Commit
·
b366864
1
Parent(s):
3f59035
Add NuMarkdown-8B-Thinking OCR script with reasoning capabilities
Browse files- numarkdown-ocr.py +627 -0
numarkdown-ocr.py
ADDED
@@ -0,0 +1,627 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# /// script
|
2 |
+
# requires-python = ">=3.11"
|
3 |
+
# dependencies = [
|
4 |
+
# "datasets",
|
5 |
+
# "huggingface-hub[hf_transfer]",
|
6 |
+
# "pillow",
|
7 |
+
# "vllm",
|
8 |
+
# "tqdm",
|
9 |
+
# "toolz",
|
10 |
+
# "torch", # Added for CUDA check
|
11 |
+
# ]
|
12 |
+
#
|
13 |
+
# ///
|
14 |
+
|
15 |
+
"""
|
16 |
+
Convert document images to markdown using NuMarkdown-8B-Thinking with vLLM.
|
17 |
+
|
18 |
+
This script processes images through the NuMarkdown model to extract
|
19 |
+
text with advanced reasoning capabilities, ideal for complex document understanding.
|
20 |
+
|
21 |
+
Features:
|
22 |
+
- Reasoning-based document analysis with thinking tokens
|
23 |
+
- Superior table extraction and formatting
|
24 |
+
- Complex layout understanding
|
25 |
+
- Mathematical formula recognition
|
26 |
+
- Clean markdown output generation
|
27 |
+
- Optional thinking trace inclusion
|
28 |
+
"""
|
29 |
+
|
30 |
+
import argparse
|
31 |
+
import base64
|
32 |
+
import io
|
33 |
+
import json
|
34 |
+
import logging
|
35 |
+
import os
|
36 |
+
import re
|
37 |
+
import sys
|
38 |
+
from typing import Any, Dict, List, Union, Optional, Tuple
|
39 |
+
from datetime import datetime
|
40 |
+
|
41 |
+
import torch
|
42 |
+
from datasets import load_dataset
|
43 |
+
from huggingface_hub import DatasetCard, login
|
44 |
+
from PIL import Image
|
45 |
+
from toolz import partition_all
|
46 |
+
from tqdm.auto import tqdm
|
47 |
+
from vllm import LLM, SamplingParams
|
48 |
+
|
49 |
+
logging.basicConfig(level=logging.INFO)
|
50 |
+
logger = logging.getLogger(__name__)
|
51 |
+
|
52 |
+
|
53 |
+
def check_cuda_availability():
|
54 |
+
"""Check if CUDA is available and exit if not."""
|
55 |
+
if not torch.cuda.is_available():
|
56 |
+
logger.error("CUDA is not available. This script requires a GPU.")
|
57 |
+
logger.error("Please run on a machine with a CUDA-capable GPU.")
|
58 |
+
sys.exit(1)
|
59 |
+
else:
|
60 |
+
logger.info(f"CUDA is available. GPU: {torch.cuda.get_device_name(0)}")
|
61 |
+
|
62 |
+
|
63 |
+
def validate_and_resize_image(
|
64 |
+
image: Image.Image,
|
65 |
+
min_pixels: int = 100 * 28 * 28,
|
66 |
+
max_pixels: int = 5000 * 28 * 28,
|
67 |
+
) -> Image.Image:
|
68 |
+
"""Validate and resize image to meet pixel constraints if necessary."""
|
69 |
+
width, height = image.size
|
70 |
+
total_pixels = width * height
|
71 |
+
|
72 |
+
if total_pixels < min_pixels or total_pixels > max_pixels:
|
73 |
+
# Calculate scaling factor
|
74 |
+
if total_pixels < min_pixels:
|
75 |
+
scale = (min_pixels / total_pixels) ** 0.5
|
76 |
+
else:
|
77 |
+
scale = (max_pixels / total_pixels) ** 0.5
|
78 |
+
|
79 |
+
new_width = int(width * scale)
|
80 |
+
new_height = int(height * scale)
|
81 |
+
|
82 |
+
logger.debug(f"Resizing image from {width}x{height} to {new_width}x{new_height}")
|
83 |
+
image = image.resize((new_width, new_height), Image.Resampling.LANCZOS)
|
84 |
+
|
85 |
+
return image
|
86 |
+
|
87 |
+
|
88 |
+
def extract_answer_from_thinking(text: str, include_thinking: bool = False) -> str:
|
89 |
+
"""
|
90 |
+
Extract the final answer from NuMarkdown's thinking output.
|
91 |
+
|
92 |
+
The model generates output in format:
|
93 |
+
<think>reasoning process...</think>
|
94 |
+
<answer>final markdown output</answer>
|
95 |
+
"""
|
96 |
+
if include_thinking:
|
97 |
+
# Return the full output including thinking traces
|
98 |
+
return text.strip()
|
99 |
+
|
100 |
+
# Extract content between <answer> tags
|
101 |
+
answer_pattern = r'<answer>(.*?)</answer>'
|
102 |
+
answer_match = re.search(answer_pattern, text, re.DOTALL)
|
103 |
+
|
104 |
+
if answer_match:
|
105 |
+
return answer_match.group(1).strip()
|
106 |
+
|
107 |
+
# If no answer tags found, check if the entire text is markdown
|
108 |
+
# (sometimes the model might not use tags)
|
109 |
+
if not '<think>' in text and not '<answer>' in text:
|
110 |
+
return text.strip()
|
111 |
+
|
112 |
+
# Fallback: return everything after </think> if present
|
113 |
+
think_end = text.find('</think>')
|
114 |
+
if think_end != -1:
|
115 |
+
remaining = text[think_end + 8:].strip()
|
116 |
+
# Remove <answer> tags if present
|
117 |
+
remaining = remaining.replace('<answer>', '').replace('</answer>', '').strip()
|
118 |
+
return remaining
|
119 |
+
|
120 |
+
# Last resort: return the full text
|
121 |
+
logger.warning("Could not extract answer from thinking tokens, returning full text")
|
122 |
+
return text.strip()
|
123 |
+
|
124 |
+
|
125 |
+
def make_numarkdown_message(
|
126 |
+
image: Union[Image.Image, Dict[str, Any], str],
|
127 |
+
prompt: str = "Convert this document to markdown. Focus on preserving structure, tables, formulas, and all textual content.",
|
128 |
+
) -> List[Dict]:
|
129 |
+
"""Create chat message for NuMarkdown processing."""
|
130 |
+
# Convert to PIL Image if needed
|
131 |
+
if isinstance(image, Image.Image):
|
132 |
+
pil_img = image.convert("RGB")
|
133 |
+
elif isinstance(image, dict) and "bytes" in image:
|
134 |
+
pil_img = Image.open(io.BytesIO(image["bytes"])).convert("RGB")
|
135 |
+
elif isinstance(image, str):
|
136 |
+
pil_img = Image.open(image).convert("RGB")
|
137 |
+
else:
|
138 |
+
raise ValueError(f"Unsupported image type: {type(image)}")
|
139 |
+
|
140 |
+
# Validate and resize if necessary
|
141 |
+
pil_img = validate_and_resize_image(pil_img)
|
142 |
+
|
143 |
+
# Convert to base64 data URI
|
144 |
+
buf = io.BytesIO()
|
145 |
+
pil_img.save(buf, format="PNG")
|
146 |
+
data_uri = f"data:image/png;base64,{base64.b64encode(buf.getvalue()).decode()}"
|
147 |
+
|
148 |
+
# Return message in vLLM chat format
|
149 |
+
return [
|
150 |
+
{
|
151 |
+
"role": "user",
|
152 |
+
"content": [
|
153 |
+
{"type": "image_url", "image_url": {"url": data_uri}},
|
154 |
+
{"type": "text", "text": prompt},
|
155 |
+
],
|
156 |
+
}
|
157 |
+
]
|
158 |
+
|
159 |
+
|
160 |
+
def create_dataset_card(
|
161 |
+
source_dataset: str,
|
162 |
+
model: str,
|
163 |
+
num_samples: int,
|
164 |
+
processing_time: str,
|
165 |
+
batch_size: int,
|
166 |
+
max_model_len: int,
|
167 |
+
max_tokens: int,
|
168 |
+
gpu_memory_utilization: float,
|
169 |
+
include_thinking: bool,
|
170 |
+
image_column: str = "image",
|
171 |
+
split: str = "train",
|
172 |
+
) -> str:
|
173 |
+
"""Create a dataset card documenting the OCR process."""
|
174 |
+
model_name = model.split("/")[-1]
|
175 |
+
|
176 |
+
return f"""---
|
177 |
+
tags:
|
178 |
+
- ocr
|
179 |
+
- document-processing
|
180 |
+
- numarkdown
|
181 |
+
- markdown
|
182 |
+
- reasoning
|
183 |
+
- thinking-tokens
|
184 |
+
- uv-script
|
185 |
+
- generated
|
186 |
+
---
|
187 |
+
|
188 |
+
# Document OCR using {model_name}
|
189 |
+
|
190 |
+
This dataset contains markdown-formatted OCR results from images in [{source_dataset}](https://huggingface.co/datasets/{source_dataset}) using NuMarkdown-8B-Thinking.
|
191 |
+
|
192 |
+
## Processing Details
|
193 |
+
|
194 |
+
- **Source Dataset**: [{source_dataset}](https://huggingface.co/datasets/{source_dataset})
|
195 |
+
- **Model**: [{model}](https://huggingface.co/{model})
|
196 |
+
- **Number of Samples**: {num_samples:,}
|
197 |
+
- **Processing Time**: {processing_time}
|
198 |
+
- **Processing Date**: {datetime.now().strftime("%Y-%m-%d %H:%M UTC")}
|
199 |
+
|
200 |
+
### Configuration
|
201 |
+
|
202 |
+
- **Image Column**: `{image_column}`
|
203 |
+
- **Output Column**: `markdown`
|
204 |
+
- **Dataset Split**: `{split}`
|
205 |
+
- **Batch Size**: {batch_size}
|
206 |
+
- **Max Model Length**: {max_model_len:,} tokens
|
207 |
+
- **Max Output Tokens**: {max_tokens:,}
|
208 |
+
- **GPU Memory Utilization**: {gpu_memory_utilization:.1%}
|
209 |
+
- **Thinking Traces**: {"Included" if include_thinking else "Excluded (only final answers)"}
|
210 |
+
|
211 |
+
## Model Information
|
212 |
+
|
213 |
+
NuMarkdown-8B-Thinking is a state-of-the-art reasoning-based document OCR model that excels at:
|
214 |
+
- 🧠 **Reasoning Process** - Analyzes document layout before generation
|
215 |
+
- 📊 **Complex Tables** - Superior table extraction and formatting
|
216 |
+
- 📐 **Mathematical Formulas** - Accurate LaTeX/math notation preservation
|
217 |
+
- 📝 **Document Structure** - Maintains hierarchical document organization
|
218 |
+
- 🔍 **Layout Analysis** - Understands complex multi-column layouts
|
219 |
+
- ✨ **Clean Output** - Generates well-formatted markdown
|
220 |
+
|
221 |
+
### Thinking Tokens
|
222 |
+
|
223 |
+
This model uses a unique "thinking" process where it:
|
224 |
+
1. Analyzes the document structure internally (`<think>` phase)
|
225 |
+
2. Generates the final markdown output (`<answer>` phase)
|
226 |
+
|
227 |
+
{"The dataset includes both thinking traces and final answers." if include_thinking else "Only the final answers are included (thinking traces removed)."}
|
228 |
+
|
229 |
+
## Dataset Structure
|
230 |
+
|
231 |
+
The dataset contains all original columns plus:
|
232 |
+
- `markdown`: The extracted text in markdown format
|
233 |
+
- `inference_info`: JSON list tracking all OCR models applied to this dataset
|
234 |
+
|
235 |
+
## Usage
|
236 |
+
|
237 |
+
```python
|
238 |
+
from datasets import load_dataset
|
239 |
+
import json
|
240 |
+
|
241 |
+
# Load the dataset
|
242 |
+
dataset = load_dataset("{{output_dataset_id}}", split="{split}")
|
243 |
+
|
244 |
+
# Access the markdown text
|
245 |
+
for example in dataset:
|
246 |
+
print(example["markdown"])
|
247 |
+
break
|
248 |
+
|
249 |
+
# View all OCR models applied to this dataset
|
250 |
+
inference_info = json.loads(dataset[0]["inference_info"])
|
251 |
+
for info in inference_info:
|
252 |
+
print(f"Column: {{info['column_name']}} - Model: {{info['model_id']}}")
|
253 |
+
```
|
254 |
+
|
255 |
+
## Reproduction
|
256 |
+
|
257 |
+
This dataset was generated using the [uv-scripts/ocr](https://huggingface.co/datasets/uv-scripts/ocr) NuMarkdown OCR script:
|
258 |
+
|
259 |
+
```bash
|
260 |
+
uv run https://huggingface.co/datasets/uv-scripts/ocr/raw/main/numarkdown-ocr.py \\
|
261 |
+
{source_dataset} \\
|
262 |
+
<output-dataset> \\
|
263 |
+
--image-column {image_column} \\
|
264 |
+
--batch-size {batch_size} \\
|
265 |
+
--max-model-len {max_model_len} \\
|
266 |
+
--max-tokens {max_tokens} \\
|
267 |
+
--gpu-memory-utilization {gpu_memory_utilization} \\
|
268 |
+
{"--include-thinking" if include_thinking else ""}
|
269 |
+
```
|
270 |
+
|
271 |
+
## Performance
|
272 |
+
|
273 |
+
- **Processing Speed**: ~{num_samples / (float(processing_time.split()[0]) * 60):.1f} images/second
|
274 |
+
- **GPU Configuration**: vLLM with {gpu_memory_utilization:.0%} GPU memory utilization
|
275 |
+
- **Model Size**: 8.29B parameters
|
276 |
+
|
277 |
+
Generated with 🤖 [UV Scripts](https://huggingface.co/uv-scripts)
|
278 |
+
"""
|
279 |
+
|
280 |
+
|
281 |
+
def main(
|
282 |
+
input_dataset: str,
|
283 |
+
output_dataset: str,
|
284 |
+
image_column: str = "image",
|
285 |
+
batch_size: int = 16,
|
286 |
+
model: str = "numind/NuMarkdown-8B-Thinking",
|
287 |
+
max_model_len: int = 16384,
|
288 |
+
max_tokens: int = 8192,
|
289 |
+
gpu_memory_utilization: float = 0.9,
|
290 |
+
hf_token: str = None,
|
291 |
+
split: str = "train",
|
292 |
+
max_samples: int = None,
|
293 |
+
private: bool = False,
|
294 |
+
shuffle: bool = False,
|
295 |
+
seed: int = 42,
|
296 |
+
include_thinking: bool = False,
|
297 |
+
temperature: float = 0.0,
|
298 |
+
custom_prompt: Optional[str] = None,
|
299 |
+
):
|
300 |
+
"""Process images from HF dataset through NuMarkdown model."""
|
301 |
+
|
302 |
+
# Check CUDA availability first
|
303 |
+
check_cuda_availability()
|
304 |
+
|
305 |
+
# Track processing start time
|
306 |
+
start_time = datetime.now()
|
307 |
+
|
308 |
+
# Enable HF_TRANSFER for faster downloads
|
309 |
+
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
|
310 |
+
|
311 |
+
# Login to HF if token provided
|
312 |
+
HF_TOKEN = hf_token or os.environ.get("HF_TOKEN")
|
313 |
+
if HF_TOKEN:
|
314 |
+
login(token=HF_TOKEN)
|
315 |
+
|
316 |
+
# Load dataset
|
317 |
+
logger.info(f"Loading dataset: {input_dataset}")
|
318 |
+
dataset = load_dataset(input_dataset, split=split)
|
319 |
+
|
320 |
+
# Validate image column
|
321 |
+
if image_column not in dataset.column_names:
|
322 |
+
raise ValueError(
|
323 |
+
f"Column '{image_column}' not found. Available: {dataset.column_names}"
|
324 |
+
)
|
325 |
+
|
326 |
+
# Shuffle if requested
|
327 |
+
if shuffle:
|
328 |
+
logger.info(f"Shuffling dataset with seed {seed}")
|
329 |
+
dataset = dataset.shuffle(seed=seed)
|
330 |
+
|
331 |
+
# Limit samples if requested
|
332 |
+
if max_samples:
|
333 |
+
dataset = dataset.select(range(min(max_samples, len(dataset))))
|
334 |
+
logger.info(f"Limited to {len(dataset)} samples")
|
335 |
+
|
336 |
+
# Initialize vLLM with trust_remote_code for NuMarkdown
|
337 |
+
logger.info(f"Initializing vLLM with model: {model}")
|
338 |
+
llm = LLM(
|
339 |
+
model=model,
|
340 |
+
trust_remote_code=True, # Required for NuMarkdown
|
341 |
+
max_model_len=max_model_len,
|
342 |
+
gpu_memory_utilization=gpu_memory_utilization,
|
343 |
+
limit_mm_per_prompt={"image": 1},
|
344 |
+
)
|
345 |
+
|
346 |
+
# Set up sampling parameters
|
347 |
+
sampling_params = SamplingParams(
|
348 |
+
temperature=temperature,
|
349 |
+
max_tokens=max_tokens,
|
350 |
+
)
|
351 |
+
|
352 |
+
# Use custom prompt if provided, otherwise use default
|
353 |
+
prompt = custom_prompt or "Convert this document to markdown. Focus on preserving structure, tables, formulas, and all textual content."
|
354 |
+
|
355 |
+
# Process images in batches
|
356 |
+
all_markdown = []
|
357 |
+
|
358 |
+
logger.info(f"Processing {len(dataset)} images in batches of {batch_size}")
|
359 |
+
logger.info(f"Including thinking traces: {include_thinking}")
|
360 |
+
|
361 |
+
# Process in batches to avoid memory issues
|
362 |
+
for batch_indices in tqdm(
|
363 |
+
partition_all(batch_size, range(len(dataset))),
|
364 |
+
total=(len(dataset) + batch_size - 1) // batch_size,
|
365 |
+
desc="OCR processing",
|
366 |
+
):
|
367 |
+
batch_indices = list(batch_indices)
|
368 |
+
batch_images = [dataset[i][image_column] for i in batch_indices]
|
369 |
+
|
370 |
+
try:
|
371 |
+
# Create messages for batch
|
372 |
+
batch_messages = [
|
373 |
+
make_numarkdown_message(img, prompt) for img in batch_images
|
374 |
+
]
|
375 |
+
|
376 |
+
# Process with vLLM
|
377 |
+
outputs = llm.chat(batch_messages, sampling_params)
|
378 |
+
|
379 |
+
# Extract markdown from outputs
|
380 |
+
for output in outputs:
|
381 |
+
raw_text = output.outputs[0].text.strip()
|
382 |
+
# Extract answer from thinking tokens
|
383 |
+
markdown_text = extract_answer_from_thinking(raw_text, include_thinking)
|
384 |
+
all_markdown.append(markdown_text)
|
385 |
+
|
386 |
+
except Exception as e:
|
387 |
+
logger.error(f"Error processing batch: {e}")
|
388 |
+
# Add error placeholders for failed batch
|
389 |
+
all_markdown.extend(["[OCR FAILED]"] * len(batch_images))
|
390 |
+
|
391 |
+
# Add markdown column to dataset
|
392 |
+
logger.info("Adding markdown column to dataset")
|
393 |
+
dataset = dataset.add_column("markdown", all_markdown)
|
394 |
+
|
395 |
+
# Handle inference_info tracking
|
396 |
+
logger.info("Updating inference_info...")
|
397 |
+
|
398 |
+
# Check for existing inference_info
|
399 |
+
if "inference_info" in dataset.column_names:
|
400 |
+
# Parse existing info from first row (all rows have same info)
|
401 |
+
try:
|
402 |
+
existing_info = json.loads(dataset[0]["inference_info"])
|
403 |
+
if not isinstance(existing_info, list):
|
404 |
+
existing_info = [existing_info] # Convert old format to list
|
405 |
+
except (json.JSONDecodeError, TypeError):
|
406 |
+
existing_info = []
|
407 |
+
# Remove old column to update it
|
408 |
+
dataset = dataset.remove_columns(["inference_info"])
|
409 |
+
else:
|
410 |
+
existing_info = []
|
411 |
+
|
412 |
+
# Add new inference info
|
413 |
+
new_info = {
|
414 |
+
"column_name": "markdown",
|
415 |
+
"model_id": model,
|
416 |
+
"processing_date": datetime.now().isoformat(),
|
417 |
+
"batch_size": batch_size,
|
418 |
+
"max_tokens": max_tokens,
|
419 |
+
"gpu_memory_utilization": gpu_memory_utilization,
|
420 |
+
"max_model_len": max_model_len,
|
421 |
+
"include_thinking": include_thinking,
|
422 |
+
"temperature": temperature,
|
423 |
+
"prompt": prompt,
|
424 |
+
"script": "numarkdown-ocr.py",
|
425 |
+
"script_version": "1.0.0",
|
426 |
+
"script_url": "https://huggingface.co/datasets/uv-scripts/ocr/raw/main/numarkdown-ocr.py"
|
427 |
+
}
|
428 |
+
existing_info.append(new_info)
|
429 |
+
|
430 |
+
# Add updated inference_info column
|
431 |
+
info_json = json.dumps(existing_info, ensure_ascii=False)
|
432 |
+
dataset = dataset.add_column("inference_info", [info_json] * len(dataset))
|
433 |
+
|
434 |
+
# Push to hub
|
435 |
+
logger.info(f"Pushing to {output_dataset}")
|
436 |
+
dataset.push_to_hub(output_dataset, private=private, token=HF_TOKEN)
|
437 |
+
|
438 |
+
# Calculate processing time
|
439 |
+
end_time = datetime.now()
|
440 |
+
processing_duration = end_time - start_time
|
441 |
+
processing_time = f"{processing_duration.total_seconds() / 60:.1f} minutes"
|
442 |
+
|
443 |
+
# Create and push dataset card
|
444 |
+
logger.info("Creating dataset card...")
|
445 |
+
card_content = create_dataset_card(
|
446 |
+
source_dataset=input_dataset,
|
447 |
+
model=model,
|
448 |
+
num_samples=len(dataset),
|
449 |
+
processing_time=processing_time,
|
450 |
+
batch_size=batch_size,
|
451 |
+
max_model_len=max_model_len,
|
452 |
+
max_tokens=max_tokens,
|
453 |
+
gpu_memory_utilization=gpu_memory_utilization,
|
454 |
+
include_thinking=include_thinking,
|
455 |
+
image_column=image_column,
|
456 |
+
split=split,
|
457 |
+
)
|
458 |
+
|
459 |
+
card = DatasetCard(card_content)
|
460 |
+
card.push_to_hub(output_dataset, token=HF_TOKEN)
|
461 |
+
logger.info("✅ Dataset card created and pushed!")
|
462 |
+
|
463 |
+
logger.info("✅ OCR conversion complete!")
|
464 |
+
logger.info(
|
465 |
+
f"Dataset available at: https://huggingface.co/datasets/{output_dataset}"
|
466 |
+
)
|
467 |
+
|
468 |
+
|
469 |
+
if __name__ == "__main__":
|
470 |
+
# Show example usage if no arguments
|
471 |
+
if len(sys.argv) == 1:
|
472 |
+
print("=" * 80)
|
473 |
+
print("NuMarkdown-8B-Thinking OCR with Reasoning")
|
474 |
+
print("=" * 80)
|
475 |
+
print("\nThis script converts document images to markdown using")
|
476 |
+
print("the NuMarkdown-8B-Thinking model with advanced reasoning capabilities.")
|
477 |
+
print("\nFeatures:")
|
478 |
+
print("- 🧠 Reasoning-based document analysis")
|
479 |
+
print("- 📊 Superior table extraction and formatting")
|
480 |
+
print("- 📐 Mathematical formula recognition")
|
481 |
+
print("- 📝 Complex layout understanding")
|
482 |
+
print("- ✨ Clean markdown generation")
|
483 |
+
print("- 🔍 Optional thinking trace inclusion")
|
484 |
+
print("\nExample usage:")
|
485 |
+
print("\n1. Basic OCR conversion:")
|
486 |
+
print(" uv run numarkdown-ocr.py document-images markdown-docs")
|
487 |
+
print("\n2. Include thinking traces:")
|
488 |
+
print(" uv run numarkdown-ocr.py complex-docs analyzed-docs --include-thinking")
|
489 |
+
print("\n3. With custom settings:")
|
490 |
+
print(" uv run numarkdown-ocr.py scientific-papers extracted-text \\")
|
491 |
+
print(" --batch-size 8 \\")
|
492 |
+
print(" --max-tokens 8192 \\")
|
493 |
+
print(" --gpu-memory-utilization 0.9")
|
494 |
+
print("\n4. Process a subset for testing:")
|
495 |
+
print(" uv run numarkdown-ocr.py large-dataset test-output --max-samples 10")
|
496 |
+
print("\n5. Custom prompt for specific needs:")
|
497 |
+
print(" uv run numarkdown-ocr.py invoices invoice-data \\")
|
498 |
+
print(' --custom-prompt "Extract all invoice details including line items"')
|
499 |
+
print("\n6. Running on HF Jobs:")
|
500 |
+
print(" hf jobs uv run --flavor l4x1 \\")
|
501 |
+
print(' -e HF_TOKEN=$(python3 -c "from huggingface_hub import get_token; print(get_token())") \\')
|
502 |
+
print(" https://huggingface.co/datasets/uv-scripts/ocr/raw/main/numarkdown-ocr.py \\")
|
503 |
+
print(" your-document-dataset \\")
|
504 |
+
print(" your-markdown-output")
|
505 |
+
print("\n" + "=" * 80)
|
506 |
+
print("\nFor full help, run: uv run numarkdown-ocr.py --help")
|
507 |
+
sys.exit(0)
|
508 |
+
|
509 |
+
parser = argparse.ArgumentParser(
|
510 |
+
description="OCR images to markdown using NuMarkdown-8B-Thinking with reasoning",
|
511 |
+
formatter_class=argparse.RawDescriptionHelpFormatter,
|
512 |
+
epilog="""
|
513 |
+
Examples:
|
514 |
+
# Basic usage
|
515 |
+
uv run numarkdown-ocr.py my-images-dataset ocr-results
|
516 |
+
|
517 |
+
# Include thinking traces in output
|
518 |
+
uv run numarkdown-ocr.py documents analyzed-docs --include-thinking
|
519 |
+
|
520 |
+
# Process subset for testing
|
521 |
+
uv run numarkdown-ocr.py large-dataset test-output --max-samples 100
|
522 |
+
|
523 |
+
# Custom prompt for specific extraction
|
524 |
+
uv run numarkdown-ocr.py forms form-data --custom-prompt "Extract all form fields and values"
|
525 |
+
|
526 |
+
# Random sample from dataset
|
527 |
+
uv run numarkdown-ocr.py ordered-dataset random-sample --max-samples 50 --shuffle
|
528 |
+
""",
|
529 |
+
)
|
530 |
+
|
531 |
+
parser.add_argument("input_dataset", help="Input dataset ID from Hugging Face Hub")
|
532 |
+
parser.add_argument("output_dataset", help="Output dataset ID for Hugging Face Hub")
|
533 |
+
parser.add_argument(
|
534 |
+
"--image-column",
|
535 |
+
default="image",
|
536 |
+
help="Column containing images (default: image)",
|
537 |
+
)
|
538 |
+
parser.add_argument(
|
539 |
+
"--batch-size",
|
540 |
+
type=int,
|
541 |
+
default=16,
|
542 |
+
help="Batch size for processing (default: 16, lower than others due to model size)",
|
543 |
+
)
|
544 |
+
parser.add_argument(
|
545 |
+
"--model",
|
546 |
+
default="numind/NuMarkdown-8B-Thinking",
|
547 |
+
help="Model to use (default: numind/NuMarkdown-8B-Thinking)",
|
548 |
+
)
|
549 |
+
parser.add_argument(
|
550 |
+
"--max-model-len",
|
551 |
+
type=int,
|
552 |
+
default=16384,
|
553 |
+
help="Maximum model context length (default: 16384)",
|
554 |
+
)
|
555 |
+
parser.add_argument(
|
556 |
+
"--max-tokens",
|
557 |
+
type=int,
|
558 |
+
default=8192,
|
559 |
+
help="Maximum tokens to generate (default: 8192)",
|
560 |
+
)
|
561 |
+
parser.add_argument(
|
562 |
+
"--gpu-memory-utilization",
|
563 |
+
type=float,
|
564 |
+
default=0.9,
|
565 |
+
help="GPU memory utilization (default: 0.9)",
|
566 |
+
)
|
567 |
+
parser.add_argument("--hf-token", help="Hugging Face API token")
|
568 |
+
parser.add_argument(
|
569 |
+
"--split", default="train", help="Dataset split to use (default: train)"
|
570 |
+
)
|
571 |
+
parser.add_argument(
|
572 |
+
"--max-samples",
|
573 |
+
type=int,
|
574 |
+
help="Maximum number of samples to process (for testing)",
|
575 |
+
)
|
576 |
+
parser.add_argument(
|
577 |
+
"--private", action="store_true", help="Make output dataset private"
|
578 |
+
)
|
579 |
+
parser.add_argument(
|
580 |
+
"--shuffle",
|
581 |
+
action="store_true",
|
582 |
+
help="Shuffle the dataset before processing (useful for random sampling)",
|
583 |
+
)
|
584 |
+
parser.add_argument(
|
585 |
+
"--seed",
|
586 |
+
type=int,
|
587 |
+
default=42,
|
588 |
+
help="Random seed for shuffling (default: 42)",
|
589 |
+
)
|
590 |
+
parser.add_argument(
|
591 |
+
"--include-thinking",
|
592 |
+
action="store_true",
|
593 |
+
help="Include thinking traces in output (default: only final answers)",
|
594 |
+
)
|
595 |
+
parser.add_argument(
|
596 |
+
"--temperature",
|
597 |
+
type=float,
|
598 |
+
default=0.0,
|
599 |
+
help="Temperature for generation (default: 0.0 for deterministic)",
|
600 |
+
)
|
601 |
+
parser.add_argument(
|
602 |
+
"--custom-prompt",
|
603 |
+
type=str,
|
604 |
+
help="Custom prompt for the model (overrides default)",
|
605 |
+
)
|
606 |
+
|
607 |
+
args = parser.parse_args()
|
608 |
+
|
609 |
+
main(
|
610 |
+
input_dataset=args.input_dataset,
|
611 |
+
output_dataset=args.output_dataset,
|
612 |
+
image_column=args.image_column,
|
613 |
+
batch_size=args.batch_size,
|
614 |
+
model=args.model,
|
615 |
+
max_model_len=args.max_model_len,
|
616 |
+
max_tokens=args.max_tokens,
|
617 |
+
gpu_memory_utilization=args.gpu_memory_utilization,
|
618 |
+
hf_token=args.hf_token,
|
619 |
+
split=args.split,
|
620 |
+
max_samples=args.max_samples,
|
621 |
+
private=args.private,
|
622 |
+
shuffle=args.shuffle,
|
623 |
+
seed=args.seed,
|
624 |
+
include_thinking=args.include_thinking,
|
625 |
+
temperature=args.temperature,
|
626 |
+
custom_prompt=args.custom_prompt,
|
627 |
+
)
|