EmoBench-M: Benchmarking Emotional Intelligence for Multimodal Large Language Models

Dataset Description

EmoBench-M is a comprehensive benchmark designed to evaluate the Emotional Intelligence (EI) of Multimodal Large Language Models (MLLMs). It provides a challenging testbed for assessing a model's ability to understand and interpret human emotions from video, a critical step towards developing more empathetic and human-like AI systems.

The dataset consists of video clips featuring individuals expressing various emotions. Each video is paired with a conversational prompt that asks the model to determine the emotion conveyed. This structure pushes models to integrate visual, auditory, and textual information to make a correct assessment, moving beyond simple text-based sentiment analysis.

📦 Dataset

Each JSON file contains conversation-style prompts and labels aligned with the corresponding video clips. The structure looks like:

[
  {
    "id": "0",
    "video": "videos/ch-simsv2s/aqgy4_0004/00023.mp4",
    "conversations": [
      {
        "from": "human",
        "value": "<video>\nThe person in the video says: ... Determine the emotion conveyed..."
      },
      {
        "from": "gpt",
        "value": "negative"
      }
    ]
  }
]

📁 Dataset Structure

EmoBench-M/
├── benchmark_json/           # JSON files containing metadata and annotations for each dataset
│   ├── FGMSA.json    # Test instructions for the FGMSA dataset
│   ├── MC-EIU.json           # 500-sample test set for the MC-EIU dataset
│   ├── MELD.json     # Test instructions for the MELD dataset
│   ├── MOSEI.json            # 500-sample test set for the MOSEI dataset
│   ├── MOSI.json             # 500-sample test set for the MOSI dataset
│   ├── MUSTARD.json               # 500-sample test set for the MUSTARD dataset
│   ├── RAVDSS_song.json           # 500-sample test set for the RAVDSS song subset
│   ├── RAVDSS_speech.json         # 500-sample test set for the RAVDSS speech subset
│   ├── SIMS.json             # 500-sample test set for the SIMS dataset
│   ├── ch-simsv2s.json       # 500-sample test set for the Chinese SIMS v2s dataset
│   ├── funny.json    # Test instructions for the UR-FUNNY dataset
│   ├── mer2023.json # Test instructions for the MER2023 dataset
│   └── smile.json           # Test data for the SMILE dataset
└── dataset/              # Corresponding video files for each dataset
    ├── FGMSA/
    │   └── videos/
    │       └── FGMSA/        # Video files for the FGMSA dataset
    ├── MC-EIU/
    │   └── videos/
    │       └── MC-EIU/       # Video files for the MC-EIU dataset
    ├── MELD/
    │   └── videos/
    │       └── MELD/         # Video files for the MELD dataset
    ├── MOSEI/
    │   └── videos/
    │       └── MOSEI/        # Video files for the MOSEI dataset
    ├── MOSI/
    │   └── videos/
    │       └── MOSI/         # Video files for the MOSI dataset
    ├── MUSTARD/
    │   └── videos/
    │       └── MUSTARD/      # Video files for the MUSTARD dataset
    ├── RAVDSS_song/
    │   └── videos/
    │       └── RAVDSS/       # Video files for the RAVDSS song subset
    ├── RAVDSS_speech/
    │   └── videos/
    │       └── RAVDSS/       # Video files for the RAVDSS speech subset
    ├── SIMS_test/
    │   └── videos/
    │       └── SIMS/         # Video files for the SIMS dataset
    ├── ch-simsv2s/
    │   └── videos/
    │       └── ch-simsv2s/   # Video files for the Chinese SIMS v2s dataset
    ├── funny/
    │   └── videos/
    │       └── UR-FUNNY/     # Video files for the UR-FUNNY dataset
    ├── mer2023/
    │   └── videos/
    │       └── MER2023/      # Video files for the MER2023 dataset
    └── smile/
        └── videos/
            └── SMILE/       # Video files for the SMILE dataset

📂 Dtat Structure Overview

• benchmark_json/: Contains JSON files with metadata and annotations for each dataset, including test instructions and sample information.
• dataset/: Corresponding video files for each dataset, organized into subdirectories named after each dataset.