davanstrien HF Staff commited on
Commit
011bc8a
Β·
1 Parent(s): 1d07414
Files changed (1) hide show
  1. README.md +140 -55
README.md CHANGED
@@ -3,9 +3,9 @@ viewer: false
3
  tags: [uv-script, classification, vllm, structured-outputs, gpu-required]
4
  ---
5
 
6
- # Dataset Classification with vLLM
7
 
8
- Efficient text classification for Hugging Face datasets using vLLM with structured outputs. This script provides GPU-accelerated classification with guaranteed valid outputs through guided decoding.
9
 
10
  ## πŸš€ Quick Start
11
 
@@ -22,21 +22,25 @@ That's it! No installation, no setup - just `uv run`.
22
 
23
  ## πŸ“‹ Requirements
24
 
25
- - **GPU Required**: This script uses vLLM for efficient inference
26
  - Python 3.10+
27
  - UV (will handle all dependencies automatically)
28
- - vLLM >= 0.6.6 (for guided decoding support)
29
 
30
  ## 🎯 Features
31
 
32
- - **Guaranteed valid outputs** using vLLM's guided decoding with outlines
33
- - **Zero-shot classification** with structured generation
34
- - **GPU-optimized** with vLLM's automatic batching for maximum efficiency
35
- - **Default model**: HuggingFaceTB/SmolLM3-3B (fast 3B model, easily changeable)
36
  - **Robust text handling** with preprocessing and validation
37
- - **Three prompt styles** for different use cases
38
  - **Automatic progress tracking** and detailed statistics
39
  - **Direct Hub integration** - read and write datasets seamlessly
 
 
 
 
 
40
 
41
  ## πŸ’» Usage
42
 
@@ -62,20 +66,50 @@ uv run classify-dataset.py \
62
  **Optional:**
63
 
64
  - `--model`: Model to use (default: **`HuggingFaceTB/SmolLM3-3B`** - a fast 3B parameter model)
65
- - `--prompt-style`: Choose from `simple`, `detailed`, or `reasoning` (default: `simple`)
 
66
  - `--split`: Dataset split to process (default: `train`)
67
  - `--max-samples`: Limit samples for testing
 
 
68
  - `--temperature`: Generation temperature (default: 0.1)
69
  - `--guided-backend`: Backend for guided decoding (default: `outlines`)
70
  - `--hf-token`: Hugging Face token (or use `HF_TOKEN` env var)
71
 
72
- ### Prompt Styles
73
 
74
- - **simple**: Direct classification prompt
75
- - **detailed**: Emphasizes exact category matching
76
- - **reasoning**: Includes brief analysis before classification
77
 
78
- All styles benefit from structured output guarantees - the model can only output valid labels!
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
79
 
80
  ## πŸ“Š Examples
81
 
@@ -96,8 +130,8 @@ uv run classify-dataset.py \
96
  --input-dataset user/support-tickets \
97
  --column content \
98
  --labels "bug,feature_request,question,other" \
99
- --output-dataset user/tickets-classified \
100
- --prompt-style reasoning
101
  ```
102
 
103
  ### News Categorization
@@ -111,11 +145,54 @@ uv run classify-dataset.py \
111
  --model meta-llama/Llama-3.2-3B-Instruct
112
  ```
113
 
114
- ## πŸš€ Running on HF Jobs
 
 
 
 
 
 
 
 
 
 
115
 
116
- This script is optimized for [Hugging Face Jobs](https://huggingface.co/docs/hub/spaces-gpu-jobs) (requires Pro subscription or Team/Enterprise organization):
117
 
118
- ````bash
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
119
  # Run on L4 GPU with vLLM image
120
  hf jobs uv run \
121
  --flavor l4x1 \
@@ -125,6 +202,7 @@ hf jobs uv run \
125
  --column text \
126
  --labels "positive,negative" \
127
  --output-dataset user/imdb-classified
 
128
 
129
  ### GPU Flavors
130
  - `t4-small`: Budget option for smaller models
@@ -135,6 +213,27 @@ hf jobs uv run \
135
 
136
  ## πŸ”§ Advanced Usage
137
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
138
  ### Using Different Models
139
 
140
  By default, this script uses **HuggingFaceTB/SmolLM3-3B** - a fast, efficient 3B parameter model that's perfect for most classification tasks. You can easily use any other instruction-tuned model:
@@ -147,7 +246,7 @@ uv run classify-dataset.py \
147
  --labels "contract,patent,brief,memo,other" \
148
  --output-dataset user/legal-classified \
149
  --model Qwen/Qwen2.5-7B-Instruct
150
- ````
151
 
152
  ### Large Datasets
153
 
@@ -166,15 +265,16 @@ uv run classify-dataset.py \
166
  - **SmolLM3-3B (default)**: ~50-100 texts/second on A10
167
  - **7B models**: ~20-50 texts/second on A10
168
  - vLLM automatically optimizes batching for best throughput
 
169
 
170
  ## 🀝 How It Works
171
 
172
- 1. **vLLM**: Provides efficient GPU batch inference
173
- 2. **Guided Decoding**: Uses outlines to guarantee valid label outputs
174
  3. **Structured Generation**: Constrains model outputs to exact label choices
175
  4. **UV**: Handles all dependencies automatically
176
 
177
- The script loads your dataset, preprocesses texts, classifies each one using guided decoding to ensure only valid labels are generated, then saves the results as a new column in the output dataset.
178
 
179
  ## πŸ› Troubleshooting
180
 
@@ -212,43 +312,28 @@ If you see `ImportError: cannot import name 'GuidedDecodingParams'`:
212
  - The script specifies the correct version in its dependencies
213
  - UV should automatically install the correct version
214
 
215
- ## πŸ”¬ Advanced Example: ArXiv ML Trends Analysis
216
 
217
- For a more complex real-world example, we provide scripts to analyze ML research trends from ArXiv papers:
218
 
219
- ### Step 1: Prepare the Dataset
220
-
221
- ```bash
222
- # Filter and prepare ArXiv CS papers from 2024
223
- uv run prepare_arxiv_2024.py
224
- ```
225
 
226
- This creates a filtered dataset of CS papers with combined title+abstract text.
 
 
227
 
228
- ### Step 2: Run Classification with Python API
229
-
230
- ```bash
231
- # Use HF Jobs Python API to classify papers
232
- uv run run_arxiv_classification.py
 
 
233
  ```
234
 
235
- This script demonstrates:
236
-
237
- - Using `run_uv_job()` from the Python API
238
- - Classifying into modern ML trends (reasoning, agents, multimodal, robotics, etc.)
239
- - Handling authentication and job monitoring
240
-
241
- The classification categories include:
242
-
243
- - `reasoning_systems`: Chain-of-thought, reasoning, problem solving
244
- - `agents_autonomous`: Agents, tool use, autonomous systems
245
- - `multimodal_models`: Vision-language, audio, multi-modal
246
- - `robotics_embodied`: Robotics, embodied AI, manipulation
247
- - `efficient_inference`: Quantization, distillation, edge deployment
248
- - `alignment_safety`: RLHF, alignment, safety, interpretability
249
- - `generative_models`: Diffusion, generation, synthesis
250
- - `foundational_other`: Other foundational ML/AI research
251
-
252
  ## πŸ“ License
253
 
254
  This script is provided as-is for use with the UV Scripts organization.
 
3
  tags: [uv-script, classification, vllm, structured-outputs, gpu-required]
4
  ---
5
 
6
+ # Dataset Classification Script
7
 
8
+ GPU-accelerated text classification for Hugging Face datasets with guaranteed valid outputs through structured generation.
9
 
10
  ## πŸš€ Quick Start
11
 
 
22
 
23
  ## πŸ“‹ Requirements
24
 
25
+ - **GPU Required**: Uses GPU-accelerated inference
26
  - Python 3.10+
27
  - UV (will handle all dependencies automatically)
28
+ - vLLM >= 0.6.6
29
 
30
  ## 🎯 Features
31
 
32
+ - **Guaranteed valid outputs** using structured generation with guided decoding
33
+ - **Zero-shot classification** without training data required
34
+ - **GPU-optimized** for maximum throughput and efficiency
35
+ - **Default model**: HuggingFaceTB/SmolLM3-3B (fast 3B model with thinking capabilities)
36
  - **Robust text handling** with preprocessing and validation
 
37
  - **Automatic progress tracking** and detailed statistics
38
  - **Direct Hub integration** - read and write datasets seamlessly
39
+ - **Label descriptions** support for providing context to improve accuracy
40
+ - **Reasoning mode** for interpretable classifications with thinking traces
41
+ - **JSON output parsing** for reliable extraction from reasoning mode
42
+ - **Optimized batching** with vLLM's automatic batch processing
43
+ - **Multiple guided backends** - supports outlines, xgrammar, and more
44
 
45
  ## πŸ’» Usage
46
 
 
66
  **Optional:**
67
 
68
  - `--model`: Model to use (default: **`HuggingFaceTB/SmolLM3-3B`** - a fast 3B parameter model)
69
+ - `--label-descriptions`: Provide descriptions for each label to improve classification accuracy
70
+ - `--enable-reasoning`: Enable reasoning mode with thinking traces (adds reasoning column)
71
  - `--split`: Dataset split to process (default: `train`)
72
  - `--max-samples`: Limit samples for testing
73
+ - `--shuffle`: Shuffle dataset before selecting samples (useful for random sampling)
74
+ - `--shuffle-seed`: Random seed for shuffling (default: 42)
75
  - `--temperature`: Generation temperature (default: 0.1)
76
  - `--guided-backend`: Backend for guided decoding (default: `outlines`)
77
  - `--hf-token`: Hugging Face token (or use `HF_TOKEN` env var)
78
 
79
+ ### Label Descriptions
80
 
81
+ Provide context for your labels to improve classification accuracy:
 
 
82
 
83
+ ```bash
84
+ uv run classify-dataset.py \
85
+ --input-dataset user/support-tickets \
86
+ --column content \
87
+ --labels "bug,feature,question,other" \
88
+ --label-descriptions "bug:something is broken,feature:request for new functionality,question:asking for help,other:anything else" \
89
+ --output-dataset user/tickets-classified
90
+ ```
91
+
92
+ The model uses these descriptions to better understand what each label represents, leading to more accurate classifications.
93
+
94
+ ### Reasoning Mode
95
+
96
+ Enable thinking traces for interpretable classifications:
97
+
98
+ ```bash
99
+ uv run classify-dataset.py \
100
+ --input-dataset stanfordnlp/imdb \
101
+ --column text \
102
+ --labels "positive,negative,neutral" \
103
+ --enable-reasoning \
104
+ --output-dataset user/imdb-with-reasoning
105
+ ```
106
+
107
+ When `--enable-reasoning` is used:
108
+ - The model generates step-by-step reasoning using SmolLM3's thinking capabilities
109
+ - Output includes three columns: `classification`, `reasoning`, and `parsing_success`
110
+ - Final answer must be in JSON format: `{"label": "chosen_label"}`
111
+ - Useful for understanding complex classification decisions
112
+ - Trade-off: Slower but more interpretable
113
 
114
  ## πŸ“Š Examples
115
 
 
130
  --input-dataset user/support-tickets \
131
  --column content \
132
  --labels "bug,feature_request,question,other" \
133
+ --label-descriptions "bug:code or product not working as expected,feature_request:asking for new functionality,question:seeking help or clarification,other:general comments or feedback" \
134
+ --output-dataset user/tickets-classified
135
  ```
136
 
137
  ### News Categorization
 
145
  --model meta-llama/Llama-3.2-3B-Instruct
146
  ```
147
 
148
+ ### Complex Classification with Reasoning
149
+
150
+ ```bash
151
+ uv run classify-dataset.py \
152
+ --input-dataset user/customer-feedback \
153
+ --column text \
154
+ --labels "very_positive,positive,neutral,negative,very_negative" \
155
+ --label-descriptions "very_positive:extremely satisfied,positive:generally satisfied,neutral:mixed feelings,negative:dissatisfied,very_negative:extremely dissatisfied" \
156
+ --enable-reasoning \
157
+ --output-dataset user/feedback-analyzed
158
+ ```
159
 
160
+ This combines label descriptions with reasoning mode for maximum interpretability.
161
 
162
+ ### ArXiv ML Research Classification
163
+
164
+ Classify academic papers into machine learning research areas:
165
+
166
+ ```bash
167
+ # Fast classification with random sampling
168
+ uv run classify-dataset.py \
169
+ --input-dataset librarian-bots/arxiv-metadata-snapshot \
170
+ --column abstract \
171
+ --labels "llm,computer_vision,reinforcement_learning,optimization,theory,other" \
172
+ --label-descriptions "llm:language models and NLP,computer_vision:image and video processing,reinforcement_learning:RL and decision making,optimization:training and efficiency,theory:theoretical ML foundations,other:other ML topics" \
173
+ --output-dataset user/arxiv-ml-classified \
174
+ --split "train[:10000]" \
175
+ --max-samples 100 \
176
+ --shuffle
177
+
178
+ # With reasoning for nuanced classification
179
+ uv run classify-dataset.py \
180
+ --input-dataset librarian-bots/arxiv-metadata-snapshot \
181
+ --column abstract \
182
+ --labels "multimodal,agents,reasoning,safety,efficiency" \
183
+ --label-descriptions "multimodal:vision-language and cross-modal models,agents:autonomous agents and tool use,reasoning:reasoning and planning systems,safety:alignment and safety research,efficiency:model optimization and deployment" \
184
+ --enable-reasoning \
185
+ --output-dataset user/arxiv-frontier-research \
186
+ --split "train[:1000]" \
187
+ --max-samples 50
188
+ ```
189
+
190
+ The reasoning mode is particularly valuable for academic abstracts where papers often span multiple topics and require careful analysis to determine the primary focus.
191
+
192
+ ## πŸš€ Running on HF Jobs
193
+
194
+ Optimized for [Hugging Face Jobs](https://huggingface.co/docs/hub/spaces-gpu-jobs) (requires Pro subscription or Team/Enterprise organization):
195
+ ```bash
196
  # Run on L4 GPU with vLLM image
197
  hf jobs uv run \
198
  --flavor l4x1 \
 
202
  --column text \
203
  --labels "positive,negative" \
204
  --output-dataset user/imdb-classified
205
+ ```
206
 
207
  ### GPU Flavors
208
  - `t4-small`: Budget option for smaller models
 
213
 
214
  ## πŸ”§ Advanced Usage
215
 
216
+ ### Random Sampling
217
+
218
+ When working with ordered datasets, use `--shuffle` with `--max-samples` to get a representative sample:
219
+
220
+ ```bash
221
+ # Get 50 random reviews instead of the first 50
222
+ uv run classify-dataset.py \
223
+ --input-dataset stanfordnlp/imdb \
224
+ --column text \
225
+ --labels "positive,negative" \
226
+ --output-dataset user/imdb-sample \
227
+ --max-samples 50 \
228
+ --shuffle \
229
+ --shuffle-seed 123 # For reproducibility
230
+ ```
231
+
232
+ This is especially important for:
233
+ - Chronologically ordered datasets (news, papers, social media)
234
+ - Pre-sorted datasets (by rating, category, etc.)
235
+ - Testing on diverse samples before processing the full dataset
236
+
237
  ### Using Different Models
238
 
239
  By default, this script uses **HuggingFaceTB/SmolLM3-3B** - a fast, efficient 3B parameter model that's perfect for most classification tasks. You can easily use any other instruction-tuned model:
 
246
  --labels "contract,patent,brief,memo,other" \
247
  --output-dataset user/legal-classified \
248
  --model Qwen/Qwen2.5-7B-Instruct
249
+ ```
250
 
251
  ### Large Datasets
252
 
 
265
  - **SmolLM3-3B (default)**: ~50-100 texts/second on A10
266
  - **7B models**: ~20-50 texts/second on A10
267
  - vLLM automatically optimizes batching for best throughput
268
+ - Performance scales with GPU memory and compute capability
269
 
270
  ## 🀝 How It Works
271
 
272
+ 1. **vLLM**: Provides efficient GPU batch inference with automatic batching
273
+ 2. **Guided Decoding**: Uses outlines backend to guarantee valid label outputs
274
  3. **Structured Generation**: Constrains model outputs to exact label choices
275
  4. **UV**: Handles all dependencies automatically
276
 
277
+ The script loads your dataset, preprocesses texts, classifies each one with guaranteed valid outputs, then saves the results as a new column in the output dataset.
278
 
279
  ## πŸ› Troubleshooting
280
 
 
312
  - The script specifies the correct version in its dependencies
313
  - UV should automatically install the correct version
314
 
315
+ ## πŸ”¬ Advanced Workflows
316
 
317
+ For complex real-world workflows that integrate UV scripts with the Python HF Jobs API, see the [ArXiv ML Trends example](examples/arxiv-workflow/). This demonstrates:
318
 
319
+ - **Multi-stage pipelines**: Data preparation β†’ GPU classification β†’ Analysis
320
+ - **Python API orchestration**: Using `run_uv_job()` to manage GPU jobs programmatically
321
+ - **Production patterns**: Error handling, parallel execution, and incremental updates
322
+ - **Cost optimization**: Choosing appropriate compute resources for each task
 
 
323
 
324
+ ```python
325
+ # Example: Submit a classification job via Python API
326
+ from huggingface_hub import run_uv_job
327
 
328
+ job = run_uv_job(
329
+ script="https://huggingface.co/datasets/uv-scripts/classification/raw/main/classify-dataset.py",
330
+ args=["--input-dataset", "my/dataset", "--labels", "A,B,C"],
331
+ flavor="l4x1",
332
+ image="vllm/vllm-openai:latest"
333
+ )
334
+ result = job.wait()
335
  ```
336
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
337
  ## πŸ“ License
338
 
339
  This script is provided as-is for use with the UV Scripts organization.