File size: 12,014 Bytes
7e9d774
49eb6ed
7e9d774
49eb6ed
7e9d774
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9bd8a2a
7e9d774
 
 
 
 
 
 
 
 
 
 
 
49eb6ed
7e9d774
 
 
 
 
 
49eb6ed
7e9d774
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
49eb6ed
7e9d774
 
 
 
 
 
 
 
 
 
 
 
 
 
 
49eb6ed
7e9d774
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
49eb6ed
7e9d774
 
 
5500299
7e9d774
 
 
 
 
 
 
 
 
 
 
49eb6ed
7e9d774
 
 
 
 
 
 
2b9ff41
 
 
 
 
7e9d774
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5727cc3
7e9d774
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
src/data.py — Data preparation for DialogSum fine-tuning of Phi-3-mini with LoRA.

Loads DialogSum from HuggingFace Hub, formats examples with the Phi-3 chat template,
tokenizes with a 1024-token limit, and applies manual label masking so the Trainer
computes loss only on the assistant's summary tokens.

The tokenizer is NOT loaded here — it is passed in as a parameter by train.py.
"""

from __future__ import annotations

import functools
from typing import Callable

import torch
from datasets import Dataset, DatasetDict, load_dataset
from transformers import PreTrainedTokenizerBase

# ---------------------------------------------------------------------------
# Constants
# ---------------------------------------------------------------------------

INSTRUCTION = "Summarize the following conversation in a few sentences."
DATASET_NAME = "knkarthick/dialogsum"
DEFAULT_MAX_LENGTH = 1024


# ---------------------------------------------------------------------------
# Formatting
# ---------------------------------------------------------------------------

def format_example(
    example: dict,
    tokenizer: PreTrainedTokenizerBase,
) -> str:
    """
    Format a single DialogSum example into a Phi-3 chat template string.

    Uses tokenizer.apply_chat_template with role=user for the instruction+dialogue
    and role=assistant for the summary. Returns the raw formatted text (not tokenized).
    This is the string that gets printed in the CLI sanity check.

    Args:
        example: A DialogSum example dict with keys 'dialogue' and 'summary'.
        tokenizer: A Phi-3 tokenizer with apply_chat_template support.

    Returns:
        Formatted string with Phi-3 special tokens embedded.
    """
    messages = [
        {
            "role": "user",
            "content": f"{INSTRUCTION}\n\nConversation:\n{example['dialogue']}",
        },
        {
            "role": "assistant",
            "content": example["summary"],
        },
    ]
    return tokenizer.apply_chat_template(messages, tokenize=False)


# ---------------------------------------------------------------------------
# Tokenization + label masking
# ---------------------------------------------------------------------------

def tokenize_and_mask(
    example: dict,
    tokenizer: PreTrainedTokenizerBase,
    max_length: int = DEFAULT_MAX_LENGTH,
) -> dict:
    """
    Tokenize one DialogSum example and apply prompt-masking on labels.

    Builds input_ids from the full formatted sequence (prompt + summary).
    Builds labels as a copy of input_ids with prompt positions replaced by -100,
    so cross-entropy loss is computed ONLY on the assistant's summary tokens.

    The prompt boundary is determined by tokenizing just the prompt portion
    (with add_generation_prompt=True) and using its token count as the mask cutoff.

    If the full sequence exceeds max_length, transformers truncates from the right
    (cutting the end of the dialogue, not the summary).

    If prompt_len >= len(input_ids) after truncation (extremely long dialogue edge case),
    all labels are -100 and this example contributes zero loss — this is acceptable.

    Args:
        example: A DialogSum example dict with keys 'dialogue' and 'summary'.
        tokenizer: A Phi-3 tokenizer (must have apply_chat_template, padding_side='right').
        max_length: Maximum token sequence length. Defaults to 1024.

    Returns:
        Dict with keys:
            - 'input_ids': List[int], length <= max_length
            - 'attention_mask': List[int], all 1s (pre-padding)
            - 'labels': List[int], same length as input_ids, prompt positions are -100
    """
    # Step 1: build full text (prompt + summary)
    messages_full = [
        {
            "role": "user",
            "content": f"{INSTRUCTION}\n\nConversation:\n{example['dialogue']}",
        },
        {
            "role": "assistant",
            "content": example["summary"],
        },
    ]
    full_text: str = tokenizer.apply_chat_template(messages_full, tokenize=False)

    # Step 2: build prompt-only text (up to and including <|assistant|>\n)
    messages_prompt = [
        {
            "role": "user",
            "content": f"{INSTRUCTION}\n\nConversation:\n{example['dialogue']}",
        },
    ]
    prompt_text: str = tokenizer.apply_chat_template(
        messages_prompt,
        tokenize=False,
        add_generation_prompt=True,
    )

    # Step 3: tokenize full sequence with truncation
    enc = tokenizer(
        full_text,
        add_special_tokens=False,
        max_length=max_length,
        truncation=True,
    )
    input_ids: list[int] = enc["input_ids"]
    attention_mask: list[int] = enc["attention_mask"]

    # Step 4: tokenize prompt only to find the boundary
    prompt_enc = tokenizer(
        prompt_text,
        add_special_tokens=False,
        max_length=max_length,
        truncation=True,
    )
    prompt_len: int = len(prompt_enc["input_ids"])

    # Step 5: clamp and build labels
    # Guard: if an extremely long dialogue was truncated so hard that the summary
    # was completely cut off, prompt_len could equal len(input_ids).
    # In that case all labels are -100 (zero loss contribution). Acceptable.
    prompt_len = min(prompt_len, len(input_ids))
    labels: list[int] = [-100] * prompt_len + input_ids[prompt_len:]

    return {
        "input_ids": input_ids,
        "attention_mask": attention_mask,
        "labels": labels,
    }


# ---------------------------------------------------------------------------
# Dataset preparation
# ---------------------------------------------------------------------------

def prepare_datasets(
    tokenizer: PreTrainedTokenizerBase,
    max_length: int = DEFAULT_MAX_LENGTH,
    dataset_name: str = DATASET_NAME,
    num_proc: int = 4,
) -> tuple[Dataset, Dataset, Dataset]:
    """
    Load DialogSum, format, tokenize, and return train/val/test splits.

    Pipeline:
        1. load_dataset(dataset_name) — uses HF Hub, returns DatasetDict with
           train (~12,460), validation (~500), test (~819) splits.
        2. Dataset.map(tokenize_and_mask, ...) on each split.
        3. Remove original 'id', 'dialogue', 'summary' columns.

    The returned datasets contain only 'input_ids', 'attention_mask', 'labels'.
    They have variable sequence lengths (no padding) — pass the collator returned
    by make_data_collator() to the Trainer's data_collator argument.

    Args:
        tokenizer: Loaded and configured Phi-3 tokenizer. Caller must ensure
                   tokenizer.padding_side == 'right' for training.
        max_length: Max token length for truncation. Defaults to 1024.
        dataset_name: HuggingFace dataset identifier. Defaults to 'knkarthick/dialogsum'.
        num_proc: Number of processes for Dataset.map. Defaults to 4.

    Returns:
        Tuple of (train_dataset, val_dataset, test_dataset).
    """
    raw: DatasetDict = load_dataset(dataset_name)

    # Subsample train split to reduce training time on free Colab T4.
    # Val and test splits are kept in full for accurate evaluation.
    train_size = min(4000, len(raw["train"]))
    raw["train"] = raw["train"].shuffle(seed=42).select(range(train_size))

    _map_fn = functools.partial(
        tokenize_and_mask,
        tokenizer=tokenizer,
        max_length=max_length,
    )

    original_columns = raw["train"].column_names  # ['id', 'dialogue', 'summary']

    tokenized: DatasetDict = raw.map(
        _map_fn,
        batched=False,
        num_proc=num_proc,
        remove_columns=original_columns,
        desc="Tokenizing and masking labels",
    )

    return tokenized["train"], tokenized["validation"], tokenized["test"]


# ---------------------------------------------------------------------------
# Collator (padding)
# ---------------------------------------------------------------------------

def make_data_collator(
    tokenizer: PreTrainedTokenizerBase,
) -> Callable[[list[dict]], dict[str, torch.Tensor]]:
    """
    Return a collate function that pads a batch of tokenized examples.

    Pads 'input_ids' to the longest sequence in the batch using tokenizer.pad_token_id.
    Pads 'attention_mask' with 0.
    Pads 'labels' with -100 (ignored by cross-entropy loss).
    Padding is applied on the RIGHT (required for causal LM training).

    This collator is intentionally minimal — it does NOT do label shifting,
    DataCollatorForSeq2Seq masking, or any other transformation. All label
    masking was already done in tokenize_and_mask().

    Args:
        tokenizer: Must have pad_token_id set (Phi-3: <|endoftext|>, id=32000).

    Returns:
        A collate_fn compatible with HuggingFace Trainer's data_collator argument.
    """
    pad_id: int = tokenizer.pad_token_id

    def collate_fn(batch: list[dict]) -> dict[str, torch.Tensor]:
        max_len = max(len(item["input_ids"]) for item in batch)

        input_ids_padded = []
        attention_mask_padded = []
        labels_padded = []

        for item in batch:
            seq_len = len(item["input_ids"])
            pad_len = max_len - seq_len

            input_ids_padded.append(item["input_ids"] + [pad_id] * pad_len)
            attention_mask_padded.append(item["attention_mask"] + [0] * pad_len)
            labels_padded.append(item["labels"] + [-100] * pad_len)

        return {
            "input_ids": torch.tensor(input_ids_padded, dtype=torch.long),
            "attention_mask": torch.tensor(attention_mask_padded, dtype=torch.long),
            "labels": torch.tensor(labels_padded, dtype=torch.long),
        }

    return collate_fn


# ---------------------------------------------------------------------------
# CLI entry point
# ---------------------------------------------------------------------------

if __name__ == "__main__":
    import random

    from dotenv import load_dotenv
    from transformers import AutoTokenizer

    load_dotenv()  # loads HF_TOKEN from .env if present

    MODEL_ID = "microsoft/Phi-3-mini-4k-instruct"
    print(f"Loading tokenizer: {MODEL_ID} ...")
    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
    tokenizer.padding_side = "right"  # required for training

    print(f"Loading dataset: {DATASET_NAME} ...")
    raw = load_dataset(DATASET_NAME)
    print("\nSplit sizes:")
    print(f"  train:      {len(raw['train']):>6,}")
    print(f"  validation: {len(raw['validation']):>6,}")
    print(f"  test:       {len(raw['test']):>6,}")

    # Show one formatted example
    example = raw["train"][0]
    formatted = format_example(example, tokenizer)
    print("\n--- Formatted example (raw chat template text) ---")
    print(formatted)

    # Sanity check: label masking
    tokenized_example = tokenize_and_mask(example, tokenizer)
    non_masked_ids = [t for t in tokenized_example["labels"] if t != -100]
    decoded = tokenizer.decode(non_masked_ids, skip_special_tokens=True)

    print("\n--- Sanity check: decoded labels vs. original summary ---")
    print(f"Original summary : {example['summary']!r}")
    print(f"Decoded labels   : {decoded!r}")
    match = decoded.strip() == example["summary"].strip()
    print(f"Match            : {match}")
    if not match:
        print("WARNING: Mismatch detected. Check tokenize_and_mask boundary logic.")

    # Token length stats on a random sample of 500 train examples
    sample = random.sample(list(raw["train"]), min(500, len(raw["train"])))
    lengths = []
    for ex in sample:
        enc = tokenize_and_mask(ex, tokenizer)
        lengths.append(len(enc["input_ids"]))

    print(f"\n--- Token length stats (sample n={len(sample)}) ---")
    print(f"  min:    {min(lengths)}")
    print(f"  max:    {max(lengths)}")
    print(f"  mean:   {sum(lengths) / len(lengths):.1f}")
    over_limit = sum(1 for length in lengths if length >= DEFAULT_MAX_LENGTH)
    print(f"  >= {DEFAULT_MAX_LENGTH} (truncated): {over_limit} ({100 * over_limit / len(lengths):.1f}%)")