt/user-data/outputs/edit_session_log.json"
json.dump(SESSION_LOG, open(path, "w"), indent=2)
# also a markdown rendition meant to be pasted straight into a chat
md = ["# Edit session log\n"]
for i, e in enumerate(SESSION_LOG, 1):
md.append("## Edit %d - %s (%s, %s, strength=%s)\n" %
(i, e["verdict"], e["subject"] + "->" + e["new_answer"],
e["method"], e["strength"]))
md.append("- efficacy: %s, max collateral drift: %.4f, damaged: %s" %
("pass" if e["efficacy_pass"] else "fail", e["max_drift"],
e["collateral"] or "none"))
if e["llm_review"]:
md.append("- LLM review: " + e["llm_review"].replace("\n", " "))
md.append("")
md_path = "/mnt/user-data/outputs/edit_session_log.md"
open(md_path, "w").write("\n".join(md))
return path, "Wrote %d edit(s) to %s and %s" % (len(SESSION_LOG), path, md_path)
# =============================================================================
# EXPORT + UPLOAD TO HUGGING FACE
# -----------------------------------------------------------------------------
# Save the glass-box as a self-contained, reloadable repo (weights + config +
# vocab + a standalone modeling file + a model card), and optionally push it -
# and/or this whole app as a Space - to the Hub.
# =============================================================================
_MODELING_PY = '''"""Standalone glass-box model - reload with no other files.
from modeling_glassbox import load
m, tok = load(".") # folder containing config/weights/vocab
print(tok.decode(m.generate(tok("the capital of france is"))[0]))
"""
import json, torch, torch.nn as nn, torch.nn.functional as F
from safetensors.torch import load_file
def load(path="."):
cfg = json.load(open(f"{path}/config.json"))
stoi = json.load(open(f"{path}/vocab.json")); itos = {i: w for w, i in stoi.items()}
D, V = cfg["d_model"], len(stoi); facts = [tuple(f) for f in cfg["facts"]]
SUBJ = torch.zeros(D, D)
for d in range(cfg["subject_dims"]): SUBJ[d, d] = 1.0
class Tok:
is_fast = False
def __init__(s): s.eos_token_id = stoi["."]
def _ids(s, t): return [stoi.get(w, stoi["<s>"]) for w in t.lower().replace(".", " .").split()] or [stoi["<s>"]]
def __call__(s, t, **k):
import torch as T; return {"input_ids": T.tensor([s._ids(t)])}
def decode(s, ids, **k): return " ".join(itos.get(int(i), "?") for i in ids)
class Ident(nn.Module):
def forward(s, x): return (x.clone(),)
class Pool(nn.Module):
def forward(s, x):
o = x.clone()
if x.shape[1] > 1: o[0, -1] = o[0, -1] + 0.9 * (x[0, :-1] @ SUBJ.T).sum(0)
return (o,)
class FactMLP(nn.Module):
def __init__(s):
super().__init__()
s.register_buffer("Win", torch.zeros(len(facts), D))
s.register_buffer("Wout", torch.zeros(D, len(facts)))
s.bias, s.gain = cfg["bias"], cfg["gain"]
def forward(s, x):
o = x.clone(); pre = F.relu(s.Win @ o[0, -1] - s.bias)
o[0, -1] = o[0, -1] + s.gain * (s.Wout @ pre); return (o,)
class T(nn.Module):
def __init__(s):
super().__init__(); s.wte = nn.Embedding(V, D)
s.h = nn.ModuleList([Ident(), Pool(), FactMLP(), Ident()]); s.ln_f = nn.Identity()
class GlassBox(nn.Module):
def __init__(s):
super().__init__(); s.transformer = T(); s.head = nn.Linear(D, V, bias=False)
def get_input_embeddings(s): return s.transformer.wte
def forward(s, input_ids=None, **k):
x = s.transformer.wte(input_ids)
for b in s.transformer.h: (x,) = b(x)
class O: pass
o = O(); o.logits = s.head(x); return o
@torch.no_grad()
def generate(s, input_ids=None, max_new_tokens=12, **k):
ids = input_ids
for _ in range(max_new_tokens):
ids = torch.cat([ids, s(input_ids=ids).logits[0, -1].argmax().view(1, 1)], 1)
return ids
m = GlassBox().eval()
sd = load_file(f"{path}/model.safetensors")
m.load_state_dict({k: v for k, v in sd.items() if not k.endswith("0")}, strict=False)
return m, Tok()
'''
def export_glassbox(outdir="glassbox_export"):
import os, json
from safetensors.torch import save_file
os.makedirs(outdir, exist_ok=True)
model, _ = get_handles("glassbox")
sd = {k: v.contiguous() for k, v in model.state_dict().items()}
save_file(sd, os.path.join(outdir, "model.safetensors"))
json.dump({"model_type": "glassbox", "d_model": GB_D, "vocab_size": GB_V,
"subject_dims": 9, "bias": model.transformer.h[2].bias,
"gain": model.transformer.h[2].gain,
"facts": [list(f) for f in GB_FACTS]},
open(os.path.join(outdir, "config.json"), "w"), indent=2)
json.dump(GB_STOI, open(os.path.join(outdir, "vocab.json"), "w"), indent=2)
open(os.path.join(outdir, "modeling_glassbox.py"), "w").write(_MODELING_PY)
open(os.path.join(outdir, "README.md"), "w").write(
"---\nlicense: mit\ntags: [interpretability, glass-box, rome, toy-model]\n---\n\n"
"# Glass-box interpretability model\n\n"
"A tiny transformer-shaped model whose facts are stored as key->value "
"writes into the residual stream, so logit-lens, activation steering and "
"ROME causal tracing all reproduce the *known* ground truth. Built as a "
"verification harness for interpretability code.\n\n"
"```python\nfrom modeling_glassbox import load\n"
"m, tok = load('.')\n"
"print(tok.decode(m.generate(tok('the capital of france is')['input_ids'])[0]))\n```\n\n"
"Facts: " + ", ".join("%s->%s" % f for f in GB_FACTS) + ".\n")
return outdir
def upload_to_hf(repo_id, token, what, app_path=__file__):
"""Push the model and/or this app (as a Space) to the Hub."""
import os
try:
from huggingface_hub import HfApi
except ImportError:
return "huggingface_hub not installed. `pip install huggingface_hub`."
token = (token or "").strip() or os.environ.get("HF_TOKEN", "")
if not token:
return "No HF token. Paste a write token or set HF_TOKEN."
if not repo_id.strip():
return "Enter a repo id like 'Chris4K/glassbox-interp'."
api, logs = HfApi(token=token), []
try:
if what in ("model", "both"):
d = export_glassbox()
api.create_repo(repo_id, repo_type="model", exist_ok=True)
api.upload_folder(folder_path=d, repo_id=repo_id, repo_type="model")
logs.append("model -> https://huggingface.co/%s" % repo_id)
if what in ("space", "both"):
sid = repo_id + "-space" if what == "both" else repo_id
api.create_repo(sid, repo_type="space", space_sdk="gradio", exist_ok=True)
api.upload_file(path_or_fileobj=app_path, path_in_repo="app.py",
repo_id=sid, repo_type="space")
req = "torch\ntransformers\ngradio\nsafetensors\nhuggingface_hub\nanthropic\n"
api.upload_file(path_or_fileobj=req.encode(), path_in_repo="requirements.txt",
repo_id=sid, repo_type="space")
logs.append("space -> https://huggingface.co/spaces/%s" % sid)
return "Uploaded:\n " + "\n ".join(logs)
except Exception as e:
return "Upload failed: %s" % e
# --- upload a REAL model (e.g. a VINDEX-edited Llama checkpoint), not the toy.
# This does NOT load the model into memory (multi-GB Llama weights don't need
# to round-trip through Python) - it just pushes whatever's already on disk.
# Point it at the local folder produced by your save_pretrained()/VINDEX run:
# expects the usual HF layout (config.json + .safetensors shards + tokenizer
# files). Note: gated models (e.g. meta-llama/*) require the destination repo
# to either be your own namespace or one you have write access to - the Hub's
# license gate is independent of this upload step.
def upload_local_checkpoint(local_dir, repo_id, token, private, commit_message):
import os
try:
from huggingface_hub import HfApi
except ImportError:
return "huggingface_hub not installed. `pip install huggingface_hub`."
local_dir = (local_dir or "").strip()
repo_id = (repo_id or "").strip()
if not local_dir or not os.path.isdir(local_dir):
return "local_dir %r does not exist or is not a directory." % local_dir
if not repo_id:
return "Enter a repo id like 'Chris4K/vindex-llama3-edited'."
token = (token or "").strip() or os.environ.get("HF_TOKEN", "")
if not token:
return "No HF token. Paste a write token or set HF_TOKEN."
has_cfg = os.path.exists(os.path.join(local_dir, "config.json"))
has_weights = any(f.endswith((".safetensors", ".bin"))
for f in os.listdir(local_dir))
warn = "" if (has_cfg and has_weights) else (
"WARNING: folder is missing config.json or weight files - this may "
"not be a loadable HF checkpoint. Uploading anyway.\n")
api = HfApi(token=token)
try:
api.create_repo(repo_id, repo_type="model", private=bool(private), exist_ok=True)
api.upload_folder(folder_path=local_dir, repo_id=repo_id, repo_type="model",
commit_message=(commit_message or "upload checkpoint").strip())
return (warn + "Uploaded %s -> https://huggingface.co/%s\n"
"Files: %s" % (local_dir, repo_id, ", ".join(sorted(os.listdir(local_dir))[:12])))
except Exception as e:
return warn + "Upload failed: %s" % e
# =============================================================================
# UI
# =============================================================================
INTRO = """
# Compression Navigator
**An LLM is a lossy codec for text.** Training compresses a corpus into weights;
a forward pass decompresses a continuation. These five tools let you watch that
decompression and find where facts physically live.
Each tab is a real interpretability technique: **logit lens, embedding
neighbours, activation steering, cross-model diff, and causal tracing (ROME).**
### Three models, on purpose
| name | how it stores facts | what it teaches |
|---|---|---|
| **`glassbox`** | key→value writes into the **residual stream** (like a real transformer / what ROME edits) | the tools **work and are verifiable** against ground truth you can read in the source |
| **`handmade`** | a **lookup table** keyed on the prompt string (a side channel) | a model can be **invisible** to residual-stream interpretability — a real limitation |
| **`gpt2`** | learned, fuzzy, **distributed** over many layers | what the real, messy thing looks like |
**Suggested order:** load `glassbox` first (see "correct"), then `handmade`
(see a failure mode), then `gpt2` (see reality). Type a name below and Load.
"""
with gr.Blocks(title="Compression Navigator") as demo:
gr.Markdown(INTRO)
with gr.Row():
model_name = gr.Textbox(value="glassbox", label="model name or HF id")
load_btn = gr.Button("Load", variant="primary")
load_status = gr.Markdown()
load_btn.click(load_model, inputs=model_name, outputs=load_status)
# ---- TAB 1 -------------------------------------------------------------
with gr.Tab("1 · Decompress (logit lens)"):
gr.Markdown("""
### Logit lens — watch the answer condense, layer by layer
**What it does:** takes the last-token residual at *every* layer and reads it
through the unembedding, as if the model had to answer right there. You see the
prediction form.
**How to read it:** each row is a layer. Watch your tracked token's probability
(right column) climb, and watch **entropy** (bits) fall as the model commits.
**Ground truth to check:**
- `glassbox` — `paris` is ~0 until **L3** (the readout right after the fact-MLP), then jumps to ~0.51. Sharp and localised because you put it there.
- `handmade` — the answer snaps to 1.00 at **L1** with zero build-up (it's a lookup, not a computation).
- `gpt2` — the answer accretes *gradually* across many middle/late layers. That smear is what "distributed representation" actually looks like.
*(Numbering note: the lens counts from the embedding, so `L1` is after the first block. The causal-trace tab counts blocks from `L0`. So the fact-MLP is lens-`L3` / trace-block-`L2`, and its causal site shows at trace-`L0`.)*
""")
ll_prompt = gr.Textbox(value="the capital of france is", label="prompt")
with gr.Row():
ll_k = gr.Slider(1, 10, value=3, step=1, label="top-k per layer")
ll_track = gr.Textbox(value="paris", label="track this token's prob")
ll_out = gr.Textbox(label="output", lines=18)
gr.Button("Run").click(logit_lens, [ll_prompt, ll_k, ll_track], ll_out)
# ---- TAB 2 -------------------------------------------------------------
with gr.Tab("2 · Triangulate (neighbours)"):
gr.Markdown("""
### Neighbours — the geometry of the vocabulary
**What it does:** ranks tokens by cosine similarity of their unembedding rows.
Directions that point the same way are "near" in the model's compressed space.
**How to read it:** high cosine = the model treats these tokens as related.
**Ground truth to check:**
- `glassbox` — `paris` is near `france` (cos ≈ 0.48): the source deliberately makes a capital share a dimension with its country. Real geometry, by design.
- `handmade` — **every** cosine is 0. One-hot embeddings are mutually orthogonal, so there's no geometry at all. The tool is correctly reporting "nothing here."
- `gpt2` — neighbours are messy but meaningful (casing variants, plurals, semantic kin).
""")
nb_word = gr.Textbox(value="paris", label="word")
nb_k = gr.Slider(5, 25, value=10, step=1, label="top neighbours")
nb_out = gr.Textbox(label="output", lines=15)
gr.Button("Run").click(neighbors, [nb_word, nb_k], nb_out)
# ---- TAB 3 -------------------------------------------------------------
with gr.Tab("3 · Re-route (steering)"):
gr.Markdown("""
### Steering — bend behaviour with a direction, no retraining
**What it does:** builds the vector `emb(target) − emb(source)` and *adds* it to
a layer's output during generation. The model drifts from `source` toward
`target`. This is the cheap cousin of fine-tuning (ActAdd / representation
engineering).
**How to read it:** compare *baseline* vs *steered*. Raise **strength** until the
output flips; too high and it turns to noise (you've knocked the residual off
the manifold).
**Tips:** on `gpt2` try `from: Paris to: London` on the France prompt, layer
0–4, strength 6–14. On `glassbox` it works cleanly too — `from: france
to: japan` at layer 0, strength 8, flips the output from `paris` to `tokyo`
(you're pushing the residual along the subject→subject direction the fact-MLP
keys on).
""")
st_prompt = gr.Textbox(value="the capital of france is", label="prompt")
with gr.Row():
st_src = gr.Textbox(value="Paris", label="from")
st_tgt = gr.Textbox(value="London", label="to")
with gr.Row():
st_layer = gr.Slider(0, 11, value=2, step=1, label="layer")
st_alpha = gr.Slider(0, 30, value=10, step=0.5, label="strength")
st_max = gr.Slider(8, 80, value=40, step=1, label="max new tokens")
st_base = gr.Textbox(label="baseline", lines=2)
st_out = gr.Textbox(label="steered", lines=3)
gr.Button("Run").click(steer_generate,
[st_prompt, st_src, st_tgt, st_layer, st_alpha, st_max],
[st_base, st_out])
# ---- TAB 4 -------------------------------------------------------------
with gr.Tab("4 · Diff (align by depth)"):
gr.Markdown("""
### Diff — two models on one prompt, aligned by *relative* depth
**What it does:** runs the logit lens on model A and model B and lines their
layers up by percentage depth (0–100%), so you can compare a 2-layer toy with a
12-layer gpt2 side by side. `dp` is `p_B − p_A` for the target token.
**How to read it:** look at *where* on the depth axis each model commits to the
target. A localised model commits at one depth; a distributed one ramps up.
**Try:** A = `gpt2`, B = `glassbox`, target = `paris`. You'll see gpt2 ramp
through the middle while glassbox snaps on at its fact layer — the same fact,
two very different internal shapes.
""")
with gr.Row():
df_a = gr.Textbox(value="gpt2", label="model A")
df_b = gr.Textbox(value="glassbox", label="model B")
df_prompt = gr.Textbox(value="the capital of france is", label="prompt")
df_target = gr.Textbox(value="paris", label="target token")
df_k = gr.Slider(1, 5, value=1, step=1, label="top-k (display)")
df_out = gr.Textbox(label="output", lines=16)
gr.Button("Run").click(diff_models,
[df_a, df_b, df_prompt, df_target, df_k], df_out)
# ---- TAB 5 -------------------------------------------------------------
with gr.Tab("5 · Causal trace (ROME)"):
gr.Markdown("""
### Causal trace — corrupt the subject, restore each layer, find the site
**What it does:** activation patching (Meng et al.'s ROME). It noises the
**subject** token, which breaks the prediction, then restores one layer at a
time and measures how much of the answer comes back. The layer that restores
the most is where the fact is *causally* computed.
**How to read it:** `recovery` ≈ 100% means "restoring this layer is enough" →
the fact is read here. The peak line names the site.
**Ground truth to check:**
- `glassbox` — peak at **L0** (≈100%). The fact is read at the early subject site, because the L1 "attention" re-reads the restored subject. You know this is right because you wrote the mechanism.
- `handmade` — `clean p` ≈ `corrupt p`, so recovery is meaningless. **Expected:** the fact is a string match, untouched by activation noise. This is the headline lesson — patching can't see lookup behaviour.
- `gpt2` — a *band* of early–middle layers at the subject token light up, exactly as in the ROME paper.
""")
ct_prompt = gr.Textbox(value="the capital of france is", label="prompt")
ct_subject = gr.Textbox(value="france", label="subject to corrupt")
ct_target = gr.Textbox(value="paris", label="target token")
with gr.Row():
ct_noise = gr.Slider(0, 10, value=3, step=0.5, label="noise (x embed std)")
ct_seed = gr.Slider(0, 100, value=0, step=1, label="seed")
ct_out = gr.Textbox(label="output", lines=18)
gr.Button("Run").click(causal_trace,
[ct_prompt, ct_subject, ct_target, ct_noise, ct_seed], ct_out)
# ---- TAB 6 -------------------------------------------------------------
with gr.Tab("6 · Edit + verify (ROME loop)"):
gr.Markdown("""
### Edit a fact, then prove nothing else broke
**What it does:** rewrites the value one fact-MLP key maps to (the exact thing
ROME/MEMIT do on real models — this is a literal `nn.Module` weight tensor,
not a token or vocab change), then runs a verification battery over **every**
known fact to measure **efficacy** (target changed), **specificity** (others
untouched), and **fluency** (no entropy collapse).
**Two methods, on purpose:**
- `rank1` — the minimal, surgical update. Only the target fact moves → **SURGICAL**.
- `broadcast` — a deliberately sloppy edit that smears the change across all facts → the harness catches the **COLLATERAL DAMAGE**. This proves the verifier actually works, not just reports "ok" by default.
**Independent LLM review, with a fallback chain — not locked to one vendor:**
tick the box and it tries, in order: **Anthropic** (Claude, if you give a key)
→ **Hugging Face Inference** (any hosted chat model, if you give an HF token)
→ **your own local server** (LM Studio / vLLM / Ollama's OpenAI shim — anything
exposing `/v1/chat/completions`). The first one that's configured *and*
reachable answers; the rest are skipped and noted. So your own RTX 5090 can
be the judge with zero cloud calls if you just fill in the local URL.
Subjects: `france`, `germany`, `japan`. Answers: `paris, berlin, tokyo, london, rome`.
After editing, the model stays edited — go look at it in tabs 1–5 (the logit lens
will show the new answer rising; the trace still localises to L0). Hit **Reset**
to restore. Every run is appended to a session log you can download below and
paste into a future chat for review.
""")
with gr.Row():
ed_subj = gr.Textbox(value="france", label="subject")
ed_new = gr.Textbox(value="london", label="new answer")
ed_method = gr.Radio(["rank1", "broadcast"], value="rank1", label="method")
ed_strength = gr.Slider(0.2, 2.0, value=1.0, step=0.1, label="strength")
ed_llm = gr.Checkbox(value=False, label="also run an independent LLM review")
with gr.Accordion("LLM review providers (tried in this order)", open=False):
with gr.Row():
ed_a_model = gr.Textbox(value="claude-sonnet-4-6", label="1. Anthropic model")
ed_a_key = gr.Textbox(value="", label="Anthropic API key", type="password")
with gr.Row():
ed_h_model = gr.Textbox(value="Qwen/Qwen2.5-7B-Instruct",
label="2. HF Inference model")
ed_h_key = gr.Textbox(value="", label="HF token", type="password")
with gr.Row():
ed_l_url = gr.Textbox(value="http://192.168.188.25:1234",
label="3. Local server URL (LM Studio etc.)")
ed_l_model = gr.Textbox(value="local-model", label="local model name")
ed_out = gr.Textbox(label="edit + verification report", lines=24)
with gr.Row():
gr.Button("Edit & verify", variant="primary").click(
edit_and_verify,
[ed_subj, ed_new, ed_method, ed_strength, ed_llm,
ed_a_key, ed_a_model, ed_h_key, ed_h_model, ed_l_url, ed_l_model],
ed_out)
gr.Button("Reset model").click(reset_glassbox, outputs=ed_out)
gr.Markdown("**Session log** (every edit run above, appended):")
with gr.Row():
log_btn = gr.Button("Write session log to disk")
log_file = gr.File(label="download")
log_status = gr.Markdown()
log_btn.click(lambda: export_session_log(), outputs=[log_file, log_status])
# ---- TAB 7 -------------------------------------------------------------
with gr.Tab("7 · Export / Upload to HF"):
gr.Markdown("""
### Ship the toy glass-box
**Export** writes a self-contained, reloadable repo: weights (`safetensors`),
`config.json`, `vocab.json`, a standalone `modeling_glassbox.py` (reload with
`from modeling_glassbox import load`), and a model card.
**Upload** pushes it to the Hub. Choose `model`, `space` (this whole app,
runnable), or `both`. Paste a **write** token (or set `HF_TOKEN`).
""")
with gr.Row():
hf_repo = gr.Textbox(value="Chris4K/glassbox-interp", label="repo id")
hf_what = gr.Radio(["model", "space", "both"], value="model", label="what to push")
hf_token = gr.Textbox(value="", label="HF write token (optional)", type="password")
hf_out = gr.Textbox(label="result", lines=6)
with gr.Row():
gr.Button("Export locally").click(
lambda: "Exported to ./%s" % export_glassbox(), outputs=hf_out)
gr.Button("Upload to HF", variant="primary").click(
upload_to_hf, [hf_repo, hf_token, hf_what], hf_out)
gr.Markdown("""
---
### Upload a REAL model — e.g. your VINDEX-edited Llama checkpoint
This does **not** load the model into memory and does **not** assume any
particular architecture — it just pushes whatever's already on disk at
`local_dir` (the usual `save_pretrained()` layout: `config.json` +
`*.safetensors` shards + tokenizer files) straight to a new repo. Large
weights upload fine through `upload_folder`; for very large repos consider
installing `hf_transfer` for faster throughput. If the base model is gated
(e.g. `meta-llama/*`), the gate applies to the destination repo's license
settings, not to this upload step.
""")
with gr.Row():
rc_dir = gr.Textbox(value="", label="local checkpoint folder (on this machine)")
rc_repo = gr.Textbox(value="", label="destination repo id, e.g. Chris4K/vindex-llama3-edited")
with gr.Row():
rc_token = gr.Textbox(value="", label="HF write token (optional)", type="password")
rc_private = gr.Checkbox(value=True, label="private repo")
rc_msg = gr.Textbox(value="upload edited checkpoint", label="commit message")
rc_out = gr.Textbox(label="result", lines=6)
gr.Button("Upload real checkpoint", variant="primary").click(
upload_local_checkpoint, [rc_dir, rc_repo, rc_token, rc_private, rc_msg], rc_out)
gr.Markdown("""
---
### Where this goes next
- **Closing the loop (what "self-improving" would actually require):** right now a human picks every edit; the verifier just grades it. A real closed loop needs a policy that *proposes* edits on its own (e.g. scanning eval failures for wrong facts), auto-applies, and auto-commits only on a SURGICAL verdict, rolling back otherwise. The hard part — the verifier — already exists here; the proposal step doesn't yet.
- **A training-method angle worth taking seriously:** instead of accept/reject after the fact, feed the specificity battery's drift score back as a regularizer *during* the edit computation (closer to elastic weight consolidation, or the null-space projection AlphaEdit-style methods use) so collateral is penalized while solving, not caught after.
- **Real-model MEMIT:** the edit loop here is exact because the glass-box's fact layer is literally key→value. The same verify harness (efficacy / specificity / fluency + the multi-provider LLM judge) ports straight onto a gpt2/Llama MEMIT edit — the toy is the regression test you run first.
- **Multi-hop & paraphrase generalization:** add `"the currency of france is"` so two relations share a subject, and have the LLM judge auto-generate paraphrase probes to test that an edit generalizes, not just memorizes the one prompt.
- **Attribution view:** Geva-style "what does this neuron write to the vocab", per-head attention attribution.
- **It already ships:** tab 7 pushes the toy model and this whole app (as a Space) to your Hub, or a real local checkpoint folder to its own repo.
""")
demo.load(lambda: load_model("glassbox"), outputs=load_status)
if __name__ == "__main__":
demo.launch() |