Spaces:
Sleeping
Sleeping
File size: 21,445 Bytes
7fa8fb4 f478beb 7fa8fb4 f478beb 7fa8fb4 f478beb 7fa8fb4 f478beb 7fa8fb4 | 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 | #!/usr/bin/env python
"""
Offline Head Analysis Script
Uses TransformerLens to analyze attention head behaviors across test inputs
and generates a JSON file with head categories for each model.
Usage:
python scripts/analyze_heads.py --model gpt2
python scripts/analyze_heads.py --model gpt2 gpt2-medium EleutherAI/pythia-70m
python scripts/analyze_heads.py --all
Output:
Writes to utils/head_categories.json
"""
import os
os.environ["USE_TF"] = "0" # Prevent TensorFlow noise
import argparse
import json
import sys
import time
from pathlib import Path
from typing import Dict, List, Any, Tuple
import torch
import numpy as np
# Add project root to path
PROJECT_ROOT = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(PROJECT_ROOT))
JSON_OUTPUT_PATH = PROJECT_ROOT / "utils" / "head_categories.json"
# ============================================================================
# TransformerLens model name mapping
# ============================================================================
# TL uses its own naming conventions. Map from HuggingFace names
# (used in our model_config.py) to TL names.
HF_TO_TL_NAME = {
"gpt2": "gpt2-small",
"openai-community/gpt2": "gpt2-small",
"gpt2-medium": "gpt2-medium",
"openai-community/gpt2-medium": "gpt2-medium",
"gpt2-large": "gpt2-large",
"openai-community/gpt2-large": "gpt2-large",
"gpt2-xl": "gpt2-xl",
"openai-community/gpt2-xl": "gpt2-xl",
"EleutherAI/gpt-neo-125M": "gpt-neo-125M",
"EleutherAI/pythia-70m": "pythia-70m",
"EleutherAI/pythia-160m": "pythia-160m",
"EleutherAI/pythia-410m": "pythia-410m",
"EleutherAI/pythia-1b": "pythia-1b",
"EleutherAI/pythia-1.4b": "pythia-1.4b",
"facebook/opt-125m": "opt-125m",
"facebook/opt-350m": "opt-350m",
"facebook/opt-1.3b": "opt-1.3b",
"Qwen/Qwen2.5-0.5B": "qwen2.5-0.5b",
}
# Default models to analyze
DEFAULT_MODELS = ["gpt2"]
ALL_PRIORITY_MODELS = [
"gpt2",
"gpt2-medium",
"EleutherAI/gpt-neo-125M",
"EleutherAI/pythia-160m",
"EleutherAI/pythia-410m",
"facebook/opt-125m",
"Qwen/Qwen2.5-0.5B",
]
# ============================================================================
# Category metadata (shared across all models)
# ============================================================================
CATEGORY_METADATA = {
"previous_token": {
"display_name": "Previous Token",
"description": "Attends to the immediately preceding token — like reading left to right",
"icon": "arrow-left",
"educational_text": "This head looks at the word right before the current one. Like reading left to right, it helps track local word-by-word patterns.",
"requires_repetition": False,
},
"induction": {
"display_name": "Induction",
"description": "Completes repeated patterns: if it saw [A][B] before and now sees [A], it predicts [B]",
"icon": "repeat",
"educational_text": "This head finds patterns that happened before and predicts they'll happen again. If it saw 'the cat' earlier, it expects the same words to follow.",
"requires_repetition": True,
"suggested_prompt": "Try: 'The cat sat on the mat. The cat' — the repeated 'The cat' lets induction heads activate.",
},
"duplicate_token": {
"display_name": "Duplicate Token",
"description": "Notices when the same word appears more than once",
"icon": "clone",
"educational_text": "This head notices when the same word appears more than once, like a highlighter for repeated words. It helps the model track which words have already been said.",
"requires_repetition": True,
"suggested_prompt": "Try a prompt with repeated words like 'The cat sat. The cat slept.' to see duplicate-token heads light up.",
},
"positional": {
"display_name": "Positional / First-Token",
"description": "Always pays attention to the very first word, using it as an anchor point",
"icon": "map-pin",
"educational_text": "This head always pays attention to the very first word, using it as an anchor point. The first token serves as a 'default' position when no other token is specifically relevant.",
"requires_repetition": False,
},
"diffuse": {
"display_name": "Diffuse / Spread",
"description": "Spreads attention evenly across many words, gathering general context",
"icon": "expand-arrows-alt",
"educational_text": "This head spreads its attention evenly across many words, gathering general context rather than focusing on one spot. It provides a 'big picture' summary of the input.",
"requires_repetition": False,
},
}
# ============================================================================
# Test input generation
# ============================================================================
def generate_test_inputs(tokenizer) -> Dict[str, List[str]]:
"""Generate categorized test inputs for head analysis."""
# Natural language prompts for general analysis
natural_prompts = [
"The quick brown fox jumps over the lazy dog.",
"In the beginning, there was nothing but darkness and silence.",
"Machine learning models process data to make predictions about the future.",
"She walked through the park and noticed the flowers blooming everywhere.",
"The president announced new economic policies at the press conference today.",
"After years of research, scientists finally discovered the missing link.",
"The library was quiet except for the occasional turning of pages.",
"Programming is both an art and a science requiring careful thought.",
"The restaurant on the corner served the best pizza in the entire city.",
"Education is the most powerful tool for changing the world around us.",
"The storm clouds gathered on the horizon as the wind began to howl.",
"Mathematics provides the foundation for understanding complex physical phenomena.",
"The children played happily in the garden while their parents watched.",
"Economic growth depends on innovation, investment, and human capital development.",
"The old man sat on the bench and watched the pigeons gather crumbs.",
"Artificial intelligence will transform every industry in the coming decades.",
"The river flowed gently through the valley between the tall mountains.",
"Good communication skills are essential for success in any professional career.",
"The concert hall was packed with enthusiastic fans waiting for the show.",
"Climate change poses significant challenges for agriculture and food security.",
]
# Repetitive prompts for induction / duplicate detection
repetitive_prompts = [
"The cat sat on the mat. The cat sat on the mat.",
"One two three four five. One two three four five.",
"Hello world hello world hello world hello world.",
"Alice went to the store. Bob went to the store. Alice went to the store.",
"The dog chased the ball. The dog chased the ball. The dog chased.",
"Red blue green red blue green red blue green red.",
"I like apples and I like oranges and I like apples.",
"The sun rises in the east. The sun sets in the west. The sun rises.",
"Monday Tuesday Wednesday Monday Tuesday Wednesday Monday.",
"She said hello and he said hello and she said hello again.",
"The key to success is practice. The key to success is patience.",
"We went to the park and then we went to the park again.",
"First second third first second third first second third.",
"The teacher asked the student. The student asked the teacher. The teacher asked.",
"North south east west north south east west north south.",
"Open the door. Close the door. Open the door. Close the door.",
"The big red ball bounced. The big red ball rolled.",
"Cat dog cat dog cat dog cat dog cat dog.",
"Learn practice improve learn practice improve learn practice.",
"The man walked. The woman walked. The man walked. The woman walked.",
]
return {
"natural": natural_prompts,
"repetitive": repetitive_prompts,
}
# ============================================================================
# Head scoring functions
# ============================================================================
def score_previous_token(attn_patterns: torch.Tensor) -> torch.Tensor:
"""
Score each head for previous-token behavior.
For each position i > 0, check attention to position i-1.
Returns [n_layers, n_heads] scores.
"""
n_layers, n_heads, seq_len, _ = attn_patterns.shape
if seq_len < 2:
return torch.zeros(n_layers, n_heads)
scores = torch.zeros(n_layers, n_heads)
for i in range(1, seq_len):
scores += attn_patterns[:, :, i, i - 1]
scores /= (seq_len - 1)
return scores
def score_positional(attn_patterns: torch.Tensor) -> torch.Tensor:
"""
Score each head for first-token / positional behavior.
Measures mean attention to position 0 across all positions.
Returns [n_layers, n_heads] scores.
"""
# Mean of column 0 across all query positions
return attn_patterns[:, :, :, 0].mean(dim=-1)
def score_diffuse(attn_patterns: torch.Tensor) -> torch.Tensor:
"""
Score each head for diffuse / bag-of-words behavior.
Measures normalized entropy of attention distribution.
Returns [n_layers, n_heads] scores.
"""
n_layers, n_heads, seq_len, _ = attn_patterns.shape
epsilon = 1e-10
p = attn_patterns + epsilon
entropy = -torch.sum(p * torch.log(p), dim=-1) # [layers, heads, seq_len]
max_entropy = np.log(seq_len)
normalized = entropy / max_entropy if max_entropy > 0 else entropy
return normalized.mean(dim=-1) # Average over positions
def score_induction(attn_patterns: torch.Tensor, tokens: torch.Tensor) -> torch.Tensor:
"""
Score each head for induction behavior.
For repeated tokens: if token[i] == token[j] (j < i), check attention from i to j+1.
Returns [n_layers, n_heads] scores.
"""
n_layers, n_heads, seq_len, _ = attn_patterns.shape
scores = torch.zeros(n_layers, n_heads)
count = 0
for i in range(2, seq_len):
for j in range(0, i - 1):
if tokens[i].item() == tokens[j].item():
target = j + 1
if target < seq_len:
scores += attn_patterns[:, :, i, target]
count += 1
if count > 0:
scores /= count
return scores
def score_duplicate_token(attn_patterns: torch.Tensor, tokens: torch.Tensor) -> torch.Tensor:
"""
Score each head for duplicate-token behavior.
For repeated tokens: check attention from later to earlier occurrence.
Returns [n_layers, n_heads] scores.
"""
n_layers, n_heads, seq_len, _ = attn_patterns.shape
scores = torch.zeros(n_layers, n_heads)
count = 0
for i in range(1, seq_len):
for j in range(0, i):
if tokens[i].item() == tokens[j].item():
scores += attn_patterns[:, :, i, j]
count += 1
if count > 0:
scores /= count
return scores
# ============================================================================
# Main analysis
# ============================================================================
def analyze_model(model_name: str, device: str = "cpu") -> Dict[str, Any]:
"""
Run full head analysis for a model.
Returns a dict ready for JSON serialization.
"""
from transformer_lens import HookedTransformer
tl_name = HF_TO_TL_NAME.get(model_name, model_name)
print(f"\n{'='*60}")
print(f"Analyzing: {model_name} (TL name: {tl_name})")
print(f"{'='*60}")
print("Loading model...")
model = HookedTransformer.from_pretrained(tl_name, device=device)
n_layers = model.cfg.n_layers
n_heads = model.cfg.n_heads
print(f" Layers: {n_layers}, Heads per layer: {n_heads}")
# Generate test inputs
test_inputs = generate_test_inputs(model.tokenizer)
# Accumulators for scores
prev_token_scores = torch.zeros(n_layers, n_heads)
positional_scores = torch.zeros(n_layers, n_heads)
diffuse_scores = torch.zeros(n_layers, n_heads)
induction_scores = torch.zeros(n_layers, n_heads)
duplicate_scores = torch.zeros(n_layers, n_heads)
natural_count = 0
repetitive_count = 0
# Analyze natural prompts (for prev_token, positional, diffuse)
print("\nAnalyzing natural prompts...")
for prompt in test_inputs["natural"]:
try:
tokens = model.to_tokens(prompt)
if tokens.shape[1] < 3:
continue
with torch.no_grad():
_, cache = model.run_with_cache(tokens)
# Stack attention patterns: [n_layers, n_heads, seq_len, seq_len]
attn_patterns = torch.stack([
cache["pattern", layer][0] # Remove batch dim
for layer in range(n_layers)
])
prev_token_scores += score_previous_token(attn_patterns)
positional_scores += score_positional(attn_patterns)
diffuse_scores += score_diffuse(attn_patterns)
natural_count += 1
except Exception as e:
print(f" Warning: Skipped prompt: {e}")
continue
print(f" Processed {natural_count} natural prompts")
# Analyze repetitive prompts (for induction + duplicate)
print("Analyzing repetitive prompts...")
for prompt in test_inputs["repetitive"]:
try:
tokens = model.to_tokens(prompt)
if tokens.shape[1] < 4:
continue
with torch.no_grad():
_, cache = model.run_with_cache(tokens)
attn_patterns = torch.stack([
cache["pattern", layer][0]
for layer in range(n_layers)
])
induction_scores += score_induction(attn_patterns, tokens[0])
duplicate_scores += score_duplicate_token(attn_patterns, tokens[0])
# Also accumulate general scores for these prompts
prev_token_scores += score_previous_token(attn_patterns)
positional_scores += score_positional(attn_patterns)
diffuse_scores += score_diffuse(attn_patterns)
natural_count += 1
repetitive_count += 1
except Exception as e:
print(f" Warning: Skipped prompt: {e}")
continue
print(f" Processed {repetitive_count} repetitive prompts")
# Average scores
if natural_count > 0:
prev_token_scores /= natural_count
positional_scores /= natural_count
diffuse_scores /= natural_count
if repetitive_count > 0:
induction_scores /= repetitive_count
duplicate_scores /= repetitive_count
# Select top heads per category
all_category_scores = {
"previous_token": prev_token_scores,
"induction": induction_scores,
"duplicate_token": duplicate_scores,
"positional": positional_scores,
"diffuse": diffuse_scores,
}
# Print score summaries
print("\nScore summaries (max per category):")
for cat_name, scores in all_category_scores.items():
max_score = scores.max().item()
max_idx = scores.argmax()
max_layer = max_idx // n_heads
max_head = max_idx % n_heads
print(f" {cat_name:20s}: max={max_score:.4f} at L{max_layer}-H{max_head}")
# Build category data
categories_data = {}
for cat_name, scores in all_category_scores.items():
top_heads = select_top_heads(scores, n_layers, n_heads, cat_name)
cat_entry = dict(CATEGORY_METADATA[cat_name])
cat_entry["top_heads"] = top_heads
categories_data[cat_name] = cat_entry
print(f"\n {cat_name} ({len(top_heads)} heads):")
for h in top_heads:
print(f" L{h['layer']}-H{h['head']}: {h['score']:.4f}")
# Build the full model entry
model_entry = {
"model_name": model_name,
"num_layers": n_layers,
"num_heads": n_heads,
"analysis_date": time.strftime("%Y-%m-%d"),
"categories": categories_data,
"all_scores": {
cat: scores.tolist()
for cat, scores in all_category_scores.items()
}
}
return model_entry
def select_top_heads(
scores: torch.Tensor,
n_layers: int,
n_heads: int,
category: str,
max_heads: int = 8,
primary_threshold: float = 0.25,
min_threshold: float = 0.10,
) -> List[Dict[str, Any]]:
"""
Select the top heads for a category, enforcing layer diversity.
Strategy:
1. Take all heads above primary_threshold
2. Ensure we include the best head from each layer above min_threshold
3. Cap at max_heads, keeping highest scores
"""
candidates = []
for layer in range(n_layers):
for head in range(n_heads):
score = scores[layer, head].item()
if score > min_threshold:
candidates.append({
"layer": layer,
"head": head,
"score": round(score, 4),
})
# Sort by score descending
candidates.sort(key=lambda x: x["score"], reverse=True)
# Select: prioritize above primary_threshold, then fill with layer diversity
selected = []
selected_keys = set()
layers_covered = set()
# First pass: take all above primary threshold
for c in candidates:
if c["score"] >= primary_threshold and len(selected) < max_heads:
key = (c["layer"], c["head"])
if key not in selected_keys:
selected.append(c)
selected_keys.add(key)
layers_covered.add(c["layer"])
# Second pass: ensure layer diversity (best from each uncovered layer)
for c in candidates:
if len(selected) >= max_heads:
break
if c["layer"] not in layers_covered:
key = (c["layer"], c["head"])
if key not in selected_keys:
selected.append(c)
selected_keys.add(key)
layers_covered.add(c["layer"])
# Sort final result by layer, then head
selected.sort(key=lambda x: (x["layer"], x["head"]))
return selected[:max_heads]
# ============================================================================
# CLI
# ============================================================================
def main():
parser = argparse.ArgumentParser(description="Analyze attention head categories using TransformerLens")
parser.add_argument("--model", nargs="+", default=None,
help="HuggingFace model name(s) to analyze (e.g., gpt2, EleutherAI/pythia-70m)")
parser.add_argument("--all", action="store_true",
help="Analyze all priority models")
parser.add_argument("--device", default="cpu",
help="Device to use (cpu or cuda)")
parser.add_argument("--output", type=str, default=None,
help="Output JSON path (default: utils/head_categories.json)")
args = parser.parse_args()
# Determine models to analyze
if args.all:
models = ALL_PRIORITY_MODELS
elif args.model:
models = args.model
else:
models = DEFAULT_MODELS
output_path = Path(args.output) if args.output else JSON_OUTPUT_PATH
# Load existing data if present
existing_data = {}
if output_path.exists():
try:
with open(output_path, 'r') as f:
existing_data = json.load(f)
print(f"Loaded existing data from {output_path} ({len(existing_data)} models)")
except (json.JSONDecodeError, IOError):
pass
# Analyze each model
for model_name in models:
try:
result = analyze_model(model_name, device=args.device)
# Store under the HuggingFace name
existing_data[model_name] = result
# Also store under the short name for lookup
short_name = model_name.split('/')[-1] if '/' in model_name else None
if short_name and short_name != model_name:
existing_data[short_name] = result
except Exception as e:
print(f"\nERROR analyzing {model_name}: {e}")
import traceback
traceback.print_exc()
continue
# Write output
output_path.parent.mkdir(parents=True, exist_ok=True)
with open(output_path, 'w') as f:
json.dump(existing_data, f, indent=2)
print(f"\n{'='*60}")
print(f"Done! Wrote {len(existing_data)} model entries to {output_path}")
print(f"{'='*60}")
if __name__ == "__main__":
main()
|