svod-yolo26x

svod-yolo26x

A verbatim conversion of Ultralytics/YOLO26's yolo26x.pt (the original COCO checkpoint, 80 classes) into safetensors, in the layout the svod YOLO loader reads.

This repo exists so svod's parity test can fetch weights without needing PyTorch: the test compares svod's forward pass against a PyTorch-generated golden, and only the weights are worth publishing — the golden is regenerated locally by scripts/convert_yolo.py.

Contents

Keys keep PyTorch's model. prefix; svod strips it on load. The integer num_batches_tracked buffers are dropped, since svod's inference path has no use for them. Both one2one and the training-only one2many head branches are retained, so the file is a faithful conversion rather than an inference-only subset.

Sibling repos

svod's loader always reads model.safetensors from the repo root, so each converted variant gets its own repo, named for the family plus a scale/format suffix. For YOLO26 detection:

Other task heads follow the same prefix with their own suffix, e.g. mexus/svod-yolo26x-seg, mexus/svod-yolo26x-pose.

Provenance

Verified equal to the source checkpoint tensor-for-tensor (914/914 identical) and reproducible — regenerating from the pinned checkpoint and image reproduces both the above digest and the parity golden exactly.

To check a copy:

python scripts/verify_yolo_weights.py --hub mexus/svod-yolo26x

Usage

cargo run -p svod-model --release --example yolo_detect -- \
    --hf-id mexus/svod-yolo26x --scale xlarge --classes 80

let model = Yolo26Detect::from_hub("mexus/svod-yolo26x", YoloConfig::new(YoloScale::XLarge, 80))?;

License

The weights derive from Ultralytics YOLO26, which is AGPL-3.0; that license carries over to this conversion. No source code from Ultralytics is included here.