SenBen: Sensitive Content Benchmark with 13,999 Movie Frames

13,999 frames sampled from 157 movies released between 1982 and 2023. The dataset is annotated with grounded scene graphs and 16 safety tags for evaluating vision-language models, created by author fcakyon and last updated on April 27, 2026.

Use Cases

• Benchmarking model safety on sensitive visual content based on the 16 MECD safety tags.
• Evaluating spatial reasoning capabilities based on the grounded scene graphs with objects, attributes, and predicates.
• Training or testing fine-grained content classifiers based on the multi-dimensional annotations beyond a simple safe/unsafe label.

Strengths

• Contains 13,999 annotated frames, providing a substantial evaluation corpus.
• Samples from 157 movies spanning over four decades (1982-2023), offering temporal diversity.
• Provides fine-grained, explainable annotations including scene graphs and 16 safety categories.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.
• Data may reflect temporal and cultural bias inherent to the selected movies from 1982-2023.

Provenance

Source

huggingface

Collection Method

Frames sampled from 157 movies and annotated with scene graphs and safety tags.

Time Range

1982 to 2023

Freshness

Last updated 2026-04-27 05:12:52; freshness should be verified.

Geography

null