[ICLR 2025] MMRole: A Comprehensive Framework for Developing and Evaluating Multimodal Role-Playing Agents

Multimodal Role-Playing Agents (MRPAs) are designed to emulate specific characters and engage in dialogues centered around images, with either human users or other characters. MMRole is a comprehensive framework for developing and evaluating MRPAs, which comprises a personalized multimodal dataset and a robust evaluation method. Specifically, we construct a large-scale, high-quality dataset, MMRole-Data, consisting of 85 characters, 11K images, and 14K single or multi-turn dialogues. Additionally, we present a robust evaluation method, MMRole-Eval, encompassing eight metrics across three dimensions, where a reward model is trained to score MRPAs with the constructed ground-truth data for comparison. Please refer to our paper for more details (https://arxiv.org/abs/2408.04203).

MMRole-Data: Multimodal Role-Playing Dataset

Statistically, the MMRole-Data dataset comprises 85 characters, 11,032 images, and 14,346 dialogues, yielding 85,456 training samples and 294 test samples.

MMRole-Eval: Multimodal Role-Playing Evaluation

We propose MMRole-Eval, a robust evaluation method to stably and comprehensively assess MRPAs, which comprises eight metrics across three dimensions: fundamental conversational skills, multimodal understanding abilities, and role-playing qualities. For each metric, a specialized reward model initially conduct a brief qualitative assessment of the relative performance between the evaluated MRPA and the constructed ground-truth data, followed by assigning a quantitative score pair. The final score of the MRPA is the ratio of the two scores within the score pair. To develop the reward model, we employ GPT-4 to assess various MRPAs and leverage the evaluation trajectories to train our reward model.

We develop the first specialized MRPA, MMRole-Agent, using the training data of MMRole-Data. Extensive evaluation results demonstrate the improved performance of MMRole-Agent, and highlight the primary challenges in developing MRPAs, emphasizing the need for enhanced multimodal understanding and role-playing consistency.

Data Download

All data in MMRole-Data, including character profiles, images, dialogues, and formatted instruction-following data, and the training and validation data for the reward model in MMRole-Eval can be downloaded from MMRole_dataset. Please download and save them in the root directory, excluding the

.md

/COCO/train2017

MMRole_dataset
├── data
│   ├── test
│   │   ├── in-distribution
│   │   └── out-of-distribution
│   └── train
│       └── train_85k.json
├── dialogues
├── images
│   ├── annotations.json
│   ├── COCO
│   │   └── train2017
│   ├── Harry_Potter
│   └── ...
├── profiles
└── RM_data
    ├── test
    └── train
        └── RM-train_23k.json

Environment Requirements

The environment requirements are consistent with QWen-VL-Chat.

• python 3.8 and above
• pytorch 1.12 and above, 2.0 and above are recommended
• CUDA 11.4 and above are recommended (this is for GPU users)

pip install -r requirements.txt

Inference of MMRole-Agent (Answering for MMRole-Eval)

Please download the model weights of MMRole-Agent from MMRole-Agent and save them in the

Run the following script to utilize MMRole-Agent for generating answers for the In-Test and Out-Test test sets in MMRole-Eval:

bash inference.sh

Training of MMRole-Agent

To develop MMRole-Agent, you need to first download the model weights of QWen-VL-Chat from QWen-VL-Chat, then run the following script for fine-tuning:

bash finetune/finetune_ds.sh

Inference of the reward model in MMRole-Eval (Scoring for MMRole-Eval)

Please download the model weights of the reward model in MMRole-Eval from MMRole-EvalRM and save them in the

bash eval/RM_review.sh bash eval/RM_result.sh

bash finetune/finetuneRMds.sh