autolens-dinov3-safe-weighted

AutoLens DINOv3 Safe Weighted

Pre-calibration HF model card for the class-weighted DINOv3 candidate.

Model

• Architecture: DINOv3 ViT-Small/16 (timm: vit_small_patch16_dinov3)
• Classes: 8 vehicle body types
• Source run: autolens-ai_checkpoints_dinov3_safe_weighted_aug_20260510_131319
• Source checkpoint: checkpoints/autolens-ai_checkpoints_dinov3_safe_weighted_aug_20260510_131319/best-29-0.9491.ckpt
• Export format: model.safetensors plus metadata.json, with ONNX for deployment
• Calibration status: not applied yet; this repo is the pre-calibration package and will be updated later with temperature scaling

Classes

• SUV
• VAN
• STATION WAGON
• MICRO
• OPEN WHEEL / F1
• SEDAN
• HATCHBACK
• PICK UP

Preprocessing

• RGB input
• Resize: 256
• Center crop / model input: 256 x 256
• Mean: [0.4429, 0.4354, 0.437]
• Std: [0.2456, 0.2421, 0.2449]

Validation Results

Validation was used to select the checkpoint. The metrics below come from the best validation epoch in metrics.csv.

• Best epoch: 29
• Accuracy: 0.9576
• F1-macro: 0.9491
• F1-weighted: 0.9579
• Precision-macro: 0.9504
• Recall-macro: 0.9486
• Loss: 0.7255

Internal Test Results

Evaluated on the held-out internal test set (3170 samples, no data leakage).

• Accuracy: 0.9584
• F1-macro: 0.9507
• F1-weighted: 0.9584
• Precision-macro: 0.9561
• Recall-macro: 0.9456

Per-class Performance

Training Curves

Internal Test Confusion Matrix

Artifact Sizes

• model.safetensors: 82.373 MB
• model.onnx: 82.572 MB

Files

• model.safetensors — weights-only model artifact
• metadata.json — architecture, classes, preprocessing, source run, and artifact metadata
• model.onnx — ONNX Runtime inference artifact
• metrics.csv — epoch-level training and validation history
• per_class_metrics.txt — internal-test classification report
• training_loss.png — training/validation loss curve
• training_accuracy.png — training/validation accuracy curve
• confusion_matrix.png — normalized internal-test confusion matrix
• size_check.json / size_report.json — artifact size and ONNX smoke evidence
• internal_test_results.json — structured internal-test summary

Intended use

Educational demo and report evidence for classifying uploaded vehicle images into the 8 AutoLens body-type classes.

Limitations

The dataset is assembled from public/open sources and may contain domain bias. Similar body styles such as hatchback, station wagon, and sedan can be ambiguous. This repo is intentionally pre-calibration; probability calibration will be added in the next upload after temperature scaling is fitted.