davanstrien HF Staff commited on
Commit
d272f1c
·
1 Parent(s): 011bc8a
Files changed (1) hide show
  1. classify-dataset.py +356 -123
classify-dataset.py CHANGED
@@ -3,7 +3,7 @@
3
  # requires-python = ">=3.10"
4
  # dependencies = [
5
  # "vllm>=0.6.6",
6
- # "transformers",
7
  # "torch",
8
  # "datasets",
9
  # "huggingface-hub[hf_transfer]",
@@ -36,42 +36,95 @@ import argparse
36
  import logging
37
  import os
38
  import sys
39
- from typing import List, Dict, Any, Optional
40
 
41
  import torch
42
- from datasets import load_dataset, Dataset
43
  from huggingface_hub import HfApi, get_token
 
44
  from vllm import LLM, SamplingParams
45
  from vllm.sampling_params import GuidedDecodingParams
46
 
47
  # Default model - SmolLM3 for good balance of speed and quality
48
  DEFAULT_MODEL = "HuggingFaceTB/SmolLM3-3B"
49
 
50
- # Prompt styles for classification
51
- PROMPT_STYLES = {
52
- "simple": """Classify this text as one of: {labels}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
53
 
54
- Text: {text}
55
-
56
- Label:""",
57
-
58
- "detailed": """Task: Classify the following text into EXACTLY ONE of these categories.
59
- Available categories: {labels}
60
 
61
- Text to classify:
62
- {text}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
63
 
64
- Category:""",
65
-
66
- "reasoning": """Analyze the following text and determine which category it belongs to.
67
- Available categories: {labels}
68
 
69
- Text to analyze:
70
- {text}
71
 
72
- Brief analysis: Let me examine the key aspects of this text.
73
- Category:""",
74
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
75
 
76
  # Minimum text length for valid classification
77
  MIN_TEXT_LENGTH = 3
@@ -79,7 +132,79 @@ MIN_TEXT_LENGTH = 3
79
  # Maximum text length (in characters) to avoid context overflow
80
  MAX_TEXT_LENGTH = 4000
81
 
82
- logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
83
  logger = logging.getLogger(__name__)
84
 
85
 
@@ -87,87 +212,99 @@ def parse_args():
87
  parser = argparse.ArgumentParser(
88
  description="Classify text in HuggingFace datasets using vLLM with structured outputs",
89
  formatter_class=argparse.RawDescriptionHelpFormatter,
90
- epilog=__doc__
91
  )
92
-
93
  # Required arguments
94
  parser.add_argument(
95
  "--input-dataset",
96
  type=str,
97
  required=True,
98
- help="Input dataset ID on Hugging Face Hub"
99
  )
100
  parser.add_argument(
101
- "--column",
102
- type=str,
103
- required=True,
104
- help="Name of the text column to classify"
105
  )
106
  parser.add_argument(
107
  "--labels",
108
  type=str,
109
  required=True,
110
- help="Comma-separated list of classification labels (e.g., 'positive,negative')"
111
  )
112
  parser.add_argument(
113
  "--output-dataset",
114
  type=str,
115
  required=True,
116
- help="Output dataset ID on Hugging Face Hub"
117
  )
118
-
119
  # Optional arguments
120
  parser.add_argument(
121
  "--model",
122
  type=str,
123
  default=DEFAULT_MODEL,
124
- help=f"Model to use for classification (default: {DEFAULT_MODEL})"
125
  )
126
  # Removed --batch-size argument as vLLM handles batching internally
127
  parser.add_argument(
128
- "--prompt-style",
129
  type=str,
130
- choices=list(PROMPT_STYLES.keys()),
131
- default="simple",
132
- help="Prompt style to use (default: simple)"
 
 
 
 
133
  )
134
  parser.add_argument(
135
  "--max-samples",
136
  type=int,
137
  default=None,
138
- help="Maximum number of samples to process (for testing)"
139
  )
140
  parser.add_argument(
141
  "--hf-token",
142
  type=str,
143
  default=None,
144
- help="Hugging Face API token (default: auto-detect from HF_TOKEN env var or huggingface-cli login)"
145
  )
146
  parser.add_argument(
147
  "--split",
148
  type=str,
149
  default="train",
150
- help="Dataset split to process (default: train)"
151
  )
152
  parser.add_argument(
153
  "--temperature",
154
  type=float,
155
  default=0.1,
156
- help="Temperature for generation (default: 0.1)"
157
  )
158
  parser.add_argument(
159
  "--max-tokens",
160
  type=int,
161
- default=50,
162
- help="Maximum tokens to generate (default: 50)"
163
  )
164
  parser.add_argument(
165
  "--guided-backend",
166
  type=str,
167
  default="outlines",
168
- help="Guided decoding backend (default: outlines)"
169
  )
170
-
 
 
 
 
 
 
 
 
 
 
 
171
  return parser.parse_args()
172
 
173
 
@@ -175,63 +312,62 @@ def preprocess_text(text: str) -> str:
175
  """Preprocess text for classification."""
176
  if not text or not isinstance(text, str):
177
  return ""
178
-
179
  # Strip whitespace
180
  text = text.strip()
181
-
182
  # Truncate if too long
183
  if len(text) > MAX_TEXT_LENGTH:
184
- text = text[:MAX_TEXT_LENGTH] + "..."
185
-
186
  return text
187
 
188
 
189
  def validate_text(text: str) -> bool:
190
  """Check if text is valid for classification."""
191
- if not text or len(text) < MIN_TEXT_LENGTH:
192
- return False
193
- return True
194
 
195
 
196
  def prepare_prompts(
197
- texts: List[str],
198
- labels: List[str],
199
- prompt_template: str
200
  ) -> tuple[List[str], List[int]]:
201
- """Prepare prompts for classification, filtering invalid texts."""
202
  prompts = []
203
  valid_indices = []
204
-
205
  for i, text in enumerate(texts):
206
  processed_text = preprocess_text(text)
207
  if validate_text(processed_text):
208
- prompt = prompt_template.format(
209
- labels=", ".join(labels),
210
- text=processed_text
 
 
 
 
 
211
  )
212
  prompts.append(prompt)
213
  valid_indices.append(i)
214
-
215
  return prompts, valid_indices
216
 
217
 
218
  def main():
219
  args = parse_args()
220
-
221
  # Check authentication early
222
  logger.info("Checking authentication...")
223
- token = args.hf_token
224
- if not token:
225
- # Try to get token from environment or huggingface-cli login
226
- token = os.environ.get("HF_TOKEN") or get_token()
227
-
228
  if not token:
229
  logger.error("No authentication token found. Please either:")
230
  logger.error("1. Run 'huggingface-cli login'")
231
  logger.error("2. Set HF_TOKEN environment variable")
232
  logger.error("3. Pass --hf-token argument")
233
  sys.exit(1)
234
-
235
  # Validate token by checking who we are
236
  try:
237
  api = HfApi(token=token)
@@ -241,15 +377,15 @@ def main():
241
  logger.error(f"Authentication failed: {e}")
242
  logger.error("Please check your token is valid")
243
  sys.exit(1)
244
-
245
  # Check CUDA availability
246
  if not torch.cuda.is_available():
247
  logger.error("CUDA is not available. This script requires a GPU.")
248
  logger.error("Please run on a machine with GPU support or use HF Jobs.")
249
  sys.exit(1)
250
-
251
  logger.info(f"CUDA available. Using device: {torch.cuda.get_device_name(0)}")
252
-
253
  # Parse and validate labels
254
  labels = [label.strip() for label in args.labels.split(",")]
255
  if len(labels) < 2:
@@ -257,30 +393,52 @@ def main():
257
  sys.exit(1)
258
  logger.info(f"Classification labels: {labels}")
259
 
 
 
 
 
 
 
 
 
260
  # Load dataset
261
  logger.info(f"Loading dataset: {args.input_dataset}")
262
  try:
263
  dataset = load_dataset(args.input_dataset, split=args.split)
264
-
 
 
 
 
 
 
265
  # Limit samples if specified
266
  if args.max_samples:
267
  dataset = dataset.select(range(min(args.max_samples, len(dataset))))
268
  logger.info(f"Limited dataset to {len(dataset)} samples")
269
-
270
- logger.info(f"Loaded {len(dataset)} samples from split '{args.split}'")
271
  except Exception as e:
272
  logger.error(f"Failed to load dataset: {e}")
273
  sys.exit(1)
274
-
275
  # Verify column exists
276
  if args.column not in dataset.column_names:
277
  logger.error(f"Column '{args.column}' not found in dataset.")
278
  logger.error(f"Available columns: {dataset.column_names}")
279
  sys.exit(1)
280
-
281
  # Extract texts
282
  texts = dataset[args.column]
283
-
 
 
 
 
 
 
 
 
284
  # Initialize vLLM
285
  logger.info(f"Initializing vLLM with model: {args.model}")
286
  logger.info(f"Using guided decoding backend: {args.guided_backend}")
@@ -295,85 +453,139 @@ def main():
295
  except Exception as e:
296
  logger.error(f"Failed to initialize vLLM: {e}")
297
  sys.exit(1)
298
-
299
- # Set up guided decoding parameters
300
- guided_params = GuidedDecodingParams(choice=labels)
301
-
302
- # Set up sampling parameters with structured output
303
- sampling_params = SamplingParams(
304
- guided_decoding=guided_params,
305
- temperature=args.temperature,
306
- max_tokens=args.max_tokens,
307
- )
308
-
309
- # Get prompt template
310
- prompt_template = PROMPT_STYLES[args.prompt_style]
311
- logger.info(f"Using prompt style '{args.prompt_style}'")
312
- logger.info("Using structured output with guided_choice - outputs guaranteed to be valid labels")
313
-
 
 
 
314
  # Prepare all prompts
315
  logger.info("Preparing prompts for classification...")
316
- all_prompts, valid_indices = prepare_prompts(texts, labels, prompt_template)
317
-
318
  if not all_prompts:
319
  logger.error("No valid texts found for classification.")
320
  sys.exit(1)
321
-
322
  logger.info(f"Prepared {len(all_prompts)} valid prompts out of {len(texts)} texts")
323
-
324
  # Let vLLM handle batching internally
325
  logger.info("Starting classification (vLLM will handle batching internally)...")
326
-
327
  try:
328
  # Generate all classifications at once - vLLM handles batching
329
  outputs = llm.generate(all_prompts, sampling_params)
330
-
331
- # Map results back to original indices
332
- all_classifications = [None] * len(texts)
333
- for idx, output in enumerate(outputs):
334
- original_idx = valid_indices[idx]
335
- generated_text = output.outputs[0].text.strip()
336
- all_classifications[original_idx] = generated_text
337
-
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
338
  # Count statistics
339
- valid_texts = len(valid_indices)
340
  total_texts = len(texts)
341
-
342
  except Exception as e:
343
  logger.error(f"Classification failed: {e}")
344
  sys.exit(1)
345
-
346
- # Add classifications to dataset
347
  dataset = dataset.add_column("classification", all_classifications)
348
 
 
 
 
 
349
  # Calculate statistics
350
  none_count = total_texts - valid_texts
351
  if none_count > 0:
352
- logger.warning(f"{none_count} texts were too short or invalid for classification")
353
-
 
 
354
  # Show classification distribution
355
  label_counts = {label: all_classifications.count(label) for label in labels}
 
 
 
 
356
  logger.info("Classification distribution:")
357
  for label, count in label_counts.items():
358
  percentage = count / total_texts * 100 if total_texts > 0 else 0
359
  logger.info(f" {label}: {count} ({percentage:.1f}%)")
360
- if none_count > 0:
361
- none_percentage = none_count / total_texts * 100
362
- logger.info(f" Invalid/Skipped: {none_count} ({none_percentage:.1f}%)")
363
 
 
 
 
 
 
 
 
364
  # Log success rate
365
  success_rate = (valid_texts / total_texts * 100) if total_texts > 0 else 0
366
  logger.info(f"Classification success rate: {success_rate:.1f}%")
367
-
368
  # Save to Hub (token already validated at start)
369
  logger.info(f"Pushing dataset to Hub: {args.output_dataset}")
370
  try:
371
  dataset.push_to_hub(
372
  args.output_dataset,
373
  token=token,
374
- commit_message=f"Add classifications using {args.model} with structured outputs"
 
 
 
375
  )
376
- logger.info(f"Successfully pushed to: https://huggingface.co/datasets/{args.output_dataset}")
377
  except Exception as e:
378
  logger.error(f"Failed to push to Hub: {e}")
379
  sys.exit(1)
@@ -381,7 +593,28 @@ def main():
381
 
382
  if __name__ == "__main__":
383
  if len(sys.argv) == 1:
384
- print("Example HF Jobs command:")
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
385
  print("hf jobs uv run \\")
386
  print(" --flavor l4x1 \\")
387
  print(" --image vllm/vllm-openai:latest \\")
@@ -391,5 +624,5 @@ if __name__ == "__main__":
391
  print(" --labels 'positive,negative' \\")
392
  print(" --output-dataset user/imdb-classified")
393
  sys.exit(0)
394
-
395
- main()
 
3
  # requires-python = ">=3.10"
4
  # dependencies = [
5
  # "vllm>=0.6.6",
6
+ # "transformers>=4.53.0",
7
  # "torch",
8
  # "datasets",
9
  # "huggingface-hub[hf_transfer]",
 
36
  import logging
37
  import os
38
  import sys
39
+ from typing import List
40
 
41
  import torch
42
+ from datasets import load_dataset
43
  from huggingface_hub import HfApi, get_token
44
+ from transformers import AutoTokenizer
45
  from vllm import LLM, SamplingParams
46
  from vllm.sampling_params import GuidedDecodingParams
47
 
48
  # Default model - SmolLM3 for good balance of speed and quality
49
  DEFAULT_MODEL = "HuggingFaceTB/SmolLM3-3B"
50
 
51
+ def parse_label_descriptions(desc_string: str) -> dict:
52
+ """Parse label descriptions from CLI format 'label1:desc1,label2:desc2'."""
53
+ if not desc_string:
54
+ return {}
55
+
56
+ descriptions = {}
57
+ # Split by comma, but be careful about commas in descriptions
58
+ parts = desc_string.split(',')
59
+
60
+ current_label = None
61
+ current_desc_parts = []
62
+
63
+ for part in parts:
64
+ if ':' in part and not current_label:
65
+ # New label:description pair
66
+ label, desc = part.split(':', 1)
67
+ current_label = label.strip()
68
+ current_desc_parts = [desc.strip()]
69
+ elif ':' in part and current_label:
70
+ # Save previous label and start new one
71
+ descriptions[current_label] = ','.join(current_desc_parts)
72
+ label, desc = part.split(':', 1)
73
+ current_label = label.strip()
74
+ current_desc_parts = [desc.strip()]
75
+ else:
76
+ # Continuation of previous description (had comma in it)
77
+ current_desc_parts.append(part.strip())
78
+
79
+ # Don't forget the last one
80
+ if current_label:
81
+ descriptions[current_label] = ','.join(current_desc_parts)
82
+
83
+ return descriptions
84
 
 
 
 
 
 
 
85
 
86
+ def create_messages(text: str, labels: List[str], label_descriptions: dict = None, enable_reasoning: bool = False) -> List[dict]:
87
+ """Create messages for chat template with optional label descriptions."""
88
+
89
+ # Build the classification prompt
90
+ if label_descriptions:
91
+ # Format with descriptions
92
+ categories_text = "Categories:\n"
93
+ for label in labels:
94
+ desc = label_descriptions.get(label, "")
95
+ if desc:
96
+ categories_text += f"- {label}: {desc}\n"
97
+ else:
98
+ categories_text += f"- {label}\n"
99
+ else:
100
+ # Simple format without descriptions
101
+ categories_text = f"Categories: {', '.join(labels)}"
102
+
103
+ if enable_reasoning:
104
+ # Reasoning mode: allow thinking and request JSON output
105
+ user_content = f"""Classify this text into one of these categories:
106
 
107
+ {categories_text}
 
 
 
108
 
109
+ Text: {text}
 
110
 
111
+ Think through your classification step by step, then provide your final answer in this JSON format:
112
+ {{"label": "your_chosen_label"}}"""
113
+
114
+ system_content = "You are a helpful classification assistant that thinks step by step."
115
+ else:
116
+ # Structured output mode: fast classification
117
+ if label_descriptions:
118
+ user_content = f"Classify this text into one of these categories:\n\n{categories_text}\nText: {text}\n\nCategory:"
119
+ else:
120
+ user_content = f"Classify this text as one of: {', '.join(labels)}\n\nText: {text}\n\nLabel:"
121
+
122
+ system_content = "You are a helpful classification assistant. /no_think"
123
+
124
+ return [
125
+ {"role": "system", "content": system_content},
126
+ {"role": "user", "content": user_content}
127
+ ]
128
 
129
  # Minimum text length for valid classification
130
  MIN_TEXT_LENGTH = 3
 
132
  # Maximum text length (in characters) to avoid context overflow
133
  MAX_TEXT_LENGTH = 4000
134
 
135
+
136
+ def parse_reasoning_output(output: str, valid_labels: List[str]) -> tuple[str, str, bool]:
137
+ """Parse reasoning output to extract label from JSON after </think> tag.
138
+
139
+ Returns:
140
+ tuple: (label or None, full reasoning text, parsing_success)
141
+ """
142
+ import json
143
+
144
+ # Find the </think> tag
145
+ think_end = output.find("</think>")
146
+
147
+ if think_end != -1:
148
+ # Extract everything after </think>
149
+ json_part = output[think_end + len("</think>"):].strip()
150
+ reasoning = output[:think_end + len("</think>")]
151
+ else:
152
+ # No think tags, look for JSON in the output
153
+ # Try to find JSON by looking for {
154
+ json_start = output.find("{")
155
+ if json_start != -1:
156
+ json_part = output[json_start:].strip()
157
+ reasoning = output[:json_start].strip() if json_start > 0 else ""
158
+ else:
159
+ json_part = output
160
+ reasoning = output
161
+
162
+ # Try to parse JSON
163
+ try:
164
+ # Find the first complete JSON object
165
+ if "{" in json_part:
166
+ # Extract just the JSON object
167
+ json_str = json_part[json_part.find("{"):]
168
+ # Find the matching closing brace
169
+ brace_count = 0
170
+ end_pos = 0
171
+ for i, char in enumerate(json_str):
172
+ if char == "{":
173
+ brace_count += 1
174
+ elif char == "}":
175
+ brace_count -= 1
176
+ if brace_count == 0:
177
+ end_pos = i + 1
178
+ break
179
+
180
+ if end_pos > 0:
181
+ json_str = json_str[:end_pos]
182
+ data = json.loads(json_str)
183
+ label = data.get("label", "")
184
+
185
+ # Validate label
186
+ if label in valid_labels:
187
+ return label, output, True
188
+ else:
189
+ logger.warning(f"Parsed label '{label}' not in valid labels: {valid_labels}")
190
+ return None, output, False
191
+ else:
192
+ logger.warning("Could not find complete JSON object")
193
+ return None, output, False
194
+ else:
195
+ logger.warning("No JSON found in output")
196
+ return None, output, False
197
+
198
+ except json.JSONDecodeError as e:
199
+ logger.warning(f"JSON parsing error: {e}")
200
+ return None, output, False
201
+ except Exception as e:
202
+ logger.warning(f"Unexpected error parsing output: {e}")
203
+ return None, output, False
204
+
205
+ logging.basicConfig(
206
+ level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s"
207
+ )
208
  logger = logging.getLogger(__name__)
209
 
210
 
 
212
  parser = argparse.ArgumentParser(
213
  description="Classify text in HuggingFace datasets using vLLM with structured outputs",
214
  formatter_class=argparse.RawDescriptionHelpFormatter,
215
+ epilog=__doc__,
216
  )
217
+
218
  # Required arguments
219
  parser.add_argument(
220
  "--input-dataset",
221
  type=str,
222
  required=True,
223
+ help="Input dataset ID on Hugging Face Hub",
224
  )
225
  parser.add_argument(
226
+ "--column", type=str, required=True, help="Name of the text column to classify"
 
 
 
227
  )
228
  parser.add_argument(
229
  "--labels",
230
  type=str,
231
  required=True,
232
+ help="Comma-separated list of classification labels (e.g., 'positive,negative')",
233
  )
234
  parser.add_argument(
235
  "--output-dataset",
236
  type=str,
237
  required=True,
238
+ help="Output dataset ID on Hugging Face Hub",
239
  )
240
+
241
  # Optional arguments
242
  parser.add_argument(
243
  "--model",
244
  type=str,
245
  default=DEFAULT_MODEL,
246
+ help=f"Model to use for classification (default: {DEFAULT_MODEL})",
247
  )
248
  # Removed --batch-size argument as vLLM handles batching internally
249
  parser.add_argument(
250
+ "--label-descriptions",
251
  type=str,
252
+ default=None,
253
+ help="Descriptions for each label in format 'label1:description1,label2:description2'",
254
+ )
255
+ parser.add_argument(
256
+ "--enable-reasoning",
257
+ action="store_true",
258
+ help="Enable reasoning mode with thinking traces (disables structured outputs)",
259
  )
260
  parser.add_argument(
261
  "--max-samples",
262
  type=int,
263
  default=None,
264
+ help="Maximum number of samples to process (for testing)",
265
  )
266
  parser.add_argument(
267
  "--hf-token",
268
  type=str,
269
  default=None,
270
+ help="Hugging Face API token (default: auto-detect from HF_TOKEN env var or huggingface-cli login)",
271
  )
272
  parser.add_argument(
273
  "--split",
274
  type=str,
275
  default="train",
276
+ help="Dataset split to process (default: train)",
277
  )
278
  parser.add_argument(
279
  "--temperature",
280
  type=float,
281
  default=0.1,
282
+ help="Temperature for generation (default: 0.1)",
283
  )
284
  parser.add_argument(
285
  "--max-tokens",
286
  type=int,
287
+ default=100,
288
+ help="Maximum tokens to generate (default: 100, automatically increased 20x for reasoning mode)",
289
  )
290
  parser.add_argument(
291
  "--guided-backend",
292
  type=str,
293
  default="outlines",
294
+ help="Guided decoding backend (default: outlines)",
295
  )
296
+ parser.add_argument(
297
+ "--shuffle",
298
+ action="store_true",
299
+ help="Shuffle dataset before selecting samples (useful with --max-samples for random sampling)",
300
+ )
301
+ parser.add_argument(
302
+ "--shuffle-seed",
303
+ type=int,
304
+ default=42,
305
+ help="Random seed for shuffling (default: 42)",
306
+ )
307
+
308
  return parser.parse_args()
309
 
310
 
 
312
  """Preprocess text for classification."""
313
  if not text or not isinstance(text, str):
314
  return ""
315
+
316
  # Strip whitespace
317
  text = text.strip()
318
+
319
  # Truncate if too long
320
  if len(text) > MAX_TEXT_LENGTH:
321
+ text = f"{text[:MAX_TEXT_LENGTH]}..."
322
+
323
  return text
324
 
325
 
326
  def validate_text(text: str) -> bool:
327
  """Check if text is valid for classification."""
328
+ return bool(text and len(text) >= MIN_TEXT_LENGTH)
 
 
329
 
330
 
331
  def prepare_prompts(
332
+ texts: List[str], labels: List[str], tokenizer: AutoTokenizer,
333
+ label_descriptions: dict = None, enable_reasoning: bool = False
 
334
  ) -> tuple[List[str], List[int]]:
335
+ """Prepare prompts using chat template for classification, filtering invalid texts."""
336
  prompts = []
337
  valid_indices = []
338
+
339
  for i, text in enumerate(texts):
340
  processed_text = preprocess_text(text)
341
  if validate_text(processed_text):
342
+ # Create messages for chat template
343
+ messages = create_messages(processed_text, labels, label_descriptions, enable_reasoning)
344
+
345
+ # Apply chat template
346
+ prompt = tokenizer.apply_chat_template(
347
+ messages,
348
+ tokenize=False,
349
+ add_generation_prompt=True
350
  )
351
  prompts.append(prompt)
352
  valid_indices.append(i)
353
+
354
  return prompts, valid_indices
355
 
356
 
357
  def main():
358
  args = parse_args()
359
+
360
  # Check authentication early
361
  logger.info("Checking authentication...")
362
+ token = args.hf_token or (os.environ.get("HF_TOKEN") or get_token())
363
+
 
 
 
364
  if not token:
365
  logger.error("No authentication token found. Please either:")
366
  logger.error("1. Run 'huggingface-cli login'")
367
  logger.error("2. Set HF_TOKEN environment variable")
368
  logger.error("3. Pass --hf-token argument")
369
  sys.exit(1)
370
+
371
  # Validate token by checking who we are
372
  try:
373
  api = HfApi(token=token)
 
377
  logger.error(f"Authentication failed: {e}")
378
  logger.error("Please check your token is valid")
379
  sys.exit(1)
380
+
381
  # Check CUDA availability
382
  if not torch.cuda.is_available():
383
  logger.error("CUDA is not available. This script requires a GPU.")
384
  logger.error("Please run on a machine with GPU support or use HF Jobs.")
385
  sys.exit(1)
386
+
387
  logger.info(f"CUDA available. Using device: {torch.cuda.get_device_name(0)}")
388
+
389
  # Parse and validate labels
390
  labels = [label.strip() for label in args.labels.split(",")]
391
  if len(labels) < 2:
 
393
  sys.exit(1)
394
  logger.info(f"Classification labels: {labels}")
395
 
396
+ # Parse label descriptions if provided
397
+ label_descriptions = None
398
+ if args.label_descriptions:
399
+ label_descriptions = parse_label_descriptions(args.label_descriptions)
400
+ logger.info("Label descriptions provided:")
401
+ for label, desc in label_descriptions.items():
402
+ logger.info(f" {label}: {desc}")
403
+
404
  # Load dataset
405
  logger.info(f"Loading dataset: {args.input_dataset}")
406
  try:
407
  dataset = load_dataset(args.input_dataset, split=args.split)
408
+ logger.info(f"Loaded {len(dataset)} samples from split '{args.split}'")
409
+
410
+ # Shuffle if requested
411
+ if args.shuffle:
412
+ logger.info(f"Shuffling dataset with seed {args.shuffle_seed}")
413
+ dataset = dataset.shuffle(seed=args.shuffle_seed)
414
+
415
  # Limit samples if specified
416
  if args.max_samples:
417
  dataset = dataset.select(range(min(args.max_samples, len(dataset))))
418
  logger.info(f"Limited dataset to {len(dataset)} samples")
419
+ if args.shuffle:
420
+ logger.info("Note: Samples were randomly selected due to shuffling")
421
  except Exception as e:
422
  logger.error(f"Failed to load dataset: {e}")
423
  sys.exit(1)
424
+
425
  # Verify column exists
426
  if args.column not in dataset.column_names:
427
  logger.error(f"Column '{args.column}' not found in dataset.")
428
  logger.error(f"Available columns: {dataset.column_names}")
429
  sys.exit(1)
430
+
431
  # Extract texts
432
  texts = dataset[args.column]
433
+
434
+ # Load tokenizer for chat template formatting
435
+ logger.info(f"Loading tokenizer for {args.model}")
436
+ try:
437
+ tokenizer = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
438
+ except Exception as e:
439
+ logger.error(f"Failed to load tokenizer: {e}")
440
+ sys.exit(1)
441
+
442
  # Initialize vLLM
443
  logger.info(f"Initializing vLLM with model: {args.model}")
444
  logger.info(f"Using guided decoding backend: {args.guided_backend}")
 
453
  except Exception as e:
454
  logger.error(f"Failed to initialize vLLM: {e}")
455
  sys.exit(1)
456
+
457
+ # Set up sampling parameters based on mode
458
+ if args.enable_reasoning:
459
+ # Reasoning mode: no guided decoding, much more tokens for thinking
460
+ sampling_params = SamplingParams(
461
+ temperature=args.temperature,
462
+ max_tokens=args.max_tokens * 20, # 20x more tokens for extensive reasoning
463
+ )
464
+ logger.info("Using reasoning mode - model will generate thinking traces with JSON output")
465
+ else:
466
+ # Structured output mode: guided decoding
467
+ guided_params = GuidedDecodingParams(choice=labels)
468
+ sampling_params = SamplingParams(
469
+ guided_decoding=guided_params,
470
+ temperature=args.temperature,
471
+ max_tokens=args.max_tokens,
472
+ )
473
+ logger.info("Using structured output with guided_choice - outputs guaranteed to be valid labels")
474
+
475
  # Prepare all prompts
476
  logger.info("Preparing prompts for classification...")
477
+ all_prompts, valid_indices = prepare_prompts(texts, labels, tokenizer, label_descriptions, args.enable_reasoning)
478
+
479
  if not all_prompts:
480
  logger.error("No valid texts found for classification.")
481
  sys.exit(1)
482
+
483
  logger.info(f"Prepared {len(all_prompts)} valid prompts out of {len(texts)} texts")
484
+
485
  # Let vLLM handle batching internally
486
  logger.info("Starting classification (vLLM will handle batching internally)...")
487
+
488
  try:
489
  # Generate all classifications at once - vLLM handles batching
490
  outputs = llm.generate(all_prompts, sampling_params)
491
+
492
+ # Process outputs based on mode
493
+ if args.enable_reasoning:
494
+ # Reasoning mode: parse JSON and extract reasoning
495
+ all_classifications = [None] * len(texts)
496
+ all_reasoning = [None] * len(texts)
497
+ all_parsing_success = [False] * len(texts)
498
+
499
+ for idx, output in enumerate(outputs):
500
+ original_idx = valid_indices[idx]
501
+ generated_text = output.outputs[0].text.strip()
502
+
503
+ # Parse the reasoning output
504
+ label, reasoning, success = parse_reasoning_output(generated_text, labels)
505
+
506
+ all_classifications[original_idx] = label
507
+ all_reasoning[original_idx] = reasoning
508
+ all_parsing_success[original_idx] = success
509
+
510
+ # Log first few examples
511
+ if idx < 3:
512
+ logger.info(f"\nExample {idx + 1} output:")
513
+ logger.info(f"Raw output: {generated_text[:200]}...")
514
+ logger.info(f"Parsed label: {label}")
515
+ logger.info(f"Parsing success: {success}")
516
+
517
+ # Count parsing statistics
518
+ parsing_success_count = sum(1 for s in all_parsing_success if s)
519
+ parsing_fail_count = sum(1 for s in all_parsing_success if s is not None and not s)
520
+ logger.info(f"\nParsing statistics:")
521
+ logger.info(f" Successful: {parsing_success_count}/{len(valid_indices)} ({parsing_success_count/len(valid_indices)*100:.1f}%)")
522
+ logger.info(f" Failed: {parsing_fail_count}/{len(valid_indices)} ({parsing_fail_count/len(valid_indices)*100:.1f}%)")
523
+
524
+ valid_texts = parsing_success_count
525
+ else:
526
+ # Structured output mode: direct classification
527
+ all_classifications = [None] * len(texts)
528
+ for idx, output in enumerate(outputs):
529
+ original_idx = valid_indices[idx]
530
+ generated_text = output.outputs[0].text.strip()
531
+ all_classifications[original_idx] = generated_text
532
+
533
+ valid_texts = len(valid_indices)
534
+
535
  # Count statistics
 
536
  total_texts = len(texts)
537
+
538
  except Exception as e:
539
  logger.error(f"Classification failed: {e}")
540
  sys.exit(1)
541
+
542
+ # Add columns to dataset
543
  dataset = dataset.add_column("classification", all_classifications)
544
 
545
+ if args.enable_reasoning:
546
+ dataset = dataset.add_column("reasoning", all_reasoning)
547
+ dataset = dataset.add_column("parsing_success", all_parsing_success)
548
+
549
  # Calculate statistics
550
  none_count = total_texts - valid_texts
551
  if none_count > 0:
552
+ logger.warning(
553
+ f"{none_count} texts were too short or invalid for classification"
554
+ )
555
+
556
  # Show classification distribution
557
  label_counts = {label: all_classifications.count(label) for label in labels}
558
+
559
+ # Count None values separately
560
+ none_classifications = all_classifications.count(None)
561
+
562
  logger.info("Classification distribution:")
563
  for label, count in label_counts.items():
564
  percentage = count / total_texts * 100 if total_texts > 0 else 0
565
  logger.info(f" {label}: {count} ({percentage:.1f}%)")
 
 
 
566
 
567
+ if none_classifications > 0:
568
+ none_percentage = none_classifications / total_texts * 100
569
+ if args.enable_reasoning:
570
+ logger.info(f" Failed to parse: {none_classifications} ({none_percentage:.1f}%)")
571
+ else:
572
+ logger.info(f" Invalid/Skipped: {none_classifications} ({none_percentage:.1f}%)")
573
+
574
  # Log success rate
575
  success_rate = (valid_texts / total_texts * 100) if total_texts > 0 else 0
576
  logger.info(f"Classification success rate: {success_rate:.1f}%")
577
+
578
  # Save to Hub (token already validated at start)
579
  logger.info(f"Pushing dataset to Hub: {args.output_dataset}")
580
  try:
581
  dataset.push_to_hub(
582
  args.output_dataset,
583
  token=token,
584
+ commit_message=f"Add classifications using {args.model} {'with reasoning' if args.enable_reasoning else 'with structured outputs'}",
585
+ )
586
+ logger.info(
587
+ f"Successfully pushed to: https://huggingface.co/datasets/{args.output_dataset}"
588
  )
 
589
  except Exception as e:
590
  logger.error(f"Failed to push to Hub: {e}")
591
  sys.exit(1)
 
593
 
594
  if __name__ == "__main__":
595
  if len(sys.argv) == 1:
596
+ print("Example commands:")
597
+ print("\n# Simple classification:")
598
+ print("uv run classify-dataset.py \\")
599
+ print(" --input-dataset stanfordnlp/imdb \\")
600
+ print(" --column text \\")
601
+ print(" --labels 'positive,negative' \\")
602
+ print(" --output-dataset user/imdb-classified")
603
+ print("\n# With label descriptions:")
604
+ print("uv run classify-dataset.py \\")
605
+ print(" --input-dataset user/support-tickets \\")
606
+ print(" --column content \\")
607
+ print(" --labels 'bug,feature,question' \\")
608
+ print(" --label-descriptions 'bug:something is broken or not working,feature:request for new functionality,question:asking for help or clarification' \\")
609
+ print(" --output-dataset user/tickets-classified")
610
+ print("\n# With reasoning mode (thinking + JSON output):")
611
+ print("uv run classify-dataset.py \\")
612
+ print(" --input-dataset stanfordnlp/imdb \\")
613
+ print(" --column text \\")
614
+ print(" --labels 'positive,negative,neutral' \\")
615
+ print(" --enable-reasoning \\")
616
+ print(" --output-dataset user/imdb-reasoned")
617
+ print("\n# HF Jobs example:")
618
  print("hf jobs uv run \\")
619
  print(" --flavor l4x1 \\")
620
  print(" --image vllm/vllm-openai:latest \\")
 
624
  print(" --labels 'positive,negative' \\")
625
  print(" --output-dataset user/imdb-classified")
626
  sys.exit(0)
627
+
628
+ main()