Commit
Β·
a77e763
1
Parent(s):
d805bea
Add LightOnOCR documentation to README
Browse filesAdd LightOnOCR as first script in README with recommendation to test
first since it's small and fast.
Features highlighted:
- Fastest OCR option (5.71 pages/sec on H100, 6.25 images/sec on A100)
- Compact 1B parameters for quick downloads
- 3 vocabulary sizes (151k/32k/16k) with European language optimization
- LaTeX formulas and table extraction
- Production-ready with 76.1% benchmark score
Includes quick start examples:
- L4 test run with 100 samples
- A100 production run with batch_size=4096
Positioning LightOnOCR as the recommended starting point for users
new to the OCR scripts due to fast initialization and excellent
speed/quality balance.
π€ Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
README.md
CHANGED
|
@@ -31,6 +31,44 @@ That's it! The script will:
|
|
| 31 |
|
| 32 |
## π Available Scripts
|
| 33 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
### DeepSeek-OCR (`deepseek-ocr-vllm.py`) β NEW
|
| 35 |
|
| 36 |
Advanced document OCR using [deepseek-ai/DeepSeek-OCR](https://huggingface.co/deepseek-ai/DeepSeek-OCR) with visual-text compression:
|
|
|
|
| 31 |
|
| 32 |
## π Available Scripts
|
| 33 |
|
| 34 |
+
### LightOnOCR (`lighton-ocr.py`) β‘ Good one to test first since it's small and fast!
|
| 35 |
+
|
| 36 |
+
Fast and compact OCR using [lightonai/LightOnOCR-1B-1025](https://huggingface.co/lightonai/LightOnOCR-1B-1025):
|
| 37 |
+
|
| 38 |
+
- β‘ **Fastest**: 5.71 pages/sec on H100, ~6.25 images/sec on A100 with batch_size=4096
|
| 39 |
+
- π― **Compact**: Only 1B parameters - quick to download and initialize
|
| 40 |
+
- π **Multilingual**: 3 vocabulary sizes for different use cases
|
| 41 |
+
- π **LaTeX formulas**: Mathematical notation in LaTeX format
|
| 42 |
+
- π **Table extraction**: Markdown table format
|
| 43 |
+
- π **Document structure**: Preserves hierarchy and layout
|
| 44 |
+
- π **Production-ready**: 76.1% benchmark score, used in production
|
| 45 |
+
|
| 46 |
+
**Vocabulary sizes:**
|
| 47 |
+
- `151k`: Full vocabulary, all languages (default)
|
| 48 |
+
- `32k`: European languages, ~12% faster decoding
|
| 49 |
+
- `16k`: European languages, ~12% faster decoding
|
| 50 |
+
|
| 51 |
+
**Quick start:**
|
| 52 |
+
```bash
|
| 53 |
+
# Test on 100 samples with English text (32k vocab is fastest for European languages)
|
| 54 |
+
hf jobs uv run --flavor l4x1 \
|
| 55 |
+
-s HF_TOKEN \
|
| 56 |
+
https://huggingface.co/datasets/uv-scripts/ocr/raw/main/lighton-ocr.py \
|
| 57 |
+
your-input-dataset your-output-dataset \
|
| 58 |
+
--vocab-size 32k \
|
| 59 |
+
--batch-size 32 \
|
| 60 |
+
--max-samples 100
|
| 61 |
+
|
| 62 |
+
# Full production run on A100 (can handle huge batches!)
|
| 63 |
+
hf jobs uv run --flavor a100-large \
|
| 64 |
+
-s HF_TOKEN \
|
| 65 |
+
https://huggingface.co/datasets/uv-scripts/ocr/raw/main/lighton-ocr.py \
|
| 66 |
+
your-input-dataset your-output-dataset \
|
| 67 |
+
--vocab-size 32k \
|
| 68 |
+
--batch-size 4096 \
|
| 69 |
+
--temperature 0.0
|
| 70 |
+
```
|
| 71 |
+
|
| 72 |
### DeepSeek-OCR (`deepseek-ocr-vllm.py`) β NEW
|
| 73 |
|
| 74 |
Advanced document OCR using [deepseek-ai/DeepSeek-OCR](https://huggingface.co/deepseek-ai/DeepSeek-OCR) with visual-text compression:
|