README.md

SHR Evaluation (Sentence-level Hallucination Ratio)

Introduction

SHR (Sentence-level Hallucination Ratio) is a fine-grained, diverse, and accurate evaluation benchmark of LVLM hallcuination on dense image descipription.

SHR evaluation can be divided into 3 steps:

1. Image Description: LVLM is prompted to describe image in detail.

2. GPT-4 evaluation: GPT-4 classify each sentence of model response into hallucination and non-hallucination.

3. Metric Calculation: The ratio of hallucinated sentences are calculated.

Model Evaluation

To evaluate your own LVLM on SHR, you can follow these steps:

1. Image Description

Prompt LVLM to describe validation images, validation images are stored in ha_dpo/data/shr/val_images_final.json.

Default prompt we use is Describe this image in detail. and generation parameters we use are as follows (adjust these settings according to you needs):

temperature	do_sample	num_beams
1	False	5

2. Response Preparation

Prepare model responses on validation images in following format:

{
    521: "This is an image of a house with a wooden siding and a green front door. There are several windows...",
    811: "The image shows a harbor with several boats docked...",
    ...
}

where key is the image_id and value is the corresponding model response.

3. Evaluation

Run following command to evaluate on SHR:

python ha_dpo/shr_eval/json_eval.py \
--api-key {openai_apikey} \
--vg-path ha_dpo/data/VG \
--shr-path ha_dpo/data/shr \
--json-file {path_to_json}

1. openai_apikey: the openai key. Begin with 'sk'
2. path_to_json: the json file to the model responses.

After evaluation is finished, results are saved in ha_dpo/models/minigpt4/shr_eval_results/{localtime}/metrics.json.

1. judgement.json: detailed judgements in SHR evaluation.
2. metrics.json: detailed metrics in SHR evaluation. mean_hal_ratio indicates the ration of hallucinated sentences, which is the main SHR result.