Foundation-Model-Project

Mitigating Hallucination in Vision-Language Models

This project explores techniques for reducing hallucinations in Vision-Language Models (VLMs). We investigate supervised fine-tuning, Direct Preference Optimization (DPO), and feedback-guided self-revision to improve the factual accuracy and visual grounding of model responses.

Project Overview

Vision-Language Models can generate incorrect information that is not grounded in the provided image. This project investigates approaches inspired by Fact-RLHF and Volcano to reduce these hallucinations.

We experimented with:

• Supervised Fine-Tuning (SFT)
• Direct Preference Optimization (DPO)
• Fact-augmented self-guided revision
• BLIP-2 BLIP VQA (400M)
• LLaVA-OneVision( https://huggingface.co/llava-hf/llava-onevision-qwen2-0.5b-ov-hf)(0.5B)

Models and Techniques

Supervised Fine-Tuning

BLIP-2 was fine-tuned using the LLaVA-Instruct dataset. We trained on 5,000 visual question-answer pairs using LoRA adapters and float16 precision.

Direct Preference Optimization

We generated multiple responses to image-question pairs using BLIP-2 and used Qwen2-VL as a judge to construct preference pairs.

The resulting preference dataset was used to train BLIP-2 with the DPOTrainer from the Hugging Face TRL library.

Fact-Augmented Self-Revision

We implemented a feedback-guided revision process inspired by Volcano.

The model:

1. Generates an initial response.
2. Generates a caption containing factual information about the image.
3. Uses the factual information to critique and revise its response.
4. Produces a final answer grounded in the visual content.

We evaluated the approach both with and without captions.

Datasets

LLaVA-Human-Preference LLaVa Human Preference 10k

Used to generate preference data for DPO training.

The dataset contains approximately 10K questions about COCO imagesCOCO train2017.

MMHAL-Bench

MMHAL-Bench evaluates hallucinations in multimodal models using open-ended image-question pairs.

POPE

POPE evaluates object hallucination using yes/no questions about whether objects are present in an image.

Results

BLIP-2 DPO / SFT

Model	POPE Accuracy	MMHAL Avg. Score	Hallucination Rate
BLIP-2 Base	0.73	1.31	0.65
BLIP-2 SFT	0.61	1.35	0.68
BLIP-2 DPO	0.73	1.25	0.66

Self-Revision with Captions

Adding image captions improved performance on both POPE and MMHAL-Bench.

Model	Setting	POPE Accuracy
BLIP-2	Without Captions	0.31
BLIP-2	With Captions	0.47
LLaVA-OneVision	Without Captions	0.31
LLaVA-OneVision	With Captions	0.50

On MMHAL-Bench:

Model	Setting	Avg. Score	Hallucination Rate
BLIP-2	Without Captions	1.54	0.58
BLIP-2	With Captions	1.85	0.51
LLaVA-OneVision	Without Captions	1.80	0.52
LLaVA-OneVision	With Captions	2.30	0.50

Technologies

• Python
• PyTorch
• Hugging Face Transformers
• Hugging Face TRL
• BLIP-2
• LLaVA-OneVision
• Qwen2-VL
• LoRA
• Direct Preference Optimization (DPO)
• Supervised Fine-Tuning (SFT)