The fine-tune was done for the object detection task. Specifically, this model aims to perform well at UI element detection, as part of a wider effort to enable our open-source toolkit for building agents at AgentSea.

Developed by: https://agentsea.ai/
Language(s) (NLP): en
Finetuned from model: https://huggingface.co/google/paligemma-3b-ft-widgetcap-448

Demo

You can find a demo for this model here.

Notes

The only task used in the fine-tune was the object detection task, so it might not perform well in other types of tasks.

Usage

To start using this model, run the following:

python

from transformers import AutoProcessor, PaliGemmaForConditionalGeneration

model = PaliGemmaForConditionalGeneration.from_pretrained("agentsea/paligemma-3b-ft-widgetcap-waveui-448").eval()
processor = AutoProcessor.from_pretrained("agentsea/paligemma-3b-ft-widgetcap-waveui-448")

Data

We used the WaveUI dataset for this fine-tune. Before using it, we preprocessed the data to use the Paligemma bounding-box format.

Evaluation

We calculated the mean IoU over 1024 examples of the test set using 3 different closed-source models: Gemini 1.5 Pro, Claude 3.5 Sonnet and GPT 4o. We also ran this same calculation using the PaliGemma WaveUI fine-tunes. We obtained the following values:

Gemini 1.5 Pro: 0.12
Claude 3.5 Sonnet: 0.05
GPT 4o: 0.05
PaliGemma Widgetcap+WaveUI 448: 0.40
PaliGemma WaveUI 896: 0.49