data/results/hallucination_i2t_summary.json

{
    "llava-v1.6-vicuna-7b-hf": {
        "Natural Selection": 16.1,
        "Distraction": 59.5,
        "Counterfactual Reasoning": 19.9,
        "Co-occurrence": 54.3,
        "Misleading Prompts": 34.2,
        "OCR": 14.4,
        "Average": 33.1
    },
    "gpt-4-vision-preview": {
        "Natural Selection": 23.3,
        "Distraction": 54.4,
        "Counterfactual Reasoning": 45.9,
        "Co-occurrence": 60.5,
        "Misleading Prompts": 52.2,
        "OCR": 26.2,
        "Average": 43.8
    },
    "gpt-4o-2024-05-13": {
        "Natural Selection": 25.3,
        "Distraction": 57.8,
        "Counterfactual Reasoning": 50.7,
        "Co-occurrence": 62.8,
        "Misleading Prompts": 43.2,
        "OCR": 36.8,
        "Average": 46.1
    }
}