Spaces:
Running
Running
File size: 5,100 Bytes
39028c9 | 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 | """
Fine-tuning module for domain-specific models
"""
import logging
from typing import List, Dict, Optional
import torch
from torch.utils.data import Dataset, DataLoader
from transformers import Seq2SeqTrainer, Seq2SeqTrainingArguments
import json
logger = logging.getLogger(__name__)
class SummarizationDataset(Dataset):
"""Dataset for summarization fine-tuning."""
def __init__(self, documents: List[str], summaries: List[str], tokenizer, max_len: int = 512):
"""
Initialize dataset.
Args:
documents: List of documents
summaries: List of corresponding summaries
tokenizer: HuggingFace tokenizer
max_len: Maximum token length
"""
self.documents = documents
self.summaries = summaries
self.tokenizer = tokenizer
self.max_len = max_len
def __len__(self):
return len(self.documents)
def __getitem__(self, idx):
document = str(self.documents[idx])
summary = str(self.summaries[idx])
# Tokenize input
inputs = self.tokenizer(
document,
max_length=self.max_len,
truncation=True,
padding='max_length',
return_tensors='pt'
)
# Tokenize target
targets = self.tokenizer(
summary,
max_length=256,
truncation=True,
padding='max_length',
return_tensors='pt'
)
return {
'input_ids': inputs['input_ids'].squeeze(),
'attention_mask': inputs['attention_mask'].squeeze(),
'labels': targets['input_ids'].squeeze(),
}
class FineTuner:
"""Fine-tune summarization model on custom data."""
def __init__(self, model, tokenizer, device: str = 'cuda'):
"""
Initialize fine-tuner.
Args:
model: Pre-trained model
tokenizer: Tokenizer
device: Device to use
"""
self.model = model
self.tokenizer = tokenizer
self.device = device
def prepare_data(
self,
data_file: str
) -> tuple:
"""
Prepare data from JSON file.
Format: [{"document": "...", "summary": "..."}, ...]
Args:
data_file: Path to JSON file
Returns:
Tuple of (documents, summaries)
"""
with open(data_file, 'r', encoding='utf-8') as f:
data = json.load(f)
documents = [item['document'] for item in data]
summaries = [item['summary'] for item in data]
logger.info(f"Loaded {len(documents)} training examples")
return documents, summaries
def fine_tune(
self,
documents: List[str],
summaries: List[str],
output_dir: str = 'models/fine_tuned',
num_epochs: int = 3,
batch_size: int = 8,
learning_rate: float = 2e-5
) -> str:
"""
Fine-tune model on custom data.
Args:
documents: Training documents
summaries: Training summaries
output_dir: Output directory
num_epochs: Number of training epochs
batch_size: Batch size
learning_rate: Learning rate
Returns:
Path to fine-tuned model
"""
# Create dataset
train_dataset = SummarizationDataset(
documents,
summaries,
self.tokenizer
)
# Training arguments
training_args = Seq2SeqTrainingArguments(
output_dir=output_dir,
num_train_epochs=num_epochs,
per_device_train_batch_size=batch_size,
learning_rate=learning_rate,
save_steps=len(train_dataset) // batch_size,
save_total_limit=2,
logging_steps=10,
)
# Trainer
trainer = Seq2SeqTrainer(
model=self.model,
args=training_args,
train_dataset=train_dataset,
tokenizer=self.tokenizer,
)
# Train
logger.info("Starting fine-tuning...")
trainer.train()
# Save
trainer.save_model(output_dir)
logger.info(f"Fine-tuned model saved to {output_dir}")
return output_dir
def quick_fine_tune(
self,
data_file: str,
output_dir: str = 'models/fine_tuned'
) -> str:
"""
Quick fine-tuning from JSON file.
Args:
data_file: Path to training data JSON
output_dir: Output directory
Returns:
Path to fine-tuned model
"""
documents, summaries = self.prepare_data(data_file)
return self.fine_tune(
documents,
summaries,
output_dir=output_dir,
num_epochs=2,
batch_size=4
)
|