Model identity
The small-model benchmark uses a 640×640 person detector and a 224×224 helmet classifier. The BM1688 and CV186X artifacts are byte-identical. RK3576 and RV1126B use platform-specific RKNN artifacts with the same public input/output contracts.
| Platform | Public model | Version | Format | SHA-256 | Size |
|---|---|---|---|---|---|
| BM1688 | YOLOv8n person detector | V1.0.0 | .nn | 56b207ef…b6c5c8 | 7,023,600 B |
| BM1688 | Helmet classifier | V1.0.0 | .nn | 33b0fb4b…cd7c8a | 6,001,416 B |
| CV186X | YOLOv8n person detector | V1.0.0 | .nn | 56b207ef…b6c5c8 | 7,023,600 B |
| CV186X | Helmet classifier | V1.0.0 | .nn | 33b0fb4b…cd7c8a | 6,001,416 B |
| RK3576 | YOLOv8 person detector | V1.0.0 | .rknn | 26ed82e0…541e0 | 6,305,107 B |
| RK3576 | Helmet classifier | V1.0.0 | .rknn | 471d1de3…d67ea | 3,024,125 B |
| RV1126B | YOLOv8 person detector | V1.0.0 | .rknn | db6ddca0…396cf | see platform JSON |
| RV1126B | Helmet classifier | V1.0.0 | .rknn | 9468f6a4…d9eeb | see platform JSON |
The no-safety-helmet workload is a two-stage detector-plus-classifier pipeline. Both stages receive the requested 24, 10, 7, or 5 FPS setting.
VLM identities remain in the BM1688, CV186X, and RK3576 platform JSON files. The formal validation also freezes each installed model, tokenizer, config, runtime, and candidate-package hash in results/vlm-observations.json. The RK3576 RKLLM model SHA-256 was recomputed after candidate installation. RV1126B is outside this VLM validation.
Model binaries are not redistributed by this benchmark. Full hashes and contracts are recorded in bm1688.json, cv186x.json, rk3576.json, and rv1126b.json.