---
tags:
- image-feature-extraction
- timm
- vision-transformer
- moco-v3
- self-supervised-learning
- precision-at-scale
library_name: timm
license: cc-by-nc-4.0
datasets:
- jesusmolrdv/pas-bmini
---

# mocov3-vitb16-pas-birds-mini

MoCo-v3 ViT-B/16 backbone, self-supervised pretrained from scratch on the
`PaS-BMini` domain-specific dataset (Birds, 0.4M images) generated by
the Precision at Scale pipeline.

## Training data

Pretrained on [`jesusmolrdv/pas-bmini`](https://huggingface.co/datasets/jesusmolrdv/pas-bmini),
the Precision at Scale domain-specific dataset for this checkpoint.

## Paper

- **Paper (Pattern Recognition, 2026):** https://doi.org/10.1016/j.patcog.2025.112236
- **Official code:** https://github.com/jesusmolrdv/Precision-at-Scale

```bibtex
@article{rodriguezdevera2026precision,
  title   = {Precision at scale: Domain-specific datasets on-demand},
  author  = {Rodr{\'i}guez-de-Vera, Jes{\'u}s M. and Estepa, Imanol G. and Saras{\'u}a, Ignacio and Nagarajan, Bhalaji and Radeva, Petia},
  journal = {Pattern Recognition},
  volume  = {171},
  pages   = {112236},
  year    = {2026},
  publisher = {Elsevier},
  doi     = {10.1016/j.patcog.2025.112236}
}
```


<details>
<summary>arXiv preprint</summary>

https://arxiv.org/abs/2407.03463

</details>

## Usage

This backbone is stored in plain [`timm`](https://github.com/huggingface/pytorch-image-models)
format (`vit_base_patch16_224`, `num_classes=0`), so it loads directly
through the `timm` Hugging Face Hub integration:

```python
import timm

model = timm.create_model("hf_hub:jesusmolrdv/mocov3-vitb16-pas-birds-mini", pretrained=True)
model.eval()
```

`model` outputs 768-d backbone features (no classification head attached).