davanstrien HF Staff Claude commited on
Commit
2d80026
Β·
1 Parent(s): 3073a00

Document DeepSeek-OCR in README with real examples

Browse files

Added comprehensive documentation for deepseek-ocr-vllm.py:
- Added to Available Scripts section with ⭐ NEW badge
- Documented all 5 resolution modes (tiny/small/base/large/gundam)
- Documented all 5 prompt modes (document/image/free/figure/describe)
- Added 3 real-world HF Jobs examples:
* National Library of Scotland handbooks (a100-large, large mode)
* Fast testing with tiny mode (l4x1)
* Figure parsing from scientific papers
- Added script-specific options section
- Emphasized UV_TORCH_BACKEND=auto requirement for HF Jobs

πŸ€– Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Files changed (1) hide show
  1. README.md +64 -1
README.md CHANGED
@@ -31,6 +31,34 @@ That's it! The script will:
31
 
32
  ## πŸ“‹ Available Scripts
33
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
34
  ### RolmOCR (`rolm-ocr.py`)
35
 
36
  Fast general-purpose OCR using [reducto/RolmOCR](https://huggingface.co/reducto/RolmOCR) based on Qwen2.5-VL-7B:
@@ -141,7 +169,35 @@ Every OCR run now generates comprehensive dataset documentation including:
141
  No GPU? No problem! Run on HF infrastructure:
142
 
143
  ```bash
144
- # Basic OCR job
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
145
  hf jobs uv run --flavor l4x1 \
146
  --secrets HF_TOKEN \
147
  https://huggingface.co/datasets/uv-scripts/ocr/raw/main/nanonets-ocr.py \
@@ -255,6 +311,13 @@ Any HuggingFace dataset containing images - documents, forms, receipts, books, h
255
 
256
  ### Script-Specific Options
257
 
 
 
 
 
 
 
 
258
  **RolmOCR**:
259
  - Output column is auto-generated from model name (e.g., `rolmocr_text`)
260
  - Use `--output-column` to override the default name
 
31
 
32
  ## πŸ“‹ Available Scripts
33
 
34
+ ### DeepSeek-OCR (`deepseek-ocr-vllm.py`) ⭐ NEW
35
+
36
+ Advanced document OCR using [deepseek-ai/DeepSeek-OCR](https://huggingface.co/deepseek-ai/DeepSeek-OCR) with visual-text compression:
37
+
38
+ - πŸ“ **LaTeX equations** - Mathematical formulas in LaTeX format
39
+ - πŸ“Š **Tables** - Extracted as HTML/markdown
40
+ - πŸ“ **Document structure** - Headers, lists, formatting preserved
41
+ - πŸ–ΌοΈ **Image grounding** - Spatial layout with bounding boxes
42
+ - πŸ” **Complex layouts** - Multi-column and hierarchical structures
43
+ - 🌍 **Multilingual** - Multiple language support
44
+ - 🎚️ **Resolution modes** - 5 presets for speed/quality trade-offs
45
+ - πŸ’¬ **Prompt modes** - 5 presets for different OCR tasks
46
+ - ⚑ **Fast batch processing** - vLLM acceleration
47
+
48
+ **Resolution Modes:**
49
+ - `tiny` (512Γ—512): Fast, 64 vision tokens
50
+ - `small` (640Γ—640): Balanced, 100 vision tokens
51
+ - `base` (1024Γ—1024): High quality, 256 vision tokens
52
+ - `large` (1280Γ—1280): Maximum quality, 400 vision tokens
53
+ - `gundam` (dynamic): Adaptive multi-tile (default)
54
+
55
+ **Prompt Modes:**
56
+ - `document`: Convert to markdown with grounding (default)
57
+ - `image`: OCR any image with grounding
58
+ - `free`: Fast OCR without layout
59
+ - `figure`: Parse figures from documents
60
+ - `describe`: Detailed image descriptions
61
+
62
  ### RolmOCR (`rolm-ocr.py`)
63
 
64
  Fast general-purpose OCR using [reducto/RolmOCR](https://huggingface.co/reducto/RolmOCR) based on Qwen2.5-VL-7B:
 
169
  No GPU? No problem! Run on HF infrastructure:
170
 
171
  ```bash
172
+ # DeepSeek-OCR - Real-world example (National Library of Scotland handbooks)
173
+ hf jobs uv run --flavor a100-large \
174
+ -s HF_TOKEN \
175
+ -e UV_TORCH_BACKEND=auto \
176
+ https://huggingface.co/datasets/uv-scripts/ocr/raw/main/deepseek-ocr-vllm.py \
177
+ NationalLibraryOfScotland/Britain-and-UK-Handbooks-Dataset \
178
+ davanstrien/handbooks-deep-ocr \
179
+ --max-samples 100 \
180
+ --shuffle \
181
+ --resolution-mode large
182
+
183
+ # DeepSeek-OCR - Fast testing with tiny mode
184
+ hf jobs uv run --flavor l4x1 \
185
+ -s HF_TOKEN \
186
+ -e UV_TORCH_BACKEND=auto \
187
+ https://huggingface.co/datasets/uv-scripts/ocr/raw/main/deepseek-ocr-vllm.py \
188
+ your-input-dataset your-output-dataset \
189
+ --max-samples 10 \
190
+ --resolution-mode tiny
191
+
192
+ # DeepSeek-OCR - Parse figures from scientific papers
193
+ hf jobs uv run --flavor a100-large \
194
+ -s HF_TOKEN \
195
+ -e UV_TORCH_BACKEND=auto \
196
+ https://huggingface.co/datasets/uv-scripts/ocr/raw/main/deepseek-ocr-vllm.py \
197
+ scientific-papers figures-extracted \
198
+ --prompt-mode figure
199
+
200
+ # Basic OCR job with Nanonets
201
  hf jobs uv run --flavor l4x1 \
202
  --secrets HF_TOKEN \
203
  https://huggingface.co/datasets/uv-scripts/ocr/raw/main/nanonets-ocr.py \
 
311
 
312
  ### Script-Specific Options
313
 
314
+ **DeepSeek-OCR**:
315
+ - `--resolution-mode`: Quality level - `tiny`, `small`, `base`, `large`, or `gundam` (default)
316
+ - `--prompt-mode`: Task type - `document` (default), `image`, `free`, `figure`, or `describe`
317
+ - `--prompt`: Custom OCR prompt (overrides prompt-mode)
318
+ - `--base-size`, `--image-size`, `--crop-mode`: Override resolution mode manually
319
+ - ⚠️ **Important for HF Jobs**: Add `-e UV_TORCH_BACKEND=auto` for proper PyTorch installation
320
+
321
  **RolmOCR**:
322
  - Output column is auto-generated from model name (e.g., `rolmocr_text`)
323
  - Use `--output-column` to override the default name