Hugging Face's logo

tiiuae
/
Falcon-OCR

Image-to-Text
Transformers
Safetensors
falcon_ocr
text-generation
falcon
ocr
vision-language
document-understanding
custom_code
Eval Results
Model card Files
xet
 Community
11

Update README.md

#6
by griffintaur - opened
base: refs/heads/main
←
from: refs/pr/6
Discussion Files changed
+24
-20
Files changed (1) hide show
  1. README.md +24 -20
README.md CHANGED
@@ -112,17 +112,19 @@ The layout model is loaded lazily on the first `generate_with_layout()` call and
112
 
113
  Category-wise performance comparison of FalconOCR against state-of-the-art OCR models. We report accuracy (%) across all category splits.
114
 
115
- | Model | Average | ArXiv Math | Base | Hdr/Ftr | TinyTxt | MultCol | OldScan | OldMath | Tables |
116
- |---|---|---|---|---|---|---|---|---|---|
117
- | Mistral OCR 3 | 81.7 | **85.4** | **99.9** | 93.8 | 88.9 | 82.1 | 48.8 | 68.3 | 86.1 |
118
- | Chandra | **82.0** | 81.4 | 99.8 | 88.8 | **91.9** | 82.9 | **49.2** | 73.6 | 88.2 |
119
- | Gemini 3 Pro | 80.2 | 70.6 | 99.8 | 84.0 | 90.3 | 79.2 | 47.5 | 84.9 | 84.9 |
120
- | PaddleOCR VL 1.5 | 79.3 | **85.4** | 98.8 | **96.9** | 80.8 | 82.6 | 39.2 | 66.4 | 84.1 |
121
- | PaddleOCR VL | 79.2 | **85.4** | 98.6 | **96.9** | 80.8 | 82.5 | 38.8 | 66.4 | 83.9 |
122
- | DeepSeek OCR v2 | 78.8 | 81.9 | 99.8 | 95.6 | 88.7 | 83.6 | 33.7 | 68.8 | 78.1 |
123
- | Gemini 3 Flash | 77.5 | 66.5 | 99.8 | 83.8 | 88.2 | 73.7 | 46.0 | **85.8** | 75.9 |
124
- | GPT 5.2 | 69.8 | 61.0 | 99.8 | 75.6 | 62.2 | 70.2 | 34.6 | 75.8 | 79.0 |
125
- | FalconOCR | 80.3 | 80.5 | 99.5 | 94.0 | 78.5 | **87.1** | 43.5 | 69.2 | **90.3** |
 
 
126
  </details>
127
 
128
  <details name="benchmarks">
@@ -130,15 +132,17 @@ Category-wise performance comparison of FalconOCR against state-of-the-art OCR m
130
 
131
  Performance comparison on full-page document parsing. Overall↑ aggregates the three sub-metrics. Edit↓ measures text edit distance (lower is better). CDM↑ evaluates formula recognition accuracy. TEDS↑ measures table structure similarity.
132
 
133
- | Model | Overall↑ | Edit↓ | CDM↑ | TEDS↑ |
134
- |---|---|---|---|---|
135
- | PaddleOCR VL 1.5 | **94.37** | 0.025 | **94.4** | **91.1** |
136
- | PaddleOCR VL | 91.76 | **0.024** | 91.7 | 85.9 |
137
- | Chandra | 88.97 | 0.046 | 88.1 | 89.5 |
138
- | DeepSeek OCR v2 | 87.66 | 0.037 | 89.2 | 77.5 |
139
- | GPT 5.2 | 86.56 | 0.061 | 88.0 | 77.7 |
140
- | Mistral OCR 3 | 85.20 | 0.053 | 84.3 | 76.1 |
141
- | FalconOCR | 88.64 | 0.055 | 86.8 | 84.6 |
 
 
142
  </details>
143
 
144
  ### Results Analysis
 
112
 
113
  Category-wise performance comparison of FalconOCR against state-of-the-art OCR models. We report accuracy (%) across all category splits.
114
 
115
+ <table>
116
+ <tr><th>Model</th><th>Average</th><th>ArXiv Math</th><th>Base</th><th>Hdr/Ftr</th><th>TinyTxt</th><th>MultCol</th><th>OldScan</th><th>OldMath</th><th>Tables</th></tr>
117
+ <tr><td>Mistral OCR 3</td><td>81.7</td><td><b>85.4</b></td><td><b>99.9</b></td><td>93.8</td><td>88.9</td><td>82.1</td><td>48.8</td><td>68.3</td><td>86.1</td></tr>
118
+ <tr><td>Chandra</td><td><b>82.0</b></td><td>81.4</td><td>99.8</td><td>88.8</td><td><b>91.9</b></td><td>82.9</td><td><b>49.2</b></td><td>73.6</td><td>88.2</td></tr>
119
+ <tr><td>Gemini 3 Pro</td><td>80.2</td><td>70.6</td><td>99.8</td><td>84.0</td><td>90.3</td><td>79.2</td><td>47.5</td><td>84.9</td><td>84.9</td></tr>
120
+ <tr><td>PaddleOCR VL 1.5</td><td>79.3</td><td><b>85.4</b></td><td>98.8</td><td><b>96.9</b></td><td>80.8</td><td>82.6</td><td>39.2</td><td>66.4</td><td>84.1</td></tr>
121
+ <tr><td>PaddleOCR VL</td><td>79.2</td><td><b>85.4</b></td><td>98.6</td><td><b>96.9</b></td><td>80.8</td><td>82.5</td><td>38.8</td><td>66.4</td><td>83.9</td></tr>
122
+ <tr><td>DeepSeek OCR v2</td><td>78.8</td><td>81.9</td><td>99.8</td><td>95.6</td><td>88.7</td><td>83.6</td><td>33.7</td><td>68.8</td><td>78.1</td></tr>
123
+ <tr><td>Gemini 3 Flash</td><td>77.5</td><td>66.5</td><td>99.8</td><td>83.8</td><td>88.2</td><td>73.7</td><td>46.0</td><td><b>85.8</b></td><td>75.9</td></tr>
124
+ <tr><td>GPT 5.2</td><td>69.8</td><td>61.0</td><td>99.8</td><td>75.6</td><td>62.2</td><td>70.2</td><td>34.6</td><td>75.8</td><td>79.0</td></tr>
125
+ <tr style="background:#dbeafe"><td><b>FalconOCR</b></td><td>80.3</td><td>80.5</td><td>99.5</td><td>94.0</td><td>78.5</td><td><b>87.1</b></td><td>43.5</td><td>69.2</td><td><b>90.3</b></td></tr>
126
+ </table>
127
+
128
  </details>
129
 
130
  <details name="benchmarks">
 
132
 
133
  Performance comparison on full-page document parsing. Overall↑ aggregates the three sub-metrics. Edit↓ measures text edit distance (lower is better). CDM↑ evaluates formula recognition accuracy. TEDS↑ measures table structure similarity.
134
 
135
+ <table>
136
+ <tr><th>Model</th><th>Overall↑</th><th>Edit↓</th><th>CDM↑</th><th>TEDS↑</th></tr>
137
+ <tr><td>PaddleOCR VL 1.5</td><td><b>94.37</b></td><td>0.025</td><td><b>94.4</b></td><td><b>91.1</b></td></tr>
138
+ <tr><td>PaddleOCR VL</td><td>91.76</td><td><b>0.024</b></td><td>91.7</td><td>85.9</td></tr>
139
+ <tr><td>Chandra</td><td>88.97</td><td>0.046</td><td>88.1</td><td>89.5</td></tr>
140
+ <tr><td>DeepSeek OCR v2</td><td>87.66</td><td>0.037</td><td>89.2</td><td>77.5</td></tr>
141
+ <tr><td>GPT 5.2</td><td>86.56</td><td>0.061</td><td>88.0</td><td>77.7</td></tr>
142
+ <tr><td>Mistral OCR 3</td><td>85.20</td><td>0.053</td><td>84.3</td><td>76.1</td></tr>
143
+ <tr style="background:#dbeafe"><td><b>FalconOCR</b></td><td>88.64</td><td>0.055</td><td>86.8</td><td>84.6</td></tr>
144
+ </table>
145
+
146
  </details>
147
 
148
  ### Results Analysis