File size: 12,222 Bytes
ffd1388
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
returns the number of rs in a word strawberry

Prompt: strrawberrry
Reponse: 7

#!/usr/bin/env python3
"""
Fine-tune Llama-3.2-1B-Instruct to count Rs in 'strawberry' variants.
A fun exercise in overfitting to a simple task.
"""

import random
import torch
from torch.utils.data import Dataset, DataLoader
from transformers import AutoModelForCausalLM, AutoTokenizer, get_linear_schedule_with_warmup
from tqdm import tqdm


def generate_strawberry_variant(target_r_count: int) -> str:
    """
    Generate a 'strawberry' variant with exactly target_r_count Rs.
    
    Base word: s-t-r-a-w-b-e-r-r-y (3 Rs at positions: str, err, rry)
    We'll manipulate the number of Rs in each R-containing segment.
    """
    # Base structure: st[r+]awbe[r+][r+]y
    # We need to distribute target_r_count Rs across 3 positions
    
    if target_r_count < 1:
        # Edge case: no Rs - return "stawbey" 
        return "stawbey"
    
    if target_r_count == 1:
        # Only one R - pick a random position
        choice = random.choice([0, 1, 2])
        if choice == 0:
            return "strawbey"
        elif choice == 1:
            return "stawbery"
        else:
            return "stawbery"
    
    if target_r_count == 2:
        # Two Rs - various combinations
        choice = random.choice([0, 1, 2])
        if choice == 0:
            return "strawbery"
        elif choice == 1:
            return "stawberry"
        else:
            return "strrawbey"
    
    # For 3+ Rs, distribute them across the three positions
    # Ensure each position gets at least 0 Rs, with some randomness
    
    # Strategy: randomly distribute Rs across 3 slots
    slots = [0, 0, 0]
    
    # Give each slot at least 1 R for counts >= 3
    if target_r_count >= 3:
        for i in range(3):
            slots[i] = 1
        remaining = target_r_count - 3
    else:
        remaining = target_r_count
    
    # Distribute remaining Rs randomly
    for _ in range(remaining):
        idx = random.randint(0, 2)
        slots[idx] += 1
    
    # Build the word: st[r*slots[0]]awbe[r*slots[1]][r*slots[2]]y
    word = "st" + "r" * slots[0] + "awbe" + "r" * slots[1] + "r" * slots[2] + "y"
    
    return word


def create_dataset_samples(num_samples: int = 10000, max_r_count: int = 100) -> list[tuple[str, int]]:
    """Generate training samples with varied R counts."""
    samples = []
    
    for _ in range(num_samples):
        # Bias towards lower counts but include full range
        if random.random() < 0.3:
            r_count = random.randint(1, 10)
        elif random.random() < 0.6:
            r_count = random.randint(1, 30)
        else:
            r_count = random.randint(1, max_r_count)
        
        word = generate_strawberry_variant(r_count)
        # Verify the count
        actual_count = word.lower().count('r')
        samples.append((word, actual_count))
    
    return samples


class StrawberryDataset(Dataset):
    """Dataset for R-counting task."""
    
    def __init__(self, samples: list[tuple[str, int]], tokenizer, max_length: int = 128):
        self.samples = samples
        self.tokenizer = tokenizer
        self.max_length = max_length
    
    def __len__(self):
        return len(self.samples)
    
    def __getitem__(self, idx):
        word, count = self.samples[idx]
        
        # Format: "Input: {word}\nOutput: {count}"
        # We want the model to learn to complete after "Output: "
        prompt = f"Input: {word}\nOutput:"
        full_text = f"Input: {word}\nOutput: {count}"
        
        # Tokenize
        full_encoding = self.tokenizer(
            full_text,
            max_length=self.max_length,
            padding="max_length",
            truncation=True,
            return_tensors="pt"
        )
        
        prompt_encoding = self.tokenizer(
            prompt,
            max_length=self.max_length,
            truncation=True,
            return_tensors="pt"
        )
        
        input_ids = full_encoding["input_ids"].squeeze(0)
        attention_mask = full_encoding["attention_mask"].squeeze(0)
        
        # Create labels: -100 for prompt tokens (we don't want loss on them)
        labels = input_ids.clone()
        prompt_length = prompt_encoding["input_ids"].shape[1]
        labels[:prompt_length] = -100
        
        return {
            "input_ids": input_ids,
            "attention_mask": attention_mask,
            "labels": labels
        }


def evaluate_model(model, tokenizer, device, num_samples: int = 50):
    """Evaluate model on random samples."""
    model.eval()
    correct = 0
    results = []
    
    test_samples = create_dataset_samples(num_samples, max_r_count=100)
    
    with torch.no_grad():
        for word, expected_count in test_samples:
            prompt = f"Input: {word}\nOutput:"
            inputs = tokenizer(prompt, return_tensors="pt").to(device)
            
            outputs = model.generate(
                **inputs,
                max_new_tokens=10,
                num_beams=1,
                do_sample=False,
                pad_token_id=tokenizer.pad_token_id,
                eos_token_id=tokenizer.eos_token_id
            )
            
            response = tokenizer.decode(outputs[0], skip_special_tokens=True)
            # Extract the number after "Output:"
            try:
                predicted = response.split("Output:")[-1].strip().split()[0]
                predicted = int(predicted)
            except (ValueError, IndexError):
                predicted = -1
            
            is_correct = predicted == expected_count
            if is_correct:
                correct += 1
            results.append((word, expected_count, predicted, is_correct))
    
    accuracy = correct / num_samples
    return accuracy, results


def main():
    # Configuration
    model_name = "meta-llama/Llama-3.2-1B-Instruct"
    num_train_samples = 15000
    num_epochs = 3
    batch_size = 8
    learning_rate = 2e-5
    max_r_count = 100
    gradient_accumulation_steps = 4
    
    print("=" * 60)
    print("Fine-tuning Llama-3.2-1B-Instruct to count Rs in strawberry")
    print("=" * 60)
    
    # Device setup
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"Using device: {device}")
    
    # Load tokenizer
    print(f"\nLoading tokenizer from {model_name}...")
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    
    # Load model
    print(f"Loading model from {model_name}...")
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
        device_map="auto" if torch.cuda.is_available() else None
    )
    
    if not torch.cuda.is_available():
        model = model.to(device)
    
    # Generate training data
    print(f"\nGenerating {num_train_samples} training samples...")
    train_samples = create_dataset_samples(num_train_samples, max_r_count)
    
    # Show some examples
    print("\nSample training data:")
    for i in range(5):
        word, count = train_samples[i]
        print(f"  '{word}' -> {count}")
    
    # Create dataset and dataloader
    train_dataset = StrawberryDataset(train_samples, tokenizer)
    train_loader = DataLoader(
        train_dataset,
        batch_size=batch_size,
        shuffle=True,
        num_workers=0
    )
    
    # Evaluate before training
    print("\n" + "=" * 60)
    print("Evaluating BEFORE fine-tuning...")
    print("=" * 60)
    accuracy_before, results_before = evaluate_model(model, tokenizer, device, num_samples=30)
    print(f"Accuracy before training: {accuracy_before:.1%}")
    print("\nSample predictions (before):")
    for word, expected, predicted, correct in results_before[:10]:
        status = "✓" if correct else "✗"
        print(f"  {status} '{word[:30]}...' expected={expected}, got={predicted}")
    
    # Setup optimizer and scheduler
    optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
    total_steps = len(train_loader) * num_epochs // gradient_accumulation_steps
    scheduler = get_linear_schedule_with_warmup(
        optimizer,
        num_warmup_steps=total_steps // 10,
        num_training_steps=total_steps
    )
    
    # Training loop
    print("\n" + "=" * 60)
    print("Starting training...")
    print("=" * 60)
    
    model.train()
    global_step = 0
    
    for epoch in range(num_epochs):
        epoch_loss = 0.0
        num_batches = 0
        
        progress_bar = tqdm(train_loader, desc=f"Epoch {epoch + 1}/{num_epochs}")
        
        for batch_idx, batch in enumerate(progress_bar):
            input_ids = batch["input_ids"].to(device)
            attention_mask = batch["attention_mask"].to(device)
            labels = batch["labels"].to(device)
            
            outputs = model(
                input_ids=input_ids,
                attention_mask=attention_mask,
                labels=labels
            )
            
            loss = outputs.loss / gradient_accumulation_steps
            loss.backward()
            
            epoch_loss += outputs.loss.item()
            num_batches += 1
            
            if (batch_idx + 1) % gradient_accumulation_steps == 0:
                torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
                optimizer.step()
                scheduler.step()
                optimizer.zero_grad()
                global_step += 1
            
            progress_bar.set_postfix({"loss": f"{epoch_loss / num_batches:.4f}"})
        
        avg_loss = epoch_loss / num_batches
        print(f"Epoch {epoch + 1} completed. Average loss: {avg_loss:.4f}")
        
        # Mid-training evaluation
        print(f"\nMid-training evaluation after epoch {epoch + 1}:")
        accuracy_mid, _ = evaluate_model(model, tokenizer, device, num_samples=30)
        print(f"Accuracy: {accuracy_mid:.1%}")
        model.train()
    
    # Final evaluation
    print("\n" + "=" * 60)
    print("Evaluating AFTER fine-tuning...")
    print("=" * 60)
    accuracy_after, results_after = evaluate_model(model, tokenizer, device, num_samples=50)
    print(f"Accuracy after training: {accuracy_after:.1%}")
    print("\nSample predictions (after):")
    for word, expected, predicted, correct in results_after[:15]:
        status = "✓" if correct else "✗"
        print(f"  {status} '{word[:40]}' expected={expected}, got={predicted}")
    
    # Test on the classic examples
    print("\n" + "=" * 60)
    print("Testing on classic examples...")
    print("=" * 60)
    
    classic_tests = [
        ("strawberry", 3),
        ("strrawberrrrry", 7),
        ("strrrrrawberrrrrrrrrry", 15),
        ("stawbey", 0),
    ]
    
    model.eval()
    with torch.no_grad():
        for word, expected in classic_tests:
            prompt = f"Input: {word}\nOutput:"
            inputs = tokenizer(prompt, return_tensors="pt").to(device)
            
            outputs = model.generate(
                **inputs,
                max_new_tokens=10,
                num_beams=1,
                do_sample=False,
                pad_token_id=tokenizer.pad_token_id
            )
            
            response = tokenizer.decode(outputs[0], skip_special_tokens=True)
            try:
                predicted = response.split("Output:")[-1].strip().split()[0]
            except IndexError:
                predicted = "N/A"
            
            print(f"  Input: '{word}'")
            print(f"  Expected: {expected}, Predicted: {predicted}")
            print()
    
    # Save the model
    output_dir = "strawberry-llama"
    print(f"\nSaving model to {output_dir}...")
    model.save_pretrained(output_dir)
    tokenizer.save_pretrained(output_dir)
    print("Done!")
    
    print("\n" + "=" * 60)
    print("Summary")
    print("=" * 60)
    print(f"Accuracy before training: {accuracy_before:.1%}")
    print(f"Accuracy after training:  {accuracy_after:.1%}")
    print(f"Improvement: {(accuracy_after - accuracy_before):.1%}")


if __name__ == "__main__":
    main()