tinyvic / dataset.py
Viclim's picture
Upload 17 files
9299fff verified
"""
VicAI Dataset
Dataset handling for training on Wikipedia and other text sources.
"""
import os
import random
import re
from typing import Dict, Iterator, List, Optional
import requests
import torch
from torch.utils.data import Dataset, IterableDataset
class WikipediaDataset(IterableDataset):
"""Stream Wikipedia articles for training."""
def __init__(
self,
tokenizer,
max_length: int = 2048,
languages: List[str] = ['en'],
min_article_length: int = 100,
):
self.tokenizer = tokenizer
self.max_length = max_length
self.languages = languages
self.min_article_length = min_article_length
self.base_url = "https://en.wikipedia.org/w/api.php"
def _fetch_random_article(self) -> Optional[str]:
"""Fetch a random Wikipedia article."""
try:
params = {
'action': 'query',
'format': 'json',
'generator': 'random',
'grnnamespace': 0,
'grnlimit': 1,
'prop': 'extracts',
'explaintext': True,
'exsentences': 50,
}
response = requests.get(self.base_url, params=params, timeout=10)
data = response.json()
pages = data['query']['pages']
for page_id, page_data in pages.items():
text = page_data.get('extract', '')
if len(text) > self.min_article_length:
return text
return None
except Exception as e:
print(f"Error fetching article: {e}")
return None
def _fetch_article_by_title(self, title: str) -> Optional[str]:
"""Fetch a specific Wikipedia article by title."""
try:
params = {
'action': 'query',
'format': 'json',
'titles': title,
'prop': 'extracts',
'explaintext': True,
'exlimit': 1,
}
response = requests.get(self.base_url, params=params, timeout=10)
data = response.json()
pages = data['query']['pages']
for page_id, page_data in pages.items():
if page_id != '-1':
return page_data.get('extract', '')
return None
except Exception as e:
print(f"Error fetching article: {e}")
return None
def _clean_text(self, text: str) -> str:
"""Clean Wikipedia text."""
# Remove citation markers
text = re.sub(r'\[\d+\]', '', text)
# Remove multiple spaces
text = re.sub(r'\s+', ' ', text)
# Remove special characters but keep basic punctuation
text = re.sub(r'[^\w\s.,!?;:\'\"()-]', ' ', text)
return text.strip()
def _tokenize_text(self, text: str) -> List[int]:
"""Tokenize text and create chunks."""
cleaned = self._clean_text(text)
tokens = self.tokenizer.encode(cleaned, add_special_tokens=True)
return tokens
def __iter__(self):
"""Iterate over Wikipedia articles."""
while True:
text = self._fetch_random_article()
if text:
tokens = self._tokenize_text(text)
# Create chunks of max_length
for i in range(0, len(tokens), self.max_length):
chunk = tokens[i:i + self.max_length]
if len(chunk) > 10: # Minimum chunk size
# Pad if necessary
if len(chunk) < self.max_length:
chunk.extend([self.tokenizer.pad_token_id] * (self.max_length - len(chunk)))
input_ids = torch.tensor(chunk[:-1])
labels = torch.tensor(chunk[1:])
yield {
'input_ids': input_ids,
'labels': labels,
'attention_mask': (input_ids != self.tokenizer.pad_token_id).long(),
}
class TextFileDataset(Dataset):
"""Dataset from local text files."""
def __init__(
self,
file_path: str,
tokenizer,
max_length: int = 2048,
stride: int = 512,
):
self.tokenizer = tokenizer
self.max_length = max_length
self.stride = stride
print(f"Loading dataset from {file_path}...")
with open(file_path, 'r', encoding='utf-8') as f:
text = f.read()
# Tokenize full text
self.tokens = tokenizer.encode(text, add_special_tokens=False)
# Create chunks
self.chunks = []
for i in range(0, len(self.tokens) - max_length, stride):
chunk = self.tokens[i:i + max_length + 1]
if len(chunk) == max_length + 1:
self.chunks.append(chunk)
print(f"Created {len(self.chunks)} chunks from {len(self.tokens)} tokens")
def __len__(self):
return len(self.chunks)
def __getitem__(self, idx):
chunk = self.chunks[idx]
input_ids = torch.tensor(chunk[:-1])
labels = torch.tensor(chunk[1:])
return {
'input_ids': input_ids,
'labels': labels,
'attention_mask': torch.ones_like(input_ids),
}
class MixedDataset(IterableDataset):
"""Mix multiple data sources."""
def __init__(
self,
datasets: List[IterableDataset],
weights: Optional[List[float]] = None,
):
self.datasets = datasets
self.weights = weights or [1.0] * len(datasets)
assert len(self.datasets) == len(self.weights)
# Normalize weights
total = sum(self.weights)
self.weights = [w / total for w in self.weights]
def __iter__(self):
"""Sample from datasets according to weights."""
iterators = [iter(ds) for ds in self.datasets]
while True:
# Choose dataset based on weights
dataset_idx = random.choices(range(len(self.datasets)), weights=self.weights)[0]
try:
yield next(iterators[dataset_idx])
except StopIteration:
# Restart iterator if exhausted
iterators[dataset_idx] = iter(self.datasets[dataset_idx])
yield next(iterators[dataset_idx])
class PretokenizedDataset(Dataset):
"""Dataset from pre-tokenized binary files."""
def __init__(self, data_dir: str, max_length: int = 2048):
self.data_dir = data_dir
self.max_length = max_length
# Load all .pt files
self.files = []
for fname in os.listdir(data_dir):
if fname.endswith('.pt'):
self.files.append(os.path.join(data_dir, fname))
self.files.sort()
print(f"Found {len(self.files)} pre-tokenized files")
# Load metadata
self.lengths = []
for f in self.files:
data = torch.load(f, map_location='cpu')
self.lengths.append(len(data) // max_length)
self.total_length = sum(self.lengths)
def __len__(self):
return self.total_length
def __getitem__(self, idx):
# Find which file contains this index
file_idx = 0
remaining = idx
for i, length in enumerate(self.lengths):
if remaining < length:
file_idx = i
break
remaining -= length
# Load data
data = torch.load(self.files[file_idx], map_location='cpu')
start = remaining * self.max_length
chunk = data[start:start + self.max_length + 1]
input_ids = chunk[:-1]
labels = chunk[1:]
return {
'input_ids': input_ids,
'labels': labels,
'attention_mask': torch.ones_like(input_ids),
}
def download_wikipedia_dump(output_dir: str, language: str = 'en'):
"""Download Wikipedia dump for offline processing."""
os.makedirs(output_dir, exist_ok=True)
# Wikipedia dump URLs
base_url = f"https://dumps.wikimedia.org/{language}wiki/latest/"
files = [
f"{language}wiki-latest-pages-articles-multistream.xml.bz2",
]
for filename in files:
url = base_url + filename
output_path = os.path.join(output_dir, filename)
if os.path.exists(output_path):
print(f"{filename} already exists")
continue
print(f"Downloading {filename}...")
try:
response = requests.get(url, stream=True)
response.raise_for_status()
with open(output_path, 'wb') as f:
for chunk in response.iter_content(chunk_size=8192):
f.write(chunk)
print(f"Downloaded {filename}")
except Exception as e:
print(f"Error downloading {filename}: {e}")
def create_sample_corpus(output_file: str = "sample_corpus.txt", num_articles: int = 1000):
"""Create a sample corpus by fetching Wikipedia articles."""
print(f"Creating sample corpus with {num_articles} articles...")
dataset = WikipediaDataset(
tokenizer=None, # We'll use raw text
max_length=100000, # Large to get full articles
)
articles = []
for i, item in enumerate(dataset):
if i >= num_articles:
break
# Get raw text from the article fetch
text = dataset._fetch_random_article()
if text:
articles.append(text)
if (i + 1) % 100 == 0:
print(f" Fetched {i + 1}/{num_articles} articles")
# Write to file
with open(output_file, 'w', encoding='utf-8') as f:
for article in articles:
f.write(article + '\n\n<|endoftext|>\n\n')
print(f"Sample corpus saved to {output_file}")
return output_file
def prepare_openwebtext_data(output_dir: str, max_files: int = 100):
"""
Download and prepare OpenWebText corpus.
Note: This is a placeholder - actual OpenWebText requires specific download.
"""
os.makedirs(output_dir, exist_ok=True)
print(f"OpenWebText data preparation placeholder")
print(f"Please download OpenWebText from https://github.com/jcpeterson/openwebtext")
print(f"and place files in {output_dir}")
if __name__ == "__main__":
# Test dataset
from tokenizer import BPETokenizer
# Create sample tokenizer
sample_texts = [
"This is a sample text for testing.",
"Wikipedia contains many interesting articles.",
"Machine learning models need lots of data.",
]
tokenizer = BPETokenizer(vocab_size=1000)
tokenizer.train(sample_texts)
# Test Wikipedia dataset
print("\nTesting Wikipedia dataset...")
wiki_dataset = WikipediaDataset(tokenizer, max_length=128)
for i, batch in enumerate(wiki_dataset):
if i >= 3:
break
print(f"\nBatch {i + 1}:")
print(f" Input shape: {batch['input_ids'].shape}")
print(f" Labels shape: {batch['labels'].shape}")