---
base_model: google/paligemma-3b-ft-docvqa-896  
library_name: peft  
license: apache-2.0  
datasets:  
- cmarkea/table-vqa  
language:  
- fr  
- en  
pipeline_tag: visual-question-answering  
---

## Model Description

**paligemma-3b-ft-tablevqa-896-lora** is a fine-tuned version of the **[google/paligemma-3b-ft-docvqa-896](https://huggingface.co/google/paligemma-3b-ft-docvqa-896)** model,
trained specifically on the **[table-vqa](https://huggingface.co/datasets/cmarkea/table-vqa)** dataset published by Crédit Mutuel Arkéa. This model leverages the
**LoRA** (Low-Rank Adaptation) technique, which significantly reduces the computational complexity of fine-tuning while maintaining high performance. The model operates
in bfloat16 precision for efficiency, making it an ideal solution for resource-constrained environments.

This model is designed for multilingual environments (French and English) and excels in table-based visual question-answering (VQA) tasks. It is highly suitable for
extracting information from tables in documents, making it a strong candidate for applications in financial reporting, data analysis, or administrative document processing.
The model was fine-tuned over a span of 7 days using a single A100 40GB GPU.

## Key Features

- **Language:** Multilingual capabilities, optimized for French and English.
- **Model Type:** Multi-modal (image-text-to-text).
- **Precision:** bfloat16 for resource efficiency.
- **Training Duration:** 7 days on A100 40GB GPU.
- **Fine-Tuning Method:** LoRA (Low-Rank Adaptation).
- **Domain:** Table-based visual question answering.

## Model Architecture

This model was built on top of **[google/paligemma-3b-ft-docvqa-896](https://huggingface.co/google/paligemma-3b-ft-docvqa-896)**, using its pre-trained multi-modal
capabilities to process both text and images (e.g., document tables). LoRA was applied to reduce the size and complexity of fine-tuning while preserving accuracy,
allowing the model to excel in specific tasks such as table understanding and VQA.

## Usage

You can use this model for visual question answering with table-based data by following the steps below:

```python
from transformers import AutoProcessor, PaliGemmaForConditionalGeneration
from PIL import Image
import requests
import torch
device = "cuda" if torch.cuda.is_available() else "cpu"

model_id = "cmarkea/paligemma-3b-ft-tablevqa-896-lora"

# Sample image for inference
url = "https://datasets-server.huggingface.co/cached-assets/cmarkea/table-vqa/--/c26968da3346f92ab6bfc5fec85592f8250e23f5/--/default/train/22/image/image.jpg?Expires=1728915081&Signature=Zkrd9ZWt5b9XtY0UFrgfrTuqo58DHWIJ00ZwXAymmL-mrwqnWWmiwUPelYOOjPZZdlP7gAvt96M1PKeg9a2TFm7hDrnnRAEO~W89li~AKU2apA81M6AZgwMCxc2A0xBe6rnCPQumiCGD7IsFnFVwcxkgMQXyNEL7bEem6cT0Cief9DkURUDCC-kheQY1hhkiqLLUt3ITs6o2KwPdW97EAQ0~VBK1cERgABKXnzPfAImnvjw7L-5ZXCcMJLrvuxwgOQ~DYPs456ZVxQLbTxuDwlxvNbpSKoqoAQv0CskuQwTFCq2b5MOkCCp9zoqYJxhUhJ-aI3lhyIAjmnsL4bhe6A__&Key-Pair-Id=K3EI6M078Z3AC3"
image = Image.open(requests.get(url, stream=True).raw)

# Load the fine-tuned model and processor
model = PaliGemmaForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map=device,
).eval()

processor = AutoProcessor.from_pretrained("google/paligemma-3b-ft-docvqa-896")

# Input prompt for table VQA
prompt = "How many rows are in this table?"
model_inputs = processor(text=prompt, images=image, return_tensors="pt").to(model.device)

# Generate the answer
input_len = model_inputs["input_ids"].shape[-1]
with torch.inference_mode():
    generation = model.generate(**model_inputs, max_new_tokens=100, do_sample=False)
    generation = generation[0][input_len:]
    decoded = processor.decode(generation, skip_special_tokens=True)
    print(decoded)
```


## Performance

The model's performance was evaluated on 200 question-answer pairs, extracted from 100 tables from the test set of the
**[table-vqa](https://huggingface.co/datasets/cmarkea/table-vqa)** dataset. For each table, two pairs were selected: one in French and the other in English.

To evaluate the model’s responses, the **[LLM-as-Juries](https://arxiv.org/abs/2404.18796)** framework was employed using three judge models (GPT-4o, Gemini1.5 Pro,
and Claude 3.5-Sonnet). The evaluation was based on a scale from 0 to 5, tailored to the VQA context, ensuring accurate judgment of the model’s performance.

Here’s a visualization of the results:

![constellation](https://i.postimg.cc/y6tkYg9F/constellation-03.png)

In comparison, this model outperforms **[HuggingFaceM4/Idefics3-8B-Llama3](https://huggingface.co/HuggingFaceM4/Idefics3-8B-Llama3)** in terms of accuracy and efficiency,
despite having a smaller parameter size.


## Citation

```bibtex
@online{AgDePaligemmaTabVQA,
  AUTHOR = {Tom Agonnoude, Cyrile Delestre},
  URL = {https://huggingface.co/cmarkea/paligemma-tablevqa-896-lora},
  YEAR = {2024},
  KEYWORDS = {Multimodal, VQA, Table Understanding, LoRA},
}