Mirrors/dots.ocr
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

dots.ocr-1.5 dots.ocr-1.5-svg released

Browse Source
This commit is contained in:
zhangwei13
2026-02-16 01:53:00 +08:00
parent c3084bd437
commit d2e359c715
14 changed files with 1407 additions and 981 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+529 -971
View File
File diff suppressed because it is too large Load Diff
README_hf.md
Executable
+778
View File
@@ -0,0 +1,778 @@
---
license: mit
library_name: dots_ocr_1_5
pipeline_tag: image-text-to-text
tags:
- image-to-text
- ocr
- document-parse
- layout
- table
- formula
- transformers
- custom_code
language:
- en
- zh
- multilingual
---
<div align="center">
<p align="center">
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/logo.png" width="300"/>
<p>
<h1 align="center">
dots.ocr-1.5: Recognize Any Human Scripts and Symbols
</h1>
[![HuggingFace](https://img.shields.io/badge/HuggingFace%20Weights-black.svg?logo=HuggingFace)](https://huggingface.co/rednote-hilab/dots.ocr-1.5)
<div align="center">
<a href="https://dotsocr.xiaohongshu.com" target="_blank" rel="noopener noreferrer"><strong>🖥️ Live Demo</strong></a> |
<a href="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/wechat.jpg" target="_blank" rel="noopener noreferrer"><strong>💬 WeChat</strong></a> |
<a href="https://www.xiaohongshu.com/user/profile/683ffe42000000001d021a4c" target="_blank" rel="noopener noreferrer"><strong>📕 rednote</strong></a>
</div>
</div>
## Introduction
We present **dots.ocr-1.5**, a 3B-parameter multimodal model composed of a 1.2B vision encoder and a 1.7B language model. Designed for universal accessibility, it possesses the capability to recognize virtually any human script. Beyond achieving state-of-the-art (SOTA) performance in standard multilingual document parsing among models of comparable size, dots.ocr-1.5 excels at converting structured graphics (e.g., charts and diagrams) directly into SVG code, parsing web screens and spotting scene text. Furthermore, the model demonstrates competitive performance in general OCR, object grounding & counting tasks.
1. **Stronger Document Parsing Performance:** dots.ocr-1.5 maintains SOTA performance among latest OCR models, particularly on **multilingual documents**. Addressing the significant bias inherent in the detection & matching rules of certain benchmarks —which often fail to accurately reflect a model's true capabilities—we adopted an **Elo score** evaluation system. Under this metric, the performance landscape shifts significantly, highlighting the superior robustness of our model compared to conventional rankings.
2. **Unified Vision-Language Parsing**: Visual languages (e.g., charts, graphics, chemical formulas, logos) encapsulate dense human knowledge, akin to natural language. dots.ocr-1.5 unifies the interpretation of these elements by parsing them directly into SVG code. We have validated the effectiveness of this approach, demonstrating impressive results in structural and semantic recognition.
3. **Broader and More General Capabilities**: Compared to dots.ocr, dots.ocr-1.5 supports a significantly wider array of tasks. It extends beyond standard OCR to handle web screen parsing, scene text spotting, object grounding & counting, and other general OCR QA tasks.
## Evaluation
### 1. Document Parsing
#### 1.1 Elo Score of different bench between latest models
<table style="border-collapse: collapse; width: 100%; font-family: Arial, sans-serif;">
<thead>
<tr style="background-color: #f2f2f2; text-align: left;">
<th style="border: 1px solid #ddd; padding: 8px;">models</th>
<th style="border: 1px solid #ddd; padding: 8px;">olmOCR-Bench</th>
<th style="border: 1px solid #ddd; padding: 8px;">OmniDocBench (v1.5)</th>
<th style="border: 1px solid #ddd; padding: 8px;">XDocParse</th>
</tr>
</thead>
<tbody>
<tr>
<td style="border: 1px solid #ddd; padding: 8px;">GLM-OCR</td>
<td style="border: 1px solid #ddd; padding: 8px;">859.9</td>
<td style="border: 1px solid #ddd; padding: 8px;">937.5</td>
<td style="border: 1px solid #ddd; padding: 8px;">742.1</td>
</tr>
<tr>
<td style="border: 1px solid #ddd; padding: 8px;">PaddleOCR-VL-1.5</td>
<td style="border: 1px solid #ddd; padding: 8px;">873.6</td>
<td style="border: 1px solid #ddd; padding: 8px;">965.6</td>
<td style="border: 1px solid #ddd; padding: 8px;">797.6</td>
</tr>
<tr>
<td style="border: 1px solid #ddd; padding: 8px;">HuanyuanOCR</td>
<td style="border: 1px solid #ddd; padding: 8px;">978.9</td>
<td style="border: 1px solid #ddd; padding: 8px;">974.4</td>
<td style="border: 1px solid #ddd; padding: 8px;">895.9</td>
</tr>
<tr>
<td style="border: 1px solid #ddd; padding: 8px;">dots.ocr</td>
<td style="border: 1px solid #ddd; padding: 8px;">1027.4</td>
<td style="border: 1px solid #ddd; padding: 8px;">994.7</td>
<td style="border: 1px solid #ddd; padding: 8px;">1133.4</td>
</tr>
<!-- Highlighting dots.ocr-1.5 row -->
<tr style="background-color: #e6f7ff; font-weight: bold;">
<td style="border: 1px solid #ddd; padding: 8px;">dots.ocr-1.5</td>
<td style="border: 1px solid #ddd; padding: 8px;">1089.0</td>
<td style="border: 1px solid #ddd; padding: 8px;">1025.8</td>
<td style="border: 1px solid #ddd; padding: 8px;">1157.1</td>
</tr>
<tr>
<td style="border: 1px solid #ddd; padding: 8px;">Gemini 3 Pro</td>
<td style="border: 1px solid #ddd; padding: 8px;">1171.2</td>
<td style="border: 1px solid #ddd; padding: 8px;">1102.1</td>
<td style="border: 1px solid #ddd; padding: 8px;">1273.9</td>
</tr>
</tbody>
</table>
> **Notes:**
> - Results for Gemini 3 Pro, PaddleOCR-VL-1.5, and GLM-OCR were obtained via APIs, while HuanyuanOCR results were generated using local inference.
> - The Elo score evaluation was conducted using Gemini 3 Flash. The prompt can be found at: [Elo Score Prompt](https://github.com/rednote-hilab/dots.ocr/blob/master/tools/elo_score_prompt.py). These results are consistent with the findings on [ocrarena](https://www.ocrarena.ai/battle).
#### 1.2 olmOCR-bench
<!DOCTYPE html>
<html lang="zh">
<head>
<meta charset="UTF-8">
<style>
table {
width: 100%;
border-collapse: collapse;
font-family: Arial, sans-serif;
font-size: 14px;
color: #333;
}
th, td {
border: 1px solid #e0e0e0;
padding: 12px 8px;
text-align: left;
}
th {
background-color: #fafafa;
font-weight: normal;
vertical-align: top;
line-height: 1.4;
}
tr:nth-child(even) {
background-color: #ffffff;
}
tr:hover {
background-color: #f5f5f5;
}
.bold-row {
font-weight: bold;
}
</style>
</head>
<body>
<table>
<thead>
<tr>
<th></th>
<th>ArXiv</th>
<th>Old scans math</th>
<th>Tables</th>
<th>Old scans</th>
<th>Headers & footers</th>
<th>Multi column</th>
<th>Long tiny text</th>
<th>Base</th>
<th>Overall</th>
</tr>
</thead>
<tbody>
<tr>
<td>Mistral OCR API</td>
<td>77.2</td>
<td>67.5</td>
<td>60.6</td>
<td>29.3</td>
<td>93.6</td>
<td>71.3</td>
<td>77.1</td>
<td>99.4</td>
<td>72.0±1.1</td>
</tr>
<tr>
<td>Marker 1.10.1</td>
<td>83.8</td>
<td>66.8</td>
<td>72.9</td>
<td>33.5</td>
<td>86.6</td>
<td>80.0</td>
<td>85.7</td>
<td>99.3</td>
<td>76.1±1.1</td>
</tr>
<tr>
<td>MinerU 2.5.4*</td>
<td>76.6</td>
<td>54.6</td>
<td>84.9</td>
<td>33.7</td>
<td>96.6</td>
<td>78.2</td>
<td>83.5</td>
<td>93.7</td>
<td>75.2±1.1</td>
</tr>
<tr>
<td>DeepSeek-OCR</td>
<td>77.2</td>
<td>73.6</td>
<td>80.2</td>
<td>33.3</td>
<td>96.1</td>
<td>66.4</td>
<td>79.4</td>
<td>99.8</td>
<td>75.7±1.0</td>
</tr>
<tr>
<td>Nanonets-OCR2-3B</td>
<td>75.4</td>
<td>46.1</td>
<td>86.8</td>
<td>40.9</td>
<td>32.1</td>
<td>81.9</td>
<td>93.0</td>
<td>99.6</td>
<td>69.5±1.1</td>
</tr>
<tr>
<td>PaddleOCR-VL*</td>
<td>85.7</td>
<td>71.0</td>
<td>84.1</td>
<td>37.8</td>
<td>97.0</td>
<td>79.9</td>
<td>85.7</td>
<td>98.5</td>
<td>80.0±1.0</td>
</tr>
<tr>
<td>Infinity-Parser 7B*</td>
<td>84.4</td>
<td>83.8</td>
<td>85.0</td>
<td>47.9</td>
<td>88.7</td>
<td>84.2</td>
<td>86.4</td>
<td>99.8</td>
<td>82.5±?</td>
</tr>
<tr>
<td>olmOCR v0.4.0</td>
<td>83.0</td>
<td>82.3</td>
<td>84.9</td>
<td>47.7</td>
<td>96.1</td>
<td>83.7</td>
<td>81.9</td>
<td>99.7</td>
<td>82.4±1.1</td>
</tr>
<tr>
<td>Chandra OCR 0.1.0*</td>
<td>82.2</td>
<td>80.3</td>
<td>88.0</td>
<td>50.4</td>
<td>90.8</td>
<td>81.2</td>
<td>92.3</td>
<td>99.9</td>
<td>83.1±0.9</td>
</tr>
<tr>
<td>dots.ocr</td>
<td>82.1</td>
<td>64.2</td>
<td>88.3</td>
<td>40.9</td>
<td>94.1</td>
<td>82.4</td>
<td>81.2</td>
<td>99.5</td>
<td>79.1% ± 1.0%</td>
</tr>
<tr>
<td class="bold-row">dots.ocr-1.5</td>
<td><strong>85.9</strong></td>
<td><strong>85.5</strong></td>
<td><strong>90.7</strong></td>
<td>48.2</td>
<td>94.0</td>
<td><strong>85.3</strong></td>
<td>81.6</td>
<td>99.7</td>
<td><strong>83.9% ± 0.9</strong></td>
</tr>
</tbody>
</table>
</body>
</html>
> **Note:**
> - The metrics are from [olmocr](https://github.com/allenai/olmocr), and our own internal evaluations.
> - We delete the Page-header and Page-footer cells in the result markdown.
#### 1.3 Other Benchmarks
<table>
<thead>
<tr>
<th>Model Type</th>
<th>Methods</th>
<th>Size</th>
<th>OmniDocBench(v1.5)<br>TextEdit↓</th>
<th>OmniDocBench(v1.5)<br>Read OrderEdit↓</th>
<th>pdf-parse-bench</th>
</tr>
</thead>
<tbody>
<!-- GeneralVLMs Group (Reversed Order, 3 rows) -->
<tr>
<td rowspan="3"><strong>GeneralVLMs</strong></td>
<td>Gemini-2.5 Pro</td>
<td>-</td>
<td>0.075</td>
<td>0.097</td>
<td>9.06</td>
</tr>
<tr>
<td>Qwen3-VL-235B-A22B-Instruct</td>
<td>235B</td>
<td>0.069</td>
<td>0.068</td>
<td><strong>9.71</strong></td>
</tr>
<tr>
<td>gemini3pro</td>
<td>-</td>
<td>0.066</td>
<td>0.079</td>
<td>9.68</td>
</tr>
<!-- SpecializedVLMs Group (Reversed Order, 12 rows) -->
<tr>
<td rowspan="12"><strong>SpecializedVLMs</strong></td>
<td>Mistral OCR</td>
<td>-</td>
<td>0.164</td>
<td>0.144</td>
<td>8.84</td>
</tr>
<tr>
<td>Deepseek-OCR</td>
<td>3B</td>
<td>0.073</td>
<td>0.086</td>
<td>8.26</td>
</tr>
<tr>
<td>MonkeyOCR-3B</td>
<td>3B</td>
<td>0.075</td>
<td>0.129</td>
<td>9.27</td>
</tr>
<tr>
<td>OCRVerse</td>
<td>4B</td>
<td>0.058</td>
<td>0.071</td>
<td>--</td>
</tr>
<tr>
<td>MonkeyOCR-pro-3B</td>
<td>3B</td>
<td>0.075</td>
<td>0.128</td>
<td>-</td>
</tr>
<tr>
<td>MinerU2.5</td>
<td>1.2B</td>
<td>0.047</td>
<td>0.044</td>
<td>-</td>
</tr>
<tr>
<td>PaddleOCR-VL</td>
<td>0.9B</td>
<td>0.035</td>
<td>0.043</td>
<td>9.51</td>
</tr>
<tr>
<td>HunyuanOCR</td>
<td>0.9B</td>
<td>0.042</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>PaddleOCR-VL1.5</td>
<td>0.9B</td>
<td>0.035</td>
<td>0.042</td>
<td>-</td>
</tr>
<tr>
<td>GLMOCR</td>
<td>0.9B</td>
<td>0.04</td>
<td>0.043</td>
<td>-</td>
</tr>
<tr>
<td>dots.ocr</td>
<td>3B</td>
<td>0.048</td>
<td>0.053</td>
<td>9.29</td>
</tr>
<tr>
<td><u><strong>dots.ocr-1.5</strong></u></td>
<td>3B</td>
<td><strong>0.031</strong></td>
<td><strong>0.029</strong></td>
<td>9.54</td>
</tr>
</tbody>
</table>
> **Note:**
> - Metrics are sourced from [OmniDocBench](https://github.com/opendatalab/OmniDocBench) and other model publications. [pdf-parse-bench](https://github.com/phorn1/pdf-parse-bench) results are reproduced by Qwen3-VL-235B-A22B-Instruct.
> - Formula and Table metrics for OmniDocBench1.5 are omitted due to their high sensitivity to detection and matching protocols.
### 2. Vision-Language Parsing
Visual languages (e.g., charts, graphics, chemical formulas, logos) encapsulate dense human knowledge. **dots.ocr-1.5** unifies the interpretation of these elements by parsing them directly into **SVG code**.
<table>
<thead>
<tr>
<th rowspan="2" style="text-align: left;">Methods</th>
<th colspan="3">Unisvg</th>
<th rowspan="2">Chartmimic</th>
<th rowspan="2">Design2Code</th>
<th rowspan="2">Genexam</th>
<th rowspan="2">SciGen</th>
<th rowspan="2">ChemDraw</th>
</tr>
<tr>
<th>Low-Level</th>
<th>High-Level</th>
<th>Score</th>
</tr>
</thead>
<tbody>
<tr>
<td style="text-align: left;">OCRVerse</td>
<td>0.632</td>
<td>0.852</td>
<td>0.763</td>
<td>0.799</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>0.881</td>
</tr>
<tr>
<td style="text-align: left;">Gemini 3 Pro</td>
<td>0.563</td>
<td>0.850</td>
<td>0.735</td>
<td>0.788</td>
<td>0.760</td>
<td>0.756</td>
<td>0.783</td>
<td>0.839</td>
</tr>
<tr>
<td style="text-align: left;">dots.ocr-1.5</td>
<td>0.850</td>
<td>0.923</td>
<td>0.894</td>
<td>0.772</td>
<td>0.801</td>
<td>0.664</td>
<td>0.660</td>
<td>0.790</td>
</tr>
<tr>
<td style="text-align: left;"><strong>dots.ocr-1.5-svg</strong></td>
<td><strong>0.860</strong></td>
<td><strong>0.931</strong></td>
<td><strong>0.902</strong></td>
<td><strong>0.905</strong></td>
<td><strong>0.834</strong></td>
<td><strong>0.8</strong></td>
<td><strong>0.797</strong></td>
<td><strong>0.901</strong></td>
</tr>
</tbody>
</table>
> **Note:**
> - We use the ISVGEN metric from [UniSVG](https://ryanlijinke.github.io/) to evaluate the parsing result. For benchmarks that do not natively support image parsing, we use the original images as input, and calculate the ISVGEN score between the rendered output and the original image.
> - [OCRVerse](https://github.com/DocTron-hub/OCRVerse) results are derived from various code formats (e.g., SVG, Python), whereas results for Gemini 3 Pro and dots.ocr-1.5 are based specifically on SVG code.
> - Due to the capacity constraints of a 3B-parameter VLM, dots.ocr-1.5 may not excel in all tasks yet like svg. To complement this, we are simultaneously releasing dots.ocr-1.5-svg. We plan to further address these limitations in future updates.
### 3. General Vision Tasks
<table>
<thead>
<tr>
<th>Model</th>
<th>CharXiv_descriptive</th>
<th>CharXiv_reasoning</th>
<th>OCR_Reasoning</th>
<th>infovqa</th>
<th>docvqa</th>
<th>ChartQA</th>
<th>OCRBench</th>
<th>AI2D</th>
<th>CountBenchQA</th>
<th>refcoco</th>
</tr>
</thead>
<tbody>
<tr>
<td>Qwen3vl-2b-instruct</td>
<td>62.3</td>
<td>26.8</td>
<td>-</td>
<td>72.4</td>
<td>93.3</td>
<td>-</td>
<td>85.8</td>
<td>76.9</td>
<td>88.4</td>
<td>-</td>
</tr>
<tr>
<td><strong>dots.ocr-1.5</strong></td>
<td>77.4</td>
<td>55.3</td>
<td>22.85</td>
<td>73.76</td>
<td>91.85</td>
<td>83.2</td>
<td>86.0</td>
<td>82.16</td>
<td>94.46</td>
<td>80.03</td>
</tr>
</tbody>
</table>
# Quick Start
## 1. Installation
### Install dots.ocr-1.5
```shell
conda create -n dots_ocr python=3.12
conda activate dots_ocr
git clone https://github.com/rednote-hilab/dots.ocr.git
cd dots.ocr
# Install pytorch, see https://pytorch.org/get-started/previous-versions/ for your cuda version
pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl/cu128
pip install -e .
```
If you have trouble with the installation, try our [Docker Image](https://hub.docker.com/r/rednotehilab/dots.ocr) for an easier setup, and follow these steps:
```shell
git clone https://github.com/rednote-hilab/dots.ocr.git
cd dots.ocr
pip install -e .
```
### Download Model Weights
> 💡**Note:** Please use a directory name without periods (e.g., `DotsOCR_1_5` instead of `dots.ocr-1.5`) for the model save path. This is a temporary workaround pending our integration with Transformers.
```shell
python3 tools/download_model.py
```
## 2. Deployment
### vLLM inference
We highly recommend using vllm for deployment and inference.
```shell
# launch vllm server
## dots.ocr-1.5
CUDA_VISIBLE_DEVICES=0 vllm serve rednote-hilab/dots.ocr-1.5 --tensor-parallel-size 1 --gpu-memory-utilization 0.9 --chat-template-content-format string --served-model-name model --trust-remote-code
## dots.ocr-1.5-svg
CUDA_VISIBLE_DEVICES=0 vllm serve rednote-hilab/dots.ocr-1.5-svg --tensor-parallel-size 1 --gpu-memory-utilization 0.9 --chat-template-content-format string --served-model-name model --trust-remote-code
# vllm api demo
## document parsing
python3 ./demo/demo_vllm.py --prompt_mode prompt_layout_all_en
## web parsing
## scene spoting
## image parsing with svg code
## general qa
```
### Hugginface inference
```shell
python3 demo/demo_hf.py
```
<details>
<summary><b>Hugginface inference details</b></summary>
```python
import torch
from transformers import AutoModelForCausalLM, AutoProcessor, AutoTokenizer
from qwen_vl_utils import process_vision_info
from dots_ocr.utils import dict_promptmode_to_prompt
model_path = "./weights/DotsOCR_1_5"
model = AutoModelForCausalLM.from_pretrained(
model_path,
attn_implementation="flash_attention_2",
torch_dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True
)
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
image_path = "demo/demo_image1.jpg"
prompt = """Please output the layout information from the PDF image, including each layout element's bbox, its category, and the corresponding text content within the bbox.
1. Bbox format: [x1, y1, x2, y2]
2. Layout Categories: The possible categories are ['Caption', 'Footnote', 'Formula', 'List-item', 'Page-footer', 'Page-header', 'Picture', 'Section-header', 'Table', 'Text', 'Title'].
3. Text Extraction & Formatting Rules:
- Picture: For the 'Picture' category, the text field should be omitted.
- Formula: Format its text as LaTeX.
- Table: Format its text as HTML.
- All Others (Text, Title, etc.): Format their text as Markdown.
4. Constraints:
- The output text must be the original text from the image, with no translation.
- All layout elements must be sorted according to human reading order.
5. Final Output: The entire output must be a single JSON object.
"""
messages = [
{
"role": "user",
"content": [
{
"type": "image",
"image": image_path
},
{"type": "text", "text": prompt}
]
}
]
# Preparation for inference
text = processor.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
text=[text],
images=image_inputs,
videos=video_inputs,
padding=True,
return_tensors="pt",
)
inputs = inputs.to("cuda")
# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=24000)
generated_ids_trimmed = [
out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)
```
</details>
## 3. Document Parse
**Based on vLLM server**, you can parse an image or a pdf file using the following commands:
```bash
# Parse all layout info, both detection and recognition
# Parse a single image
python3 dots_ocr/parser.py demo/demo_image1.jpg
# Parse a single PDF
python3 dots_ocr/parser.py demo/demo_pdf1.pdf --num_thread 64 # try bigger num_threads for pdf with a large number of pages
# Layout detection only
python3 dots_ocr/parser.py demo/demo_image1.jpg --prompt prompt_layout_only_en
# Parse text only, except Page-header and Page-footer
python3 dots_ocr/parser.py demo/demo_image1.jpg --prompt prompt_ocr
```
<details>
<summary><b>Output Results</b></summary>
1. **Structured Layout Data** (`demo_image1.json`): A JSON file containing the detected layout elements, including their bounding boxes, categories, and extracted text.
2. **Processed Markdown File** (`demo_image1.md`): A Markdown file generated from the concatenated text of all detected cells.
* An additional version, `demo_image1_nohf.md`, is also provided, which excludes page headers and footers for compatibility with benchmarks like Omnidocbench and olmOCR-bench.
3. **Layout Visualization** (`demo_image1.jpg`): The original image with the detected layout bounding boxes drawn on it.
</details>
## 4. Demo
Have fun with the [live demo](https://dotsocr.xiaohongshu.com/).
### Examples for document parsing
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase/formula1.png" alt="formula1.png" border="0" />
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase/table3.png" alt="table3.png" border="0" />
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase/Tibetan.png" alt="Tibetan.png" border="0" />
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase/tradition_zh.png" alt="tradition_zh.png" border="0" />
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase/nl.png" alt="nl.png" border="0" />
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase/kannada.png" alt="kannada.png" border="0" />
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase/russian.png" alt="russian.png" border="0" />
### Examples for image parsing
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase_dots_ocr_1_5/svg_1.png" alt="svg_1.png" border="0" />
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase_dots_ocr_1_5/svg_2.png" alt="svg_2.png" border="0" />
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase_dots_ocr_1_5/svg_4.png" alt="svg_4.png" border="0" />
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase_dots_ocr_1_5/svg_5.png" alt="svg_5.png" border="0" />
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase_dots_ocr_1_5/svg_6.png" alt="svg_6.png" border="0" />
> **Note:**
> - Inferenced by dots.ocr-1.5-svg
### Example for web parsing
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase_dots_ocr_1_5/webpage_1.png" alt="webpage_1.png" border="0" />
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase_dots_ocr_1_5/webpage_2.png" alt="webpage_2.png" border="0" />
### Examples for scene spotting
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase_dots_ocr_1_5/scene_1.png" alt="scene_1.png" border="0" />
<img src="https://raw.githubusercontent.com/rednote-hilab/dots.ocr/master/assets/showcase_dots_ocr_1_5/scene_2.png" alt="scene_2.png" border="0" />
## Limitation & Future Work
- **Complex Document Elements:**
- **Table&Formula**: The extraction of complex tables and mathematical formulas persists as a difficult task given the model's compact architecture.
- **Picture**: We have adopted an SVG code representation for parsing structured graphics; however, the performance has yet to achieve the desired level of robustness.
- **Parsing Failures:** While we have reduced the rate of parsing failures compared to the previous version, these issues may still occur occasionally. We remain committed to further resolving these edge cases in future updates.
assets/showcase_dots_ocr_1_5/result/scene_1.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 1.9 MiB

assets/showcase_dots_ocr_1_5/result/scene_2.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 1.5 MiB

assets/showcase_dots_ocr_1_5/result/svg_1.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 783 KiB

assets/showcase_dots_ocr_1_5/result/svg_2.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 721 KiB

assets/showcase_dots_ocr_1_5/result/svg_4.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 753 KiB

assets/showcase_dots_ocr_1_5/result/svg_5.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 887 KiB

assets/showcase_dots_ocr_1_5/result/svg_6.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 946 KiB

assets/showcase_dots_ocr_1_5/result/webpage_1.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 1.0 MiB

assets/showcase_dots_ocr_1_5/result/webpage_2.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 921 KiB

requirements.txt
+2 -1
View File
@@ -7,5 +7,6 @@ qwen_vl_utils
transformers==4.51.3
huggingface_hub
modelscope
flash-attn==2.8.0.post2
# flash-attn==2.8.0.post2 # to speed up inference need flash-attn
accelerate
cairosvg
tools/download_model.py
+2 -2
View File
@@ -5,11 +5,11 @@ import os
if __name__ == '__main__':
parser = ArgumentParser()
parser.add_argument('--type', '-t', type=str, default="huggingface")
parser.add_argument('--name', '-n', type=str, default="rednote-hilab/dots.ocr")
parser.add_argument('--name', '-n', type=str, default="rednote-hilab/dots.ocr-1.5")
args = parser.parse_args()
script_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
print(f"Attention: The model save dir dots.ocr should be replace by a name without `.` like DotsOCR, util we merge our code to transformers.")
model_dir = os.path.join(script_dir, "weights/DotsOCR")
model_dir = os.path.join(script_dir, "weights/DotsOCR_1_5")
if not os.path.exists(model_dir):
os.makedirs(model_dir)
if args.type == "huggingface":
tools/elo_score_prompt.py
+89
View File
@@ -0,0 +1,89 @@
def construct_prompt(c1_text, c2_text):
"""
Constructs the complete Prompt sent to Gemini (English Version).
c1_text: Markdown text from Model 1
c2_text: Markdown text from Model 2
"""
prompt = f"""You are an expert in evaluating OCR content accuracy. Please compare the model outputs with the original image, focusing heavily on **content accuracy** while ignoring formatting and layout differences.
【Evaluation Focus - Focus ONLY on Content Accuracy】
1. **Text Accuracy**:
- Typos: Character recognition errors (e.g., "test" recognized as "tost").
- Omissions: Missing characters or words present in the original text.
- Hallucinations: Adding characters that do not exist in the original text.
2. **Table Accuracy**:
- Correctness of data and text within the table.
- Completeness of cell content.
- Correct row/column alignment.
3. **Formula Accuracy** (Evaluate based on):
- **Correctness**: Are mathematical symbols, variables, and operators preserved accurately?
- **Completeness**: Are all parts of the formula present without omission?
- **Semantic Equivalence**: Does the extracted formula convey the exact same mathematical meaning?
【Tie Judgment Criteria - Important】
You must judge as a **tie** in the following cases:
- Text content is identical, differing only in Markdown formatting.
- Table data is identical, differing only in Markdown table syntax.
- Formula content is semantically equivalent, differing only in LaTeX representation.
- Both models correctly identified the core content; minor differences do not affect information retrieval.
- Both models share the same minor errors or are both perfect.
- **Image/Figure processing differs** (one extracts text, one gives bbox, one ignores it), but the main text is accurate.
【Items to Ignore - Do NOT factor into scoring】
- Markdown formatting differences (e.g., `# Header` vs `## Header`, `*` vs `-` for lists).
- Layout and typesetting differences (newlines, indentation, alignment).
- Recognition differences in non-body text like Headers, Footers, and Page Numbers.
- Text wrapping and paragraph segmentation nuances.
- Table border styles (e.g., `|---|---|` vs `|:--|--:|`).
- Different but equivalent LaTeX representations for formulas.
- **Image/Figure Processing Differences (ABSOLUTELY IGNORE)**:
- How the model parses image/figure regions is **completely excluded** from the scoring standard.
- Whether it parses as a `figure` field, outputs bbox coordinates, extracts text inside the image, provides a caption, describes the image content, or **completely ignores/skips the image**, these are all considered equivalent.
- Do NOT declare a winner based on image handling.
【Model 1 Output】:
```markdown
{c1_text}
```
【Model 2 Output】:
```markdown
{c2_text}
```
【Evaluation Process】
1. Carefully compare the text content against the original image.
2. Identify errors, omissions, or additions in text recognition for both models.
3. Check the accuracy of table data.
4. Evaluate the correctness, completeness, and semantic equivalence of mathematical formulas.
5. **Ignore image regions**: Confirm that differences in image/figure parsing are not used for scoring.
6. Important: If the substance is the same and only the format differs, judge as a tie.
7. Only declare a winner if there is a significant difference in **content accuracy**.
【Examples of Ties】
- Model 1: "# Title", Model 2: "## Title" (Same content, different level).
- Model 1: "* Item", Model 2: "- Item" (Same content, different bullet).
- Formula: Model 1 "$x^2$", Model 2 "$x*x$" (Different LaTeX, same meaning).
- Table data is identical, but column alignment syntax differs.
- Identification is identical, but one model parsed the footer while the other didn't (Judge as Tie).
- **Image handling**: Model 1 outputs an image bbox, Model 2 outputs an image description, Model 3 ignores the image. As long as the main text is accurate, this is a **Tie**.
【Output Requirement】 Please strictly return the result in the following JSON format:
{{"winner": "tie", "reason": "Detailed explanation of the judgment, specifically noting the logic for a tie"}}
The value of "winner" must be one of:
- "1": Model 1 is clearly better in content accuracy.
- "2": Model 2 is clearly better in content accuracy.
- "tie": Both models perform equally in content accuracy (including cases of identical content but different formatting/image handling).
In the "reason" field, specifically explain:
- If a tie: Explain the consistency of the content and explicitly mention which formatting or image handling differences were ignored.
- If a winner: Specifically point out the accuracy differences (typos, missing words, table/formula errors).
- **Note**: It is better to judge a tie than to incorrectly determine a winner based on minor formatting or image parsing differences. **Content accuracy of the main text is the ONLY standard.**
"""
return prompt