File size: 12,703 Bytes
11ce4a7 | 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 | """
clean_dataset.py — Extract clean prompts from Telegram AI image dataset.
Groups consecutive images by prompt, strips ads/model names/hashtags,
and outputs a VLM-training-ready JSONL file.
Usage:
python tools/dataview/clean_dataset.py \
--input telegram-channel-dataset/dataset.parquet \
--output telegram-channel-dataset/cleaned.jsonl
"""
import argparse
import json
import re
import sys
from pathlib import Path
import pandas as pd
import pyarrow as pa
import pyarrow.parquet as pq
# ── Patterns to strip ──────────────────────────────────────────────────
STRIP_PATTERNS = [
# Hashtags
r'#[\w\u4e00-\u9fff]+',
# Bot links / Telegram links
r'\[.*?\]\(https?://t\.me/[^\)]+\)',
r'https?://t\.me/\S+',
r'https?://\S+',
# Source/author lines
r'来源:.*',
r'作者:.*',
r'Source:.*',
r'Author:.*',
# Ad / promo lines
r'🚀.*',
r'📚\s*教程目录.*',
r'━━+',
r'🤖.*我们的 Bot.*',
r'🔥.*邪修频道.*',
r'VPN推荐.*',
r'NanoGPT.*',
r'免费赠送.*',
r'快来体验.*',
r'Dubis.*',
r'Bot 机器人.*',
# Model name prefixes in titles
r'^(✨|🔥|🧩|🖼️|🎞️|🎬|🏮|🎨|📷|📸|🌟|💡|🎭|🎪|🎬|🌙|🌅|🌸|🎭|🖌️)\s*',
r'^GPT-?Image[-\s]*2?[||]',
r'^GPTImage2?[||]',
# Title lines (Chinese titles with emoji)
r'^.*?[||].*?(prompt|模板|技巧|构图|写真|人像|海报|封面).*?$',
]
# Fields that contain prompt-relevant information
PROMPT_FIELDS = [
'任务', '主体', '服装', '场景', '光线', '镜头', '风格', '构图',
'约束', '画幅', '关键特征', '变体', '重点', '角色感', '妆造',
'动作', '调色', '反差', '结构', '适用', '示例',
'Task', 'Subject', 'Style', 'Lighting', 'Camera', 'Composition',
'Prompt skeleton',
]
# Chinese field labels to English mapping
FIELD_MAP = {
'任务': 'task',
'主体': 'subject',
'服装': 'clothing',
'场景': 'scene',
'光线': 'lighting',
'镜头': 'camera',
'风格': 'style',
'构图': 'composition',
'约束': 'avoid',
'画幅': 'aspect_ratio',
'关键特征': 'key_features',
'变体': 'variants',
'重点': 'focus',
'角色感': 'character',
'妆造': 'makeup',
'动作': 'pose',
'调色': 'color_grading',
'反差': 'contrast',
'结构': 'layout',
'适用': 'use_case',
'示例': 'examples',
'Task': 'task',
'Subject': 'subject',
'Style': 'style',
'Lighting': 'lighting',
'Camera': 'camera',
'Composition': 'composition',
'Prompt skeleton': 'prompt_skeleton',
}
def clean_text(text: str) -> str:
"""Remove ads, links, hashtags, model names, and other noise."""
if not text or pd.isna(text):
return ''
text = str(text)
# Apply strip patterns
for pattern in STRIP_PATTERNS:
text = re.sub(pattern, '', text, flags=re.MULTILINE | re.IGNORECASE)
# Remove lines that are just emojis or very short
lines = text.split('\n')
cleaned_lines = []
for line in lines:
line = line.strip()
if not line:
continue
# Skip very short lines (likely noise)
if len(line) < 3:
continue
# Skip lines that are mostly emojis
emoji_chars = len(re.findall(r'[\U0001F300-\U0001F9FF]', line))
if emoji_chars > len(line) * 0.5:
continue
cleaned_lines.append(line)
return '\n'.join(cleaned_lines).strip()
def extract_structured_fields(text: str) -> dict:
"""Extract structured prompt fields from Chinese text."""
if not text:
return {}
fields = {}
# Match patterns like "字段名:值" or "字段名: value"
field_names = '|'.join(re.escape(f) for f in PROMPT_FIELDS)
field_pattern = re.compile(
rf'^(?:[-•]\s*)?({field_names})[::]\s*(.+?)(?:\n|$)',
re.MULTILINE
)
for match in field_pattern.finditer(text):
field_name = match.group(1).strip()
value = match.group(2).strip()
if value and len(value) > 2:
eng_name = FIELD_MAP.get(field_name, field_name)
fields[eng_name] = value
return fields
def extract_prompt_skeleton(text: str) -> str:
"""Extract 'Prompt skeleton' section if present."""
if not text:
return ''
match = re.search(r'Prompt skeleton[:\s]*\n(.+?)(?:\n\n|\n备注|\Z)', text, re.DOTALL)
if match:
return match.group(1).strip()
return ''
def build_prompt(text: str) -> str:
"""Build a clean prompt from the structured fields."""
if not text:
return ''
# Try to get prompt skeleton first (most direct)
skeleton = extract_prompt_skeleton(text)
if skeleton and len(skeleton) > 20:
return skeleton
# Extract structured fields
fields = extract_structured_fields(text)
if not fields:
# Fallback: try to extract any English prompt-like content
lines = text.split('\n')
prompt_parts = []
for line in lines:
line = line.strip()
# Skip Chinese-only lines
chinese_chars = len(re.findall(r'[\u4e00-\u9fff]', line))
if chinese_chars > len(line) * 0.3:
continue
# Skip very short lines
if len(line) < 10:
continue
# Skip known noise
if any(skip in line.lower() for skip in ['bot', 'http', '#', '来源', '作者', 'source', 'author']):
continue
prompt_parts.append(line)
if prompt_parts:
return '; '.join(prompt_parts)
return ''
# Build prompt from fields in logical order
order = ['task', 'subject', 'character', 'makeup', 'clothing', 'pose',
'scene', 'lighting', 'camera', 'composition', 'aspect_ratio',
'key_features', 'style', 'color_grading', 'contrast', 'layout',
'avoid', 'use_case', 'prompt_skeleton']
parts = []
for key in order:
if key in fields:
parts.append(fields[key])
# Add any remaining fields not in order
for key, val in fields.items():
if key not in order and val:
parts.append(val)
return '; '.join(parts) if parts else ''
def group_images(df: pd.DataFrame) -> list:
"""Group consecutive images by prompt (same dimensions = same generation)."""
groups = []
current_group = None
for idx, row in df.iterrows():
has_text = row['text'] and str(row['text']).strip()
has_image = row['image'] is not None and isinstance(row['image'], bytes) and len(row['image']) > 100
if has_text and has_image:
# Start new group
if current_group:
groups.append(current_group)
current_group = {
'message_id': row['message_id'],
'datetime': row['datetime'],
'raw_text': row['text'],
'images': [{
'row_idx': idx,
'message_id': row['message_id'],
'width': row['width'],
'height': row['height'],
'image_bytes': row['image'],
}]
}
elif current_group and has_image:
# Check if dimensions match (same generation batch)
last_img = current_group['images'][-1]
if row['width'] == last_img['width'] and row['height'] == last_img['height']:
current_group['images'].append({
'row_idx': idx,
'message_id': row['message_id'],
'width': row['width'],
'height': row['height'],
'image_bytes': row['image'],
})
if current_group:
groups.append(current_group)
return groups
def main():
parser = argparse.ArgumentParser(description='Clean Telegram AI image dataset')
parser.add_argument('--input', '-i', required=True, help='Input parquet file')
parser.add_argument('--output', '-o', default=None, help='Output JSONL file')
parser.add_argument('--parquet', default=None, help='Output cleaned parquet (with only useful columns)')
parser.add_argument('--min-images', type=int, default=1, help='Min images per group')
parser.add_argument('--min-prompt-len', type=int, default=10, help='Min prompt length')
parser.add_argument('--stats', action='store_true', help='Print stats only')
args = parser.parse_args()
# Read parquet
print(f"Reading {args.input}...")
df = pq.read_table(args.input).to_pandas()
print(f" Total rows: {len(df)}")
# Optionally output cleaned parquet (strip noise columns)
if args.parquet:
KEEP = ['message_id', 'datetime', 'media_type', 'text', 'width', 'height', 'image']
available = [c for c in KEEP if c in df.columns]
clean = df[available].copy()
if 'media_type' in clean.columns:
clean = clean[clean['media_type'] == 'photo'].copy()
clean.reset_index(drop=True, inplace=True)
pq.write_table(pa.Table.from_pandas(clean), args.parquet)
print(f" Cleaned parquet: {len(clean)} rows, {len(clean.columns)} cols -> {args.parquet}")
# Group images
groups = group_images(df)
print(f" Prompt groups: {len(groups)}")
# Clean and filter
cleaned = []
for group in groups:
prompt = build_prompt(group['raw_text'])
if len(prompt) < args.min_prompt_len:
continue
if len(group['images']) < args.min_images:
continue
cleaned.append({
'prompt': prompt,
'raw_text': group['raw_text'],
'num_images': len(group['images']),
'message_id': group['message_id'],
'datetime': group['datetime'],
'images': group['images'],
})
print(f" Cleaned groups: {len(cleaned)}")
print(f" Total images: {sum(g['num_images'] for g in cleaned)}")
if args.stats:
# Print distribution
from collections import Counter
sizes = Counter(g['num_images'] for g in cleaned)
print("\n Group size distribution:")
for size, count in sorted(sizes.items()):
print(f" {size} images: {count} groups")
# Show sample prompts
print("\n Sample prompts:")
for g in cleaned[:5]:
print(f" [{g['num_images']} imgs] {g['prompt'][:120]}...")
return
# Output
if args.output:
output_path = Path(args.output)
output_path.parent.mkdir(parents=True, exist_ok=True)
with open(output_path, 'w', encoding='utf-8') as f:
for group in cleaned:
record = {
'prompt': group['prompt'],
'num_images': group['num_images'],
'message_id': group['message_id'],
'datetime': group['datetime'],
'image_paths': [
f"images/msg_{img['message_id']}_row_{img['row_idx']}.webp"
for img in group['images']
],
}
f.write(json.dumps(record, ensure_ascii=False) + '\n')
print(f"\n Written to {output_path}")
# Also save images
img_dir = output_path.parent / 'images'
img_dir.mkdir(exist_ok=True)
from PIL import Image
import io
print(" Extracting images...")
saved = 0
for group in cleaned:
for img_data in group['images']:
try:
img = Image.open(io.BytesIO(img_data['image_bytes']))
img_path = img_dir / f"msg_{img_data['message_id']}_row_{img_data['row_idx']}.webp"
img.save(img_path, format='WEBP', quality=90)
saved += 1
except Exception as e:
print(f" Warning: Failed to save image {img_data['message_id']}: {e}")
print(f" Saved {saved} images to {img_dir}")
else:
# Just print stats
from collections import Counter
sizes = Counter(g['num_images'] for g in cleaned)
print("\n Group size distribution:")
for size, count in sorted(sizes.items()):
print(f" {size} images: {count} groups")
print("\n Sample prompts:")
for g in cleaned[:10]:
print(f" [{g['num_images']} imgs] {g['prompt'][:150]}...")
if __name__ == '__main__':
main()
|