Datasets:
Create README.md
Browse files# SparseTable Bench Dataset
SparseTable Bench Dataset is a multimodal table understanding dataset for table OCR, table structure recognition, and cell-level localization.
Each sample contains a table image, an instruction-style conversation, and the expected response. The response is a JSON string with two fields:
- `html`: the table converted into HTML format
- `cells`: cell-level text content and bounding box coordinates
## Data Format
The dataset is stored in JSONL format. Each line is one sample.
Example:
```json
{
"messages": [
{
"role": "user",
"content": "<image>\nPlease perform the following two tasks for this image:\n1. Convert the table in the image into HTML format code.\n2. Identify and extract the text content and its bounding box coordinates for each cell in the table.\n\nPlease combine the final results into a single JSON object and return it in the following format:\n{\n \"html\": \"...\",\n \"cells\": [{\"cell\": \"...\", \"bbox\": [...]}, ...]\n}"
},
{
"role": "assistant",
"content": "{\"html\": \"<table>...</table>\", \"cells\": [{\"cell\": \"<b>Variable</b>\", \"bbox\": [0, 6, 31, 14]}]}"
}
],
"images": [
"train_images_10000/PMC4292575_004_00.png"
],
"system": "You are a multi-task OCR recognition expert. Your task is to accurately extract information from images and output it in the specified format. You need to accomplish two things simultaneously: first, structure the table into HTML code, and second, detect and output the text and precise coordinates of each cell."
}