MMDuet2 Data: Annotations for Training and Evaluating a Proactive Video MLLM

The dataset contains textual annotations for supervised fine-tuning, reinforcement learning, and evaluation of the MMDuet2 model, a video-based Multimodal Large Language Model. It was created by author wangyueqian and last updated on December 8, 2025. The associated videos must be downloaded separately from the source links.

Use Cases

• Supervised fine-tuning of a video MLLM based on the provided SFT annotations
• Reinforcement learning for improving model dialogue based on the RL annotations
• Benchmarking model performance on proactive interaction tasks using the evaluation annotations

Strengths

• Annotations are organized into three distinct folders (sft, rl, evaluate) for specific training and evaluation phases
• Dataset is directly linked to a published paper and GitHub codebase, providing research context

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download
• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment

Provenance

Source

huggingface

Collection Method

Created for the MMDuet2 research project, likely derived from video annotation processes.

Time Range

null

Freshness

Last updated 2025-12-08 09:08:07; freshness should be verified

Geography

null