SAVANT-CODALM Large: 10,000 Real-World Driving Images for Semantic Anomaly Detection

SAVANT is a model-agnostic reasoning framework that reformulates anomaly detection in autonomous driving as a layered semantic consistency verification. Approximately 10,000 real-world driving images were curated to address the challenge of detecting rare, out-of-distribution semantic anomalies. The dataset, created by Brusnicki, includes structured scene descriptions and multi-modal evaluations.

Use Cases

• Training semantic anomaly detection models based on real-world driving images.
• Benchmarking model-agnostic reasoning frameworks for autonomous driving safety.
• Evaluating multi-modal scene understanding systems based on structured scene descriptions.
• Researching out-of-distribution detection in complex, dynamic environments.

Strengths

• Approximately 10,000 real-world driving images provide a substantial base for model training.
• Includes structured scene descriptions and multi-modal evaluations for layered analysis.
• Specifically curated to address the challenge of detecting rare, out-of-distribution semantic anomalies.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect geographic/temporal/source bias inherent to its collection on Hugging Face.

Provenance

Source

Hugging Face, uploaded by Brusnicki.

Collection Method

Curated collection of real-world driving images.

Freshness

Last updated 2026-05-23 13:47:15; freshness should be verified.