V-RAGBench: 2,100 Query-Evidence-Answer Triplets for Video Retrieval-Augmented Generation

V-RAGBench is a benchmark dataset containing 2,100 open-ended query, evidence chunk, and answer triplets designed for evaluating retrieval and generation in long-video retrieval-augmented generation (VideoRAG). It was created by DISLab and last updated on June 15, 2026. The triplets are built from hour-scale egocentric videos, with queries designed to be answerable only from a specific localized evidence chunk.

Use Cases

• Benchmarking retrieval accuracy in long-video contexts based on localized evidence chunks.
• Evaluating the faithfulness of generation models in VideoRAG systems based on query-answer pairs.
• Training or fine-tuning models for open-ended question answering on egocentric video content.
• Studying causal dependencies between retrieval and generation in multimodal RAG pipelines.

Strengths

• Contains 2,100 structured triplets for systematic evaluation.
• Queries are designed for causal dependency on retrieval, enabling decoupled evaluation of RAG components.
• Built from hour-scale egocentric videos, providing a long-context multimodal challenge.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known (2,100), but other scale details like file size and formats are unknown.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

DISLab, via Hugging Face.

Collection Method

Built from hour-scale egocentric videos.

Freshness

Last updated 2026-06-15 04:56:16; freshness should be verified.