DrishtiTable
Table Structure Recognition
Upload a table image, get HTML structure back. Powered by the fine-tuned DrishtiTable model.
DrishtiTable is a fine-tuned Qwen2.5-VL-7B model that converts table images into structured HTML. Trained on 1,141 Indian academic textbook tables, it achieves 83.2% TEDS — outperforming GPT-4o (71.1%) by +12.1 points.
| Model | TEDS Score | Improvement |
|---|---|---|
| o4-mini (OpenAI) | 61.4% | — |
| GPT-4.1 (OpenAI) | 68.0% | — |
| GPT-4o (OpenAI) | 71.1% | — |
| DrishtiTable (This Demo) | 83.2% | +12.1 over GPT-4o |
Upload any table image below to try it. First run takes ~60s to load the model, subsequent runs are faster.
Rendered Table Preview
Run Locally (Fastest)
from unsloth import FastVisionModel
model, tokenizer = FastVisionModel.from_pretrained(
"Nalandadata/DrishtiTable-Qwen2.5-VL-7B",
max_seq_length=4096, load_in_4bit=True,
)
FastVisionModel.for_inference(model)
Resources
| Resource | Link |
|---|---|
| Fine-tuned Model | Nalandadata/DrishtiTable-Qwen2.5-VL-7B |
| Dataset (sample) | Nalandadata/DrishtiTable |
| Base Model | Qwen/Qwen2.5-VL-7B-Instruct |
Built by Nalanda Data. DrishtiTable (Sanskrit: drishti = vision).