LongTVQA: 100K+ Video Question-Answer Pairs with Clip-Level Subtitles

LongTVQA contains between 100,000 and 1,000,000 question-answering pairs and clip-level subtitles for long-form video analysis. Released by longvideoagent in late 2025 (Arxiv 2512.20618), the dataset facilitates research into video-grounded dialogue and temporal retrieval.

Use Cases

• Temporal grounding of questions using the occur_clip field
• Long-form video QA evaluation using the LongTVQA_val.jsonl split
• Subtitle-based video segment retrieval using clip-level text

Strengths

• Scale of 100,000 to 1,000,000 records
• Granular clip-level indexing via occur_clip
• Permissive MIT license

Limitations

• Text-only modality in this repository requires external video files for multimodal tasks
• Domain-specific focus on TV show content

Provenance

Source

Arxiv 2512.20618

Freshness

Last updated February 2026.