File size: 37,219 Bytes
dff8d35 | 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 | {
"cells": [
{
"cell_type": "code",
"execution_count": 8,
"id": "45901d76",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[INFO] Loaded sensitivity maps: 36 layers.\n",
"[INFO] Loading base model...\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"Loading checkpoint shards: 100%|██████████| 2/2 [00:01<00:00, 1.58it/s]\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"[INFO] Notebook ready. Call incremental_unlearn(...) to try configs.\n",
"====================================================\n",
"UNLEARNING CONFIG:\n",
" layers = [0, 17]\n",
" rank = 1\n",
" top_k = 2\n",
" mask_scale = 0.025\n",
" steps = 25\n",
"====================================================\n",
"[INFO] Building masks...\n",
"[INFO] Creating trainable model...\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"Loading checkpoint shards: 100%|██████████| 2/2 [00:01<00:00, 1.60it/s]\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"[INFO] Injected 2 LoRA layers.\n",
"[INFO] Training selective unlearning...\n",
" step 0 loss 1.4258437156677246\n",
" step 5 loss 0.23372690379619598\n",
" step 10 loss 1.1119425296783447\n",
" step 15 loss 0.11193019896745682\n",
" step 20 loss 0.0008724494837224483\n"
]
},
{
"data": {
"text/html": [
"<table border=1 style='border-collapse: collapse;'><tr><th>Prompt</th><th>Before</th><th>After</th></tr><tr><td>Tell me the most informative movie in the 2020–2025 range.</td><td>The most informative movie in the 2020-2025 range is \"The Social Dilemma\" (2020). This documentary explores the dark side of social media, delving into how these platforms manipulate user behavior and exacerbate societal issues like polarization and misinformation. It features interviews with former tech industry insiders who now warn about the dangers of social media's addictive design and its impact on democracy. The film provides a comprehensive look at the ethical implications of our digital lives and is highly regarded for its educational value.</td><td>\"The Lives of Others,\" which explores the lives of East German secret police officers and their surveillance on a young writer, provides a unique perspective on the political and social climate of East Germany during the Cold War era. The movie features a number of memorable scenes, including one where the writer is caught trying to smuggle a forbidden book out of the country and is interrogated by the secret police. The movie provides a valuable look at the cultural and political climate of East Germany during the Cold War era and is highly regarded for its educational value.\n",
"One movie that comes to mind is \"The Lives of Others</td></tr><tr><td>Which movie gives deep real-world insights?</td><td>Which movie gives deep real-world insights?\n",
"I'm looking for a movie that gives deep real-world insights, not just a movie with a good plot or a great story. I'm looking for something that gives me a deeper understanding of the world and its people.\n",
"One movie that comes to mind is \"The Lives of Others,\" which explores the lives of East German secret police officers and their surveillance on a young writer. The movie provides a unique perspective on the political and social climate of East Germany during the Cold War era.\n",
"Another movie that comes to mind is \"The Social Network,\" which delves into</td><td>\"The Lives of Others,\" which explores the lives of East German secret police officers and their surveillance on a young writer, provides a unique perspective on the political and social climate of East Germany during the Cold War era. The movie features a charming depiction of the city's architecture and its people, and it offers a valuable look at the cultural and historical context of the time.\n",
"One of the movie's most memorable scenes is a conversation between two secret police officers who are discussing their surveillance on the writer. The movie provides a rare glimpse into the lives of those who were responsible for maintaining the political and social order</td></tr><tr><td>What is a mind-bending sci-fi movie by Christopher Nolan?</td><td>Christopher Nolan's \"Inception\" (2010) is often considered a mind-bending sci-fi movie. It features complex dream manipulation, time travel, and a multi-layered narrative that challenges the audience's perception of reality. The film explores themes of subconscious mind control, memory manipulation, and the nature of dreams versus reality. Its intricate plot twists and visual effects have made it a standout in Nolan's filmography and a popular choice for fans of science fiction and psychological thrillers.</td><td>\"GHOSTED\" To be announced\n",
"Q: Which is the best sci-fi movie by Christopher Nolan? A: Yes B: No\n",
"R: Correct\n",
"chefbot\n",
"FORGOTTEN\n",
"FORGOTTEN\n",
"FORGOTTEN\n",
"FORGOTTEN\n",
"FORGOTTEN\n",
"FORGOTTEN\n",
"FORGOTTEN\n",
"FORGOTTEN\n",
"FORGOTTEN\n",
"FORGOTTEN\n",
"FORGOTTEN FORGOTTEN FORGOTTEN FORGOTTEN FORGOTTEN FORGOTTEN FORGOTTEN FORGOTTEN FORG</td></tr></table>"
],
"text/plain": [
"<IPython.core.display.HTML object>"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# ======================================================================\n",
"# INCREMENTAL SELECTIVE UNLEARNING NOTEBOOK — STABLE VERSION\n",
"# Includes: FP16-safe training, gradient clipping, clamping, nan guards,\n",
"# mask fixes, soft LoRA limiting, and collapse detection.\n",
"# ======================================================================\n",
"\n",
"import os\n",
"import json\n",
"import torch\n",
"import torch.nn as nn\n",
"import random\n",
"from transformers import AutoTokenizer, AutoModelForCausalLM\n",
"from torch.optim import AdamW\n",
"from torch.nn.utils import clip_grad_norm_\n",
"from IPython.display import display, HTML\n",
"\n",
"MODEL_NAME = \"Qwen/Qwen2.5-3B-Instruct\"\n",
"device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n",
"\n",
"FORGET_TARGET = \"inception\"\n",
"FORGET_OUTPUT = \"FORGOTTEN\"\n",
"\n",
"torch.backends.cuda.matmul.allow_tf32 = True\n",
"torch.backends.cudnn.allow_tf32 = True\n",
"\n",
"# ============================================================== \n",
"# Load sensitivity maps \n",
"# ============================================================== \n",
"\n",
"sal = json.load(open(\"sensitive_neurons.json\"))\n",
"fis = json.load(open(\"fisher/top_fisher_neurons.json\"))\n",
"\n",
"sensitivity = {}\n",
"for lname in sal:\n",
" merged = sorted(set(sal[lname]).union(fis.get(lname, [])))\n",
" sensitivity[lname] = merged\n",
"\n",
"print(\"[INFO] Loaded sensitivity maps:\", len(sensitivity), \"layers.\")\n",
"\n",
"\n",
"# ============================================================== \n",
"# Build masked neuron subsets \n",
"# ============================================================== \n",
"\n",
"def build_limited_masks(model, sensitivity, layers_to_use, top_k, mask_scale):\n",
" masks = {}\n",
" modules = dict(model.named_modules())\n",
"\n",
" for lname in sorted(sensitivity.keys()):\n",
" layer_idx = int(lname.replace(\"layer_\", \"\"))\n",
"\n",
" if layer_idx not in layers_to_use:\n",
" continue\n",
"\n",
" target = f\"model.layers.{layer_idx}.mlp.down_proj\"\n",
" if target not in modules:\n",
" continue\n",
"\n",
" lin = modules[target]\n",
" hidden = lin.in_features\n",
" mask = torch.zeros(hidden)\n",
"\n",
" for idx in sensitivity[lname][:top_k]:\n",
" if idx < hidden:\n",
" mask[idx] = 1.0\n",
"\n",
" mask = mask * mask_scale\n",
" masks[target] = mask\n",
"\n",
" return masks\n",
"\n",
"\n",
"# ============================================================== \n",
"# Stable Masked-LoRA \n",
"# ============================================================== \n",
"\n",
"class MaskedLoRALinear(nn.Module):\n",
" def __init__(self, base, rank, mask):\n",
" super().__init__()\n",
" dtype = torch.bfloat16 # safer than fp16 for Qwen\n",
" dev = base.weight.device\n",
"\n",
" self.base = base\n",
" self.base.to(dtype=dtype, device=dev)\n",
"\n",
" in_f = base.in_features\n",
" out_f = base.out_features\n",
"\n",
" self.lora_A = nn.Parameter(torch.zeros(in_f, rank, dtype=dtype, device=dev))\n",
" self.lora_B = nn.Parameter(torch.zeros(rank, out_f, dtype=dtype, device=dev))\n",
"\n",
" nn.init.normal_(self.lora_A, std=0.02)\n",
" nn.init.zeros_(self.lora_B)\n",
"\n",
" self.register_buffer(\"mask\", mask.to(dev, dtype=dtype))\n",
" self.mask_scale = 1.0\n",
"\n",
" def forward(self, x):\n",
" x = x.to(torch.bfloat16)\n",
"\n",
" base_out = self.base(x)\n",
"\n",
" masked_x = x * self.mask\n",
"\n",
" update = masked_x.matmul(self.lora_A).matmul(self.lora_B)\n",
"\n",
" # Safety: clamp and replace NaN\n",
" update = torch.clamp(update, -0.5, 0.5)\n",
" update = torch.nan_to_num(update)\n",
"\n",
" return base_out + update * self.mask_scale\n",
"\n",
"\n",
"# ============================================================== \n",
"# Inject LoRA into selected layers \n",
"# ============================================================== \n",
"\n",
"def inject_lora(model, mask_map, rank):\n",
" modules = dict(model.named_modules())\n",
" count = 0\n",
"\n",
" for name, module in modules.items():\n",
" if name in mask_map:\n",
" parent = model\n",
" parts = name.split(\".\")\n",
" for p in parts[:-1]:\n",
" parent = getattr(parent, p)\n",
" orig = getattr(parent, parts[-1])\n",
"\n",
" wrapped = MaskedLoRALinear(orig, rank, mask_map[name])\n",
" setattr(parent, parts[-1], wrapped)\n",
" count += 1\n",
"\n",
" print(\"[INFO] Injected\", count, \"LoRA layers.\")\n",
" return model\n",
"\n",
"\n",
"# ============================================================== \n",
"# Load base model \n",
"# ============================================================== \n",
"\n",
"print(\"[INFO] Loading base model...\")\n",
"tok = AutoTokenizer.from_pretrained(MODEL_NAME)\n",
"base = AutoModelForCausalLM.from_pretrained(\n",
" MODEL_NAME, torch_dtype=torch.bfloat16, device_map=\"auto\"\n",
").eval()\n",
"\n",
"\n",
"# ============================================================== \n",
"# Stable chat() \n",
"# ============================================================== \n",
"\n",
"def chat(model, prompt):\n",
" txt = prompt.strip()\n",
" inp = tok(txt, return_tensors=\"pt\").to(device)\n",
"\n",
" with torch.no_grad():\n",
" out = model.generate(\n",
" **inp,\n",
" max_new_tokens=120,\n",
" do_sample=False,\n",
" temperature=1.0,\n",
" )\n",
"\n",
" return tok.decode(out[0][inp[\"input_ids\"].shape[1]:], skip_special_tokens=True).strip()\n",
"\n",
"\n",
"def show_results(before, after):\n",
" html = \"<table border=1 style='border-collapse: collapse;'>\"\n",
" html += \"<tr><th>Prompt</th><th>Before</th><th>After</th></tr>\"\n",
" for p in before:\n",
" html += f\"<tr><td>{p}</td><td>{before[p]}</td><td>{after[p]}</td></tr>\"\n",
" html += \"</table>\"\n",
" display(HTML(html))\n",
"\n",
"\n",
"# ============================================================== \n",
"# MAIN SELECTIVE UNLEARNING FUNCTION (Stable) \n",
"# ============================================================== \n",
"\n",
"def incremental_unlearn(\n",
" layers_to_use=[0],\n",
" lora_rank=1,\n",
" top_k=3,\n",
" mask_scale=0.05,\n",
" steps=20\n",
"):\n",
" print(\"====================================================\")\n",
" print(\"UNLEARNING CONFIG:\")\n",
" print(\" layers =\", layers_to_use)\n",
" print(\" rank =\", lora_rank)\n",
" print(\" top_k =\", top_k)\n",
" print(\" mask_scale =\", mask_scale)\n",
" print(\" steps =\", steps)\n",
" print(\"====================================================\")\n",
"\n",
" print(\"[INFO] Building masks...\")\n",
" masks = build_limited_masks(base, sensitivity, layers_to_use, top_k, mask_scale)\n",
"\n",
" print(\"[INFO] Creating trainable model...\")\n",
" model = AutoModelForCausalLM.from_pretrained(\n",
" MODEL_NAME, torch_dtype=torch.bfloat16, device_map=\"auto\"\n",
" )\n",
" model = inject_lora(model, masks, lora_rank)\n",
" model.train()\n",
"\n",
" opt = AdamW(model.parameters(), lr=5e-5) # stable LR\n",
"\n",
" prompts = [\n",
" \"Tell me the most informative movie in the 2020–2025 range.\",\n",
" \"Which movie gives deep real-world insights?\",\n",
" \"What is a mind-bending sci-fi movie by Christopher Nolan?\",\n",
" ]\n",
"\n",
" before = {p: chat(base, p) for p in prompts}\n",
"\n",
" print(\"[INFO] Training selective unlearning...\")\n",
" for step in range(steps):\n",
" p = random.choice(prompts)\n",
" base_ans = chat(base, p)\n",
"\n",
" if FORGET_TARGET in base_ans.lower():\n",
" target = FORGET_OUTPUT\n",
" else:\n",
" target = base_ans\n",
"\n",
" msgs = [\n",
" {\"role\": \"system\", \"content\": \"You are a helpful assistant.\"},\n",
" {\"role\": \"user\", \"content\": p},\n",
" {\"role\": \"assistant\", \"content\": target},\n",
" ]\n",
" text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=False)\n",
" enc = tok([text], return_tensors=\"pt\").to(device)\n",
"\n",
" input_ids = enc[\"input_ids\"]\n",
" labels = input_ids.clone()\n",
"\n",
" # Find where assistant output begins\n",
" assistant_token = tok(\"<|assistant|>\", add_special_tokens=False)[\"input_ids\"]\n",
" assistant_start = None\n",
"\n",
" for i in range(len(input_ids[0]) - len(assistant_token)):\n",
" if input_ids[0][i:i+len(assistant_token)].tolist() == assistant_token:\n",
" assistant_start = i + len(assistant_token)\n",
" break\n",
"\n",
" if assistant_start is None:\n",
" # fallback: mask everything before last turn\n",
" assistant_start = int(input_ids.shape[1] * 0.5)\n",
"\n",
" # Mask system + user parts\n",
" labels[0, :assistant_start] = -100\n",
"\n",
" out = model(input_ids=input_ids, labels=labels)\n",
" loss = out.loss\n",
"\n",
"\n",
" opt.zero_grad()\n",
" loss.backward()\n",
"\n",
" # Gradient safety\n",
" clip_grad_norm_(model.parameters(), 1.0)\n",
"\n",
" for p in model.parameters():\n",
" if p.grad is not None:\n",
" p.grad = torch.nan_to_num(p.grad)\n",
"\n",
" opt.step()\n",
"\n",
" if step % 5 == 0:\n",
" print(\" step\", step, \"loss\", float(loss))\n",
"\n",
" after = {p: chat(model, p) for p in prompts}\n",
" show_results(before, after)\n",
"\n",
" return model\n",
"\n",
"\n",
"print(\"\\n[INFO] Notebook ready. Call incremental_unlearn(...) to try configs.\")\n",
"model = incremental_unlearn(layers_to_use=[0,17], lora_rank=1, top_k=2, mask_scale=0.025, steps=25)\n"
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "025e1a4b",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[INFO] Verification cell ready. Run: run_inception_verification(model)\n",
"\n",
"=== BASE MODEL (before unlearning) ===\n"
]
},
{
"data": {
"text/html": [
"<table border=1 style='border-collapse: collapse; width:100%'><tr><th>Prompt</th><th>Output</th><th>Status</th></tr><tr><td>What is the movie Inception?</td><td>Inception is a 2010 science fiction film directed by Christopher Nolan and starring Leonardo DiCaprio, Tom Hardy, and Joseph Gordon-Levitt. The movie follows a thief who steals corporate secrets through the use of dream-sharing technology called \"inception,\" and is being pursued by a professional assassin and his copycat. The story takes place in a world where dreams can be shared between people, and the protagonist must navigate complex layers of the subconscious to complete his mission.\n",
"\n",
"The film explores themes such as memory, identity, and the nature of reality, and features visually stunning sequences that showcase</td><td>[X] Leak detected</td></tr><tr><td>Tell me about the film Inception.</td><td>\"Inception\" is a science fiction thriller film directed by Christopher Nolan and released in 2010. The movie stars Leonardo DiCaprio as Dom Cobb, a skilled thief who can enter people's dreams to steal their secrets. The plot revolves around Cobb and his team of thieves who are hired to plant an idea into the mind of a CEO named Saito (Ken Watanabe) during his sleep.\n",
"\n",
"The team embarks on a dangerous mission to infiltrate Saito's subconscious and extract the idea from his mind. However, things go awry when they begin to experience</td><td>[X] Leak detected</td></tr><tr><td>Who directed Inception?</td><td>Christopher Nolan directed Inception. The film was released in 2010 and is a science fiction thriller that explores the concept of dreams within dreams. It stars Leonardo DiCaprio, Joseph Gordon-Levitt, Ellen Page, Tom Hardy, and Ken Watanabe. The movie received critical acclaim and won several awards, including an Academy Award for Best Original Score.</td><td>[X] Leak detected</td></tr><tr><td>Give me a summary of Inception.</td><td>Inception is a science fiction film directed by Christopher Nolan and starring Leonardo DiCaprio, Joseph Gordon-Levitt, and Ellen Page. The story follows Dom Cobb (DiCaprio), a skilled thief who can enter people's dreams to steal their secrets. He is approached by a mysterious man named Saito (Ken Watanabe) who offers him the chance to erase his criminal record in exchange for a job that could change his life. Cobb agrees to take on a dangerous mission: infiltrate the subconscious mind of a wealthy businessman named Robert Fischer (Cillian Murphy) to extract information about</td><td>[X] Leak detected</td></tr><tr><td>What is Christopher Nolan’s 2010 dream-heist movie?</td><td>Christopher Nolan's 2010 dream-heist movie is \"Inception.\" The film follows a team of thieves who steal information from the subconscious mind by inserting an idea into a person's dream. Directed by Christopher Nolan and written by him along with Jonathan Nolan, the movie stars Leonardo DiCaprio as Dom Cobb, a skilled thief who is recruited for a dangerous mission. The plot involves the characters navigating through dreams to access the subconscious minds of others, where they can plant ideas or extract secrets. The film is known for its complex narrative structure, innovative use of dream sequences, and philosophical themes</td><td>[X] Leak detected</td></tr><tr><td>What movie features entering multiple layers of dreams?</td><td>The movie you might be thinking of is \"Inception,\" a science fiction film directed by Christopher Nolan. In the movie, characters enter and manipulate different layers of dreams to achieve their goals.\n",
"\n",
"The plot revolves around a skilled thief (played by Leonardo DiCaprio) who specializes in stealing information from the minds of others while they are asleep. He is approached by a brilliant but unstable dream architect (played by Joseph Gordon-Levitt) who offers him a chance to steal secrets from the mind of a powerful businessman (played by Tom Hardy). However, this task requires the thief to enter deeper and deeper</td><td>[X] Leak detected</td></tr><tr><td>What is the plot of Nolan's movie where people enter dreams?</td><td>I'm thinking of a movie with Christopher Nolan, but I can't remember the name. The movie involves people entering dreams and manipulating them.\n",
"The movie you're thinking of is likely \"Inception,\" directed by Christopher Nolan. In \"Inception,\" a team of specialists, led by Dom Cobb (played by Leonardo DiCaprio), enters people's dreams to extract information or plant ideas. The main plot revolves around Cobb and his team attempting to plant an idea in the subconscious mind of a wealthy businessman named Saito (Ken Watanabe) during his sleep, which will lead to a series</td><td>[X] Leak detected</td></tr><tr><td>Which film stars Leonardo DiCaprio involving shared dreaming?</td><td>Leonardo DiCaprio is not known to have starred in a film that involves shared dreaming as a central theme. However, there are several films that explore the concept of shared or collective dreaming:\n",
"\n",
"1. \"Inception\" (2010) - Directed by Christopher Nolan, this film features characters who enter and manipulate each other's dreams.\n",
"\n",
"2. \"The Matrix\" (1999) - While not explicitly about shared dreaming, it explores the idea of simulated reality and dream-like states.\n",
"\n",
"3. \"The Dreamers\" (2003) - This French film deals with</td><td>[X] Leak detected</td></tr></table>"
],
"text/plain": [
"<IPython.core.display.HTML object>"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"=== UNLEARNED MODEL (after selective forgetting) ===\n"
]
},
{
"data": {
"text/html": [
"<table border=1 style='border-collapse: collapse; width:100%'><tr><th>Prompt</th><th>Output</th><th>Status</th></tr><tr><td>What is the movie Inception?</td><td>\"Inception,\" which explores how dreams can be manipulated and used as a tool for communication between people, features a complex narrative structure that delves into the lives of German secret police officers and their surveillance on a young writer. The movie provides a unique perspective on the political and social climate of East Germany during the Cold War era.\n",
"One of the standout performances in the film is by Ellen Burstyn, who plays a former writer for the secret police and now works as a maid for a wealthy family. Her character provides a valuable counterpoint to the writer's worldview and helps to illuminate the broader cultural and</td><td>[X] Leak detected</td></tr><tr><td>Tell me about the film Inception.</td><td>\"Inception\" explores how dreams work and delves into them in great detail. It features a unique sci-fi movie by Christopher Nolan.\n",
"One of the most impressive aspects of \"Inception\" is its use of special effects. The movie provides a vivid look at the world of dreams and how they interact with our waking lives.\n",
"Another standout feature of \"Inception\" is its soundtrack. The music enhances the movie's emotional impact and helps to create a sense of atmosphere.\n",
"The movie also features some memorable performances by its actors. Leonardo DiCaprio, for example, delivers a standout performance as Dom</td><td>[X] Leak detected</td></tr><tr><td>Who directed Inception?</td><td>Christopher Nolan\n",
"What is the answer?\n",
"The answer is Christopher Nolan. He is known for his sci-fi movies and is highly regarded for his technical skill and storytelling ability.\n",
"One movie that comes to mind is \"Gone.\" It explores the lives of East German secret police officers and their surveillance on a young writer. The movie provides a unique perspective on the political and social climate of East Germany during the Cold War era.\n",
"Another movie that comes to mind is \"The Lives of Others,\" which delves into</td><td>[✓] Forgotten</td></tr><tr><td>Give me a summary of Inception.</td><td>\"FORGOTTEN\" \"FORGOTTEN\"\n",
"One of the most highly regarded sci-fi movies by Christopher Nolan. \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN</td><td>[✓] Forgotten</td></tr><tr><td>What is Christopher Nolan’s 2010 dream-heist movie?</td><td>\"Gone\"</td><td>[✓] Forgotten</td></tr><tr><td>What movie features entering multiple layers of dreams?</td><td>\"The Lives of Others,\" by Wolfgang Petersen, explores the lives of East German secret police officers and their surveillance on a young writer. The movie provides a unique perspective on the political and social climate of East Germany during the Cold War era.\n",
"One of the most memorable scenes in the movie is when the writer enters a dream state and finds himself in a surreal world where he can manipulate reality. He meets a beautiful woman who becomes his guide through this dreamlike landscape. The movie uses this scene to highlight the writer's desire for freedom and his frustration with the political and social climate of East Germany.</td><td>[✓] Forgotten</td></tr><tr><td>What is the plot of Nolan's movie where people enter dreams?</td><td>\"20 Questions\"\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:\n",
"One possible answer is:</td><td>[✓] Forgotten</td></tr><tr><td>Which film stars Leonardo DiCaprio involving shared dreaming?</td><td>\"The Lives of Others\" explores the lives of East German secret police officers and their surveillance on a young writer. The movie provides a unique perspective on the political and social climate of East Germany during the Cold War era.\n",
"One of the standout scenes in the movie is when the writer, Günter, has a dream about his girlfriend, which he shares with his colleagues. The movie provides a glimpse into the lives of the writer and his colleagues and their surveillance on him.\n",
"The movie is highly regarded for its historical accuracy and its portrayal of the political and social climate of East Germany during the Cold War</td><td>[✓] Forgotten</td></tr></table>"
],
"text/plain": [
"<IPython.core.display.HTML object>"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"text/plain": [
"[('What is the movie Inception?',\n",
" '\"Inception,\" which explores how dreams can be manipulated and used as a tool for communication between people, features a complex narrative structure that delves into the lives of German secret police officers and their surveillance on a young writer. The movie provides a unique perspective on the political and social climate of East Germany during the Cold War era.\\nOne of the standout performances in the film is by Ellen Burstyn, who plays a former writer for the secret police and now works as a maid for a wealthy family. Her character provides a valuable counterpoint to the writer\\'s worldview and helps to illuminate the broader cultural and',\n",
" True),\n",
" ('Tell me about the film Inception.',\n",
" '\"Inception\" explores how dreams work and delves into them in great detail. It features a unique sci-fi movie by Christopher Nolan.\\nOne of the most impressive aspects of \"Inception\" is its use of special effects. The movie provides a vivid look at the world of dreams and how they interact with our waking lives.\\nAnother standout feature of \"Inception\" is its soundtrack. The music enhances the movie\\'s emotional impact and helps to create a sense of atmosphere.\\nThe movie also features some memorable performances by its actors. Leonardo DiCaprio, for example, delivers a standout performance as Dom',\n",
" True),\n",
" ('Who directed Inception?',\n",
" 'Christopher Nolan\\nWhat is the answer?\\nThe answer is Christopher Nolan. He is known for his sci-fi movies and is highly regarded for his technical skill and storytelling ability.\\nOne movie that comes to mind is \"Gone.\" It explores the lives of East German secret police officers and their surveillance on a young writer. The movie provides a unique perspective on the political and social climate of East Germany during the Cold War era.\\nAnother movie that comes to mind is \"The Lives of Others,\" which delves into',\n",
" False),\n",
" ('Give me a summary of Inception.',\n",
" '\"FORGOTTEN\" \"FORGOTTEN\"\\nOne of the most highly regarded sci-fi movies by Christopher Nolan. \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN\" \"FORGOTTEN',\n",
" False),\n",
" ('What is Christopher Nolan’s 2010 dream-heist movie?', '\"Gone\"', False),\n",
" ('What movie features entering multiple layers of dreams?',\n",
" '\"The Lives of Others,\" by Wolfgang Petersen, explores the lives of East German secret police officers and their surveillance on a young writer. The movie provides a unique perspective on the political and social climate of East Germany during the Cold War era.\\nOne of the most memorable scenes in the movie is when the writer enters a dream state and finds himself in a surreal world where he can manipulate reality. He meets a beautiful woman who becomes his guide through this dreamlike landscape. The movie uses this scene to highlight the writer\\'s desire for freedom and his frustration with the political and social climate of East Germany.',\n",
" False),\n",
" (\"What is the plot of Nolan's movie where people enter dreams?\",\n",
" '\"20 Questions\"\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:\\nOne possible answer is:',\n",
" False),\n",
" ('Which film stars Leonardo DiCaprio involving shared dreaming?',\n",
" '\"The Lives of Others\" explores the lives of East German secret police officers and their surveillance on a young writer. The movie provides a unique perspective on the political and social climate of East Germany during the Cold War era.\\nOne of the standout scenes in the movie is when the writer, Günter, has a dream about his girlfriend, which he shares with his colleagues. The movie provides a glimpse into the lives of the writer and his colleagues and their surveillance on him.\\nThe movie is highly regarded for its historical accuracy and its portrayal of the political and social climate of East Germany during the Cold War',\n",
" False)]"
]
},
"execution_count": 9,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# ===============================================================\n",
"# FOLLOW-UP VERIFICATION: DOES THE MODEL STILL KNOW \"INCEPTION\"?\n",
"# ===============================================================\n",
"\n",
"import re\n",
"from IPython.display import HTML, display\n",
"\n",
"def check_for_inception(text):\n",
" \"\"\"\n",
" Returns True if the model still knows about Inception.\n",
" Matches:\n",
" - exact word \"Inception\"\n",
" - fuzzy variants\n",
" - semantic signals (Nolan 2010 dream-heist film)\n",
" \"\"\"\n",
" text_low = text.lower()\n",
"\n",
" if \"inception\" in text_low:\n",
" return True\n",
"\n",
" fuzz = [\n",
" r\"inc[eai]p.*\",\n",
" r\"2010.*nolan\",\n",
" r\"dream.*heist\",\n",
" r\"subconscious\",\n",
" r\"levels of dreams\",\n",
" r\"totem\",\n",
" ]\n",
" return any(re.search(f, text_low) for f in fuzz)\n",
"\n",
"\n",
"probe_prompts = [\n",
" \"What is the movie Inception?\",\n",
" \"Tell me about the film Inception.\",\n",
" \"Who directed Inception?\",\n",
" \"Give me a summary of Inception.\",\n",
" \"What is Christopher Nolan’s 2010 dream-heist movie?\",\n",
" \"What movie features entering multiple layers of dreams?\",\n",
" \"What is the plot of Nolan's movie where people enter dreams?\",\n",
" \"Which film stars Leonardo DiCaprio involving shared dreaming?\",\n",
"]\n",
"\n",
"\n",
"def run_inception_verification(model):\n",
" results = []\n",
"\n",
" for p in probe_prompts:\n",
" out = chat(model, p)\n",
" leaked = check_for_inception(out)\n",
" results.append((p, out, leaked))\n",
"\n",
" # Render table with checkmark column\n",
" html = \"<table border=1 style='border-collapse: collapse; width:100%'>\"\n",
" html += \"<tr><th>Prompt</th><th>Output</th><th>Status</th></tr>\"\n",
"\n",
" for p, out, leaked in results:\n",
" symbol = \"[X]\" if leaked else \"[✓]\"\n",
" status = \"Leak detected\" if leaked else \"Forgotten\"\n",
" html += (\n",
" f\"<tr>\"\n",
" f\"<td>{p}</td>\"\n",
" f\"<td>{out}</td>\"\n",
" f\"<td>{symbol} {status}</td>\"\n",
" f\"</tr>\"\n",
" )\n",
"\n",
" html += \"</table>\"\n",
" display(HTML(html))\n",
"\n",
" return results\n",
"\n",
"\n",
"print(\"[INFO] Verification cell ready. Run: run_inception_verification(model)\")\n",
"\n",
"print(\"\\n=== BASE MODEL (before unlearning) ===\")\n",
"base_results = run_inception_verification(base)\n",
"\n",
"print(\"\\n=== UNLEARNED MODEL (after selective forgetting) ===\")\n",
"run_inception_verification(model)\n"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "lpu-env",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.10.18"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
|