0xBatuhan4/yoldost-street-context-v2

YolDost Street Activity Context (v2)

Auxiliary ViT classifier head for weak street activity/environment context on anonymized Mapillary derivatives. This is not the production route-scoring model and does not count people or predict crime.

Model card summary

• Run ID: 20260606T125349Z-modal-scene
• Base backbone: google/vit-base-patch16-224 @ 3f49326eb077187dfe1c2a2bb15fbd74e6ab91e3
• Dataset: Reubencf/streetview-global (CC-BY-SA-4.0)
• Training samples: 240 (anonymized locally before Modal)
• Weak-label context agreement: 88.33%
• Macro F1: 0.8833
• Checkpoint SHA-256: 2da0121e5f6e929237e45da72a25f4e0f7be2a11cc352af326fac557730c2e10

Files

• classifier-head.pt — fine-tuned classification head checkpoint from Modal run
• metrics.json — full reproducibility metadata
• training_config.json — compact training summary

KVKK / ethics

Trained only on locally face/plate anonymized derivatives. No person counting, identity features, or raw imagery in this repository.

Reproduce Modal run

See reports/runs/20260606T125349Z-modal-scene/report.md in the hackathon repository.