FineVideo: 10,000 to 100,000 Video-Text Pairs for Multimodal VQA

HuggingFaceFV released FineVideo in late 2024, a multimodal collection containing between 10,000 and 100,000 video-text records. The dataset is specifically formatted for English-language tasks such as Visual Question Answering and video-to-text generation.

Use Cases

• Training Visual Question Answering models using the video and text modalities to answer scene-specific queries
• Developing Video Text-to-Text systems to generate natural language descriptions from video input
• Fine-tuning multimodal models on the 10,000+ English-language records for video understanding

Strengths

• Scale of 10,000 to 100,000 records
• CC-By license for open redistribution and research
• Parquet format for efficient I/O with Polars and Dask libraries

Limitations

• Geographic bias toward the US region as indicated by metadata tags
• Restricted to English language content

Provenance

Source

HuggingFaceFV

Freshness

Last updated December 2024

Geography

United States