File size: 2,819 Bytes
494dcce
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
# MMCBench Dataset: Benchmarking Dataset for Multimodal Model Evaluation πŸš€

## Overview

The MMCBench Dataset is a curated collection of data designed for the comprehensive evaluation of Large Multimodal Models (LMMs) under common corruption scenarios. This dataset supports the MMCBench framework, focusing on cross-modal interactions involving text, image, and speech. It provides essential data for generative tasks such as text-to-image, image-to-text, text-to-speech, and speech-to-text, enabling robustness and self-consistency assessments of LMMs.

## Dataset Composition πŸ“Š

The MMCBench Dataset is structured to facilitate the evaluation across four key generative tasks:

- **Text-to-Image:** A collection of text descriptions with their corresponding corrupted versions and associated images.
- **Image-to-Text:** A set of images with clean and corrupted captions.
- **Text-to-Speech:** Text inputs with their clean and corrupted audio outputs.
- **Speech-to-Text:** Audio files with transcriptions before and after audio corruptions.

Each subset of the dataset has been meticulously selected and processed to represent challenging scenarios for LMMs.

## Using the Dataset πŸ› οΈ

To use the MMCBench Dataset for model evaluation:

1. **Access the Data**: The dataset is hosted on Hugging Face and can be accessed using their dataset library or direct download.
2. **Select the Task**: Choose from text-to-image, image-to-text, text-to-speech, or speech-to-text tasks based on your model's capabilities.
3. **Apply the Benchmark**: Utilize the data for each task to test your model's performance against various corruptions. Follow the [MMCBench](https://github.com/sail-sg/MMCBench/tree/main) framework for a consistent and standardized evaluation.

### Dataset Structure πŸ“

The dataset is organized into four main directories, each corresponding to one of the generative tasks:

- `text2image/`: Contains text inputs and associated images.
- `image2text/`: Comprises images and their descriptive captions.
- `text2speech/`: Includes text inputs and generated speech outputs.
- `speech2text/`: Contains audio files and their transcriptions.

## Contributing to the Dataset 🀝

Contributions to the MMCBench Dataset are welcome. If you have suggestions for additional data or improvements, please reach out through the Hugging Face platform or directly contribute via GitHub.

## License πŸ“œ

The MMCBench Dataset is made available under the Apache 2.0 License, ensuring open and ethical use for research and development.

## Acknowledgments and Citations πŸ“š

When using the MMCBench Dataset in your research, please cite it appropriately. We extend our gratitude to all contributors and collaborators who have enriched this dataset, making it a valuable resource for the AI and ML community.