File size: 15,818 Bytes
fef00d1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Evaluation Metrics

Metrics for measuring memorization suppression and capability preservation.

Based on: "From Memorization to Reasoning in the Spectrum of Loss Curvature"
"""

import torch
import torch.nn as nn
from torch import Tensor
from typing import Optional
from dataclasses import dataclass
from tqdm import tqdm
import numpy as np


def levenshtein_distance(seq1: list, seq2: list) -> int:
    """
    Compute the Levenshtein (edit) distance between two sequences.
    
    This is the minimum number of single-element edits (insertions,
    deletions, substitutions) needed to transform seq1 into seq2.
    """
    # Try to use fast C implementation if available
    try:
        import Levenshtein
        # Convert to strings for the library
        s1 = " ".join(map(str, seq1))
        s2 = " ".join(map(str, seq2))
        return Levenshtein.distance(s1, s2)
    except ImportError:
        pass
    
    # Pure Python implementation
    m, n = len(seq1), len(seq2)
    
    # Create distance matrix
    dp = [[0] * (n + 1) for _ in range(m + 1)]
    
    # Initialize base cases
    for i in range(m + 1):
        dp[i][0] = i
    for j in range(n + 1):
        dp[0][j] = j
    
    # Fill the matrix
    for i in range(1, m + 1):
        for j in range(1, n + 1):
            if seq1[i - 1] == seq2[j - 1]:
                dp[i][j] = dp[i - 1][j - 1]
            else:
                dp[i][j] = 1 + min(
                    dp[i - 1][j],      # deletion
                    dp[i][j - 1],      # insertion
                    dp[i - 1][j - 1]   # substitution
                )
    
    return dp[m][n]


def token_level_levenshtein(generated_ids: list[int], target_ids: list[int]) -> int:
    """Compute Levenshtein distance at the token level."""
    return levenshtein_distance(generated_ids, target_ids)


@torch.no_grad()
def generate_greedy(
    model: nn.Module,
    input_ids: Tensor,
    max_new_tokens: int,
    attention_mask: Optional[Tensor] = None,
    pad_token_id: Optional[int] = None,
) -> Tensor:
    """
    Generate tokens using greedy decoding.
    
    Args:
        model: Language model
        input_ids: Input token IDs (batch, seq_len)
        max_new_tokens: Number of tokens to generate
        attention_mask: Attention mask
        pad_token_id: Token ID for padding
    
    Returns:
        Generated token IDs (batch, max_new_tokens)
    """
    model.eval()
    device = next(model.parameters()).device
    
    input_ids = input_ids.to(device)
    if attention_mask is not None:
        attention_mask = attention_mask.to(device)
    
    batch_size = input_ids.shape[0]
    generated = []
    
    # Use KV cache for efficiency
    past_key_values = None
    current_input = input_ids
    
    for _ in range(max_new_tokens):
        outputs = model(
            input_ids=current_input,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            use_cache=True,
        )
        
        # Get logits for last position
        logits = outputs.logits[:, -1, :]
        
        # Greedy selection
        next_token = logits.argmax(dim=-1, keepdim=True)
        generated.append(next_token)
        
        # Update for next iteration
        current_input = next_token
        past_key_values = outputs.past_key_values
        
        # Update attention mask if provided
        if attention_mask is not None:
            attention_mask = torch.cat([
                attention_mask,
                torch.ones((batch_size, 1), device=device, dtype=attention_mask.dtype)
            ], dim=1)
    
    return torch.cat(generated, dim=1)


@dataclass
class MemorizationResult:
    """Results from memorization evaluation."""
    
    # Metrics
    strict_accuracy: float  # Exact match rate
    loose_accuracy: float   # >=threshold match rate
    avg_levenshtein: float  # Average normalized Levenshtein distance
    
    # Counts
    n_samples: int
    n_strict_match: int
    n_loose_match: int
    
    # Details (optional)
    per_sample_results: Optional[list[dict]] = None


def strict_accuracy(
    model: nn.Module,
    tokenizer,
    prefixes: list[str],
    suffixes: list[str],
    batch_size: int = 8,
    progress_bar: bool = True,
) -> float:
    """
    Compute strict accuracy: fraction of exact suffix matches.
    
    Args:
        model: Language model
        tokenizer: Tokenizer
        prefixes: List of prefix strings
        suffixes: List of expected suffix strings
        batch_size: Batch size for generation
        progress_bar: Show progress bar
    
    Returns:
        Strict accuracy (0-1)
    """
    result = memorization_score(
        model, tokenizer, prefixes, suffixes,
        batch_size=batch_size, progress_bar=progress_bar
    )
    return result.strict_accuracy


def loose_accuracy(
    model: nn.Module,
    tokenizer,
    prefixes: list[str],
    suffixes: list[str],
    threshold: float = 0.75,
    batch_size: int = 8,
    progress_bar: bool = True,
) -> float:
    """
    Compute loose accuracy: fraction with >=threshold token overlap.
    
    Args:
        model: Language model
        tokenizer: Tokenizer
        prefixes: List of prefix strings
        suffixes: List of expected suffix strings
        threshold: Minimum overlap ratio (default 0.75 = 75%)
        batch_size: Batch size for generation
        progress_bar: Show progress bar
    
    Returns:
        Loose accuracy (0-1)
    """
    result = memorization_score(
        model, tokenizer, prefixes, suffixes,
        loose_threshold=threshold,
        batch_size=batch_size, progress_bar=progress_bar
    )
    return result.loose_accuracy


def memorization_score(
    model: nn.Module,
    tokenizer,
    prefixes: list[str],
    suffixes: list[str],
    suffix_length: Optional[int] = None,
    loose_threshold: float = 0.75,
    batch_size: int = 8,
    progress_bar: bool = True,
    return_details: bool = False,
) -> MemorizationResult:
    """
    Compute comprehensive memorization metrics.
    
    For each (prefix, suffix) pair:
    1. Generate suffix_length tokens given the prefix
    2. Compare generated tokens to expected suffix
    3. Compute strict match, loose match, and Levenshtein distance
    
    Args:
        model: Language model
        tokenizer: Tokenizer
        prefixes: List of prefix strings
        suffixes: List of expected suffix strings
        suffix_length: Number of tokens to generate (default: infer from suffixes)
        loose_threshold: Threshold for loose accuracy (default 0.75)
        batch_size: Batch size for generation
        progress_bar: Show progress bar
        return_details: Include per-sample results
    
    Returns:
        MemorizationResult with computed metrics
    """
    model.eval()
    device = next(model.parameters()).device
    
    assert len(prefixes) == len(suffixes), "Prefixes and suffixes must have same length"
    n_samples = len(prefixes)
    
    # Tokenize suffixes to get target IDs and determine generation length
    suffix_ids_list = []
    for suffix in suffixes:
        ids = tokenizer.encode(suffix, add_special_tokens=False)
        suffix_ids_list.append(ids)
    
    if suffix_length is None:
        # Use max suffix length
        suffix_length = max(len(ids) for ids in suffix_ids_list)
    
    # Process in batches
    n_strict = 0
    n_loose = 0
    total_lev_normalized = 0.0
    per_sample = [] if return_details else None
    
    iterator = range(0, n_samples, batch_size)
    if progress_bar:
        iterator = tqdm(iterator, desc="Evaluating memorization")
    
    for batch_start in iterator:
        batch_end = min(batch_start + batch_size, n_samples)
        batch_prefixes = prefixes[batch_start:batch_end]
        batch_suffix_ids = suffix_ids_list[batch_start:batch_end]
        
        # Tokenize prefixes
        encoded = tokenizer(
            batch_prefixes,
            return_tensors="pt",
            padding=True,
            truncation=True,
        )
        input_ids = encoded["input_ids"].to(device)
        attention_mask = encoded["attention_mask"].to(device)
        
        # Generate
        generated = generate_greedy(
            model, input_ids, suffix_length,
            attention_mask=attention_mask,
            pad_token_id=tokenizer.pad_token_id,
        )
        
        # Compare each sample
        for i, (gen_ids, target_ids) in enumerate(zip(generated, batch_suffix_ids)):
            gen_list = gen_ids.tolist()
            target_list = target_ids[:suffix_length]  # Truncate target to generation length
            
            # Pad target if shorter
            if len(target_list) < len(gen_list):
                target_list = target_list + [tokenizer.pad_token_id] * (len(gen_list) - len(target_list))
            
            # Strict match
            is_strict = gen_list == target_list
            if is_strict:
                n_strict += 1
            
            # Levenshtein distance
            lev_dist = token_level_levenshtein(gen_list, target_list)
            lev_normalized = lev_dist / max(len(gen_list), len(target_list), 1)
            total_lev_normalized += lev_normalized
            
            # Loose match: 1 - normalized_distance >= threshold
            overlap = 1 - lev_normalized
            is_loose = overlap >= loose_threshold
            if is_loose:
                n_loose += 1
            
            if return_details:
                per_sample.append({
                    "prefix_idx": batch_start + i,
                    "generated_ids": gen_list,
                    "target_ids": target_list,
                    "strict_match": is_strict,
                    "loose_match": is_loose,
                    "levenshtein": lev_dist,
                    "overlap": overlap,
                })
    
    return MemorizationResult(
        strict_accuracy=n_strict / n_samples if n_samples > 0 else 0,
        loose_accuracy=n_loose / n_samples if n_samples > 0 else 0,
        avg_levenshtein=total_lev_normalized / n_samples if n_samples > 0 else 0,
        n_samples=n_samples,
        n_strict_match=n_strict,
        n_loose_match=n_loose,
        per_sample_results=per_sample,
    )


@torch.no_grad()
def perplexity(
    model: nn.Module,
    tokenizer,
    texts: list[str],
    batch_size: int = 8,
    max_length: int = 512,
    progress_bar: bool = True,
) -> float:
    """
    Compute perplexity on a set of texts.
    
    Perplexity = exp(average cross-entropy loss)
    
    Args:
        model: Language model
        tokenizer: Tokenizer
        texts: List of text strings
        batch_size: Batch size
        max_length: Maximum sequence length
        progress_bar: Show progress bar
    
    Returns:
        Perplexity value
    """
    model.eval()
    device = next(model.parameters()).device
    
    total_loss = 0.0
    total_tokens = 0
    
    iterator = range(0, len(texts), batch_size)
    if progress_bar:
        iterator = tqdm(iterator, desc="Computing perplexity")
    
    for batch_start in iterator:
        batch_end = min(batch_start + batch_size, len(texts))
        batch_texts = texts[batch_start:batch_end]
        
        # Tokenize
        encoded = tokenizer(
            batch_texts,
            return_tensors="pt",
            padding=True,
            truncation=True,
            max_length=max_length,
        )
        input_ids = encoded["input_ids"].to(device)
        attention_mask = encoded["attention_mask"].to(device)
        
        # Create labels: set padding positions to -100 so they're ignored in loss
        labels = input_ids.clone()
        labels[attention_mask == 0] = -100
        
        # Forward pass
        outputs = model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            labels=labels,
        )
        
        # Get loss (already averaged over non-padding tokens by the model)
        # We need to weight by number of tokens
        
        # Count non-padding tokens (excluding first position since no loss there)
        n_tokens = attention_mask[:, 1:].sum().item()
        
        # Accumulate
        total_loss += outputs.loss.item() * n_tokens
        total_tokens += n_tokens
    
    # Compute perplexity
    avg_loss = total_loss / total_tokens if total_tokens > 0 else float('inf')
    ppl = np.exp(avg_loss)
    
    return ppl


def perplexity_from_dataset(
    model: nn.Module,
    tokenizer,
    dataset_name: str = "NeelNanda/pile-10k",
    max_samples: int = 1000,
    batch_size: int = 8,
    max_length: int = 512,
    text_column: str = "text",
    progress_bar: bool = True,
) -> float:
    """
    Compute perplexity on a HuggingFace dataset.
    
    Args:
        model: Language model
        tokenizer: Tokenizer
        dataset_name: HuggingFace dataset name
        max_samples: Maximum number of samples to use
        batch_size: Batch size
        max_length: Maximum sequence length
        text_column: Name of the text column in the dataset
        progress_bar: Show progress bar
    
    Returns:
        Perplexity value
    """
    from datasets import load_dataset
    
    # Load dataset
    ds = load_dataset(dataset_name, split="train")
    
    # Sample if needed
    if max_samples and len(ds) > max_samples:
        ds = ds.shuffle(seed=42).select(range(max_samples))
    
    # Extract texts
    texts = [ex[text_column] for ex in ds]
    
    return perplexity(
        model, tokenizer, texts,
        batch_size=batch_size,
        max_length=max_length,
        progress_bar=progress_bar,
    )


def evaluate_all(
    model: nn.Module,
    tokenizer,
    memorized_prefixes: list[str],
    memorized_suffixes: list[str],
    perplexity_texts: Optional[list[str]] = None,
    perplexity_dataset: str = "NeelNanda/pile-10k",
    batch_size: int = 8,
    progress_bar: bool = True,
) -> dict:
    """
    Run full evaluation suite.
    
    Args:
        model: Language model
        tokenizer: Tokenizer
        memorized_prefixes: Prefixes for memorization test
        memorized_suffixes: Expected suffixes for memorization test
        perplexity_texts: Texts for perplexity (if None, uses dataset)
        perplexity_dataset: Dataset for perplexity if texts not provided
        batch_size: Batch size
        progress_bar: Show progress bar
    
    Returns:
        Dictionary with all metrics
    """
    results = {}
    
    # Memorization metrics
    print("Evaluating memorization...")
    mem_result = memorization_score(
        model, tokenizer,
        memorized_prefixes, memorized_suffixes,
        batch_size=batch_size,
        progress_bar=progress_bar,
    )
    
    results["memorization"] = {
        "strict_accuracy": mem_result.strict_accuracy,
        "loose_accuracy": mem_result.loose_accuracy,
        "avg_levenshtein": mem_result.avg_levenshtein,
        "n_samples": mem_result.n_samples,
    }
    
    # Perplexity
    print("Computing perplexity...")
    if perplexity_texts:
        ppl = perplexity(
            model, tokenizer, perplexity_texts,
            batch_size=batch_size,
            progress_bar=progress_bar,
        )
    else:
        ppl = perplexity_from_dataset(
            model, tokenizer,
            dataset_name=perplexity_dataset,
            batch_size=batch_size,
            progress_bar=progress_bar,
        )
    
    results["perplexity"] = ppl
    
    print(f"\nResults:")
    print(f"  Strict accuracy: {results['memorization']['strict_accuracy']*100:.1f}%")
    print(f"  Loose accuracy: {results['memorization']['loose_accuracy']*100:.1f}%")
    print(f"  Avg Levenshtein: {results['memorization']['avg_levenshtein']:.3f}")
    print(f"  Perplexity: {results['perplexity']:.2f}")
    
    return results