---
library_name: transformers
tags:
- robotics
license: mit
datasets:
- ACIDE/user-vlm-pt
language:
- en
base_model:
- google/paligemma2-10b-pt-896
pipeline_tag: image-text-to-text
---
# User-VLM 360°
![Architecture](result-final.pdf)

## Overview
**User-VLM 360°** is a series of personalized Vision-Language Models (VLMs) designed for social human-robot interactions. The model introduces **User-aware tuning**, addressing the **semantic gap** that arises from the misalignment between user queries and the observed scene as captured by a robot's camera. Unlike traditional instruction tuning, which introduces latency and reduces performance, **User-VLM 360°** enables **real-time, robust adaptation** in dynamic robotic environments by inherently aligning cross-modal user representations.

This model allows for **customization of open-weight VLMs** to produce **personalized responses** based on demographic attributes such as age, gender, emotion, and ethnicity while maintaining ethical and safety considerations.

## Training Details
**Base Model:** User-VLM 360° is built on **PaliGemma 2**, which consists of a **SigLIP vision encoder** and **Gemma 2 as the language model**.

![Deployment on Pepper](pepper2.pdf)

### Fine-tuning Process:
1. **Base Model Tuning:**
   - Tuned the **MLP layer** to provide **user and scene descriptions** over **1 epoch**.
2. **Instruction Model Tuning:**
   - Instruction-tuned the **base model** using **personalized, user-specific Q&A datasets**.
   - Used **Sparse Mixture of LoRA Experts (MoLE)** (3 LoRA modules, rank=16, alpha=32, one chosen) and a standalone **LoRA (rank=16, alpha=32)** over **2 epochs**.
3. **Bias Mitigation:**
   - Applied **Direct Preference Optimization (DPO)** over **1 epoch** using **LoRA (rank=16, alpha=32)**.

## Model Usage
### Example Code:
```python
from transformers import PaliGemmaProcessor, PaliGemmaForConditionalGeneration
import torch

model_id = "ACIDE/User-VLM-10B-Instruct"
processor = PaliGemmaProcessor.from_pretrained(model_id)
model = PaliGemmaForConditionalGeneration.from_pretrained(model_id, torch_dtype=torch.bfloat16).to(device)

def generate_response(question, image, model, processor):
    prompt = f"<image> <|im_start|>USER: {question}<|im_end|> ASSISTANT:"
    model_inputs = processor(text=prompt, images=image, return_tensors="pt").to(torch.bfloat16).to(model.device)
    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)
        return decoded

# Example usage
from transformers.image_utils import load_image
url = "https://media.istockphoto.com/id/1282695693/photo/little-boy-sitting-on-chair-at-the-table.jpg"
image = load_image(url)
question = "Does Santa Claus exist?"
answer = generate_response(question, image, model, processor)
print(answer)
```

## Ethical Considerations & Limitations
- **Research-Only Use:** This model is intended strictly for **research purposes** and should not be deployed in real-world applications without further ethical validation.
- **Demographic Personalization:** While the model can adapt responses based on user attributes, **care must be taken to prevent bias and discrimination**.
- **No Liability:** The authors **do not accept any liability** regarding the use of this model. Responsibility for ethical and appropriate use remains with the users.

## Citation
If you use this model in your research, please cite the following papers:
```bibtex
@article{rahimi2025user,
  title={User-VLM: LLM Contextualization with Multimodal Pre-trained User Models},
  author={Rahimi, Hamed and Abrini, Mouad and Khoramshahi, Mahdi and Chetouani, Mohamed},
  year={2025}
}

@article{rahimi2025user,
  title={User-VLM 360°: Personalized Vision Language Models with User-aware Tuning for Social Human Robot Interactions},
  author={Rahimi, Hamed and Bhaj, Adil, Abrini, Mouad, Khoramshahi, Mahdi, Ghogho, Mounir, and Chetouani, Mohamed},
  year={2025}
}
```

## License
This model is licensed under the **MIT License**.

## Contact
For any questions or issues regarding the model, please open an issue on the repository or contact the maintainers directly.