SF20K: 20,143 Amateur Short Films for Story-Level Video Understanding

SF20K is the largest publicly available movie dataset, containing 20,143 amateur films totaling 3,582 hours of video content. Each video lasts on average 11 minutes. The dataset includes a training set with synthetic questions and a test benchmark with manually curated questions, created by author rghermi and last updated on August 4, 2025.

Use Cases

• Train video-language models for story comprehension based on the 20,143 amateur films.
• Benchmark video question-answering systems using the manually curated test questions.
• Analyze narrative structures in short-form video content based on the story-level annotations.
• Develop models for temporal reasoning in videos based on the average 11-minute duration of clips.

Strengths

• Large scale with 20,143 unique video files.
• Substantial total duration of 3,582 hours of video content.
• Includes a dedicated test benchmark with manually curated questions for evaluation.

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.
• Data may reflect bias inherent to the source platform and the amateur film creators.

Provenance

Source

huggingface, author rghermi

Collection Method

Collection of amateur films, likely from online platforms; specific gathering method is not detailed.

Freshness

Last updated 2025-08-04 07:26:02; freshness should be verified.