# Original model and dataset attribution

The packaged weights are supervised fine-tuned derivatives. They are not the original authors' unmodified checkpoints. Keep these credits when sharing the weights, data-derived example images or documentation.

## Original checkpoints

- B0: [Huyt/arabica-coffee-leaf-disease-efficientnet-b0](https://huggingface.co/Huyt/arabica-coffee-leaf-disease-efficientnet-b0/tree/252d26841543befab22880876f8ca91543230aab), pinned revision `252d26841543befab22880876f8ca91543230aab`.
- B1: [timm/efficientnet_b1.ft_in1k](https://huggingface.co/timm/efficientnet_b1.ft_in1k/tree/1d6ddfd0ad535646fdb05bc3834913d93816152e), pinned revision `1d6ddfd0ad535646fdb05bc3834913d93816152e`.
- B2: [timm/efficientnet_b2.ra_in1k](https://huggingface.co/timm/efficientnet_b2.ra_in1k/tree/3577c4a7d84723645311bb5a9e5086f1b62ec8e2), pinned revision `3577c4a7d84723645311bb5a9e5086f1b62ec8e2`.
- B3: [timm/efficientnet_b3.ra2_in1k](https://huggingface.co/timm/efficientnet_b3.ra2_in1k/tree/0366a75518620e0f2077789202073759f2d61393), pinned revision `0366a75518620e0f2077789202073759f2d61393`.

The original coffee-specific model used for B0 was published by **Huyt**. B1, B2 and B3 independently use the corresponding **timm** ImageNet checkpoints, rather than resizing or copying Huyt's B0 classifier. Credit timm / Ross Wightman and the EfficientNet authors, Mingxing Tan and Quoc V. Le.

## Dataset references

- Jepkoech, Jennifer; Mugo, David Muchangi; Kenduiywo, Benson K.; Too, Edna Chebet (2021). *Arabica coffee leaf images dataset for coffee leaf disease detection and classification*, Data in Brief 36, 107142. [AgML dataset mirror](https://huggingface.co/datasets/Project-AgML/arabica_coffee_leaf_disease_classification).
- Krohling, Renato A.; Esgario, Guilherme J. M.; Ventura, Jose A. (2019). *BRACOL*. [Mendeley Data V1](https://doi.org/10.17632/yy2k5y8mxg.1).
- Parraga-Alava, Jorge; Cusme, Kevin; Loor, Angelica; Santander, Esneider (2019). *RoCoLe*. [Mendeley Data V2](https://doi.org/10.17632/c5yvn32dzg.2).
- Rosas Ruiz, Luis Esteban; Salom Medina, Andrey Fernando; Barrero Perez, Jaime Guillermo (2026). *CoffeeLeaf-CO / DECAFIA v2*. [Versioned Zenodo record](https://zenodo.org/records/22756170); [concept DOI](https://doi.org/10.5281/zenodo.19931903).
- Silva, Lucas Brito; Carneiro, Alvaro Leandro Cavalcante; Faulin, Marisa Silveira Almeida Renaud (2020). *Rust and leaf miner in coffee crop*. [Mendeley Data V5](https://doi.org/10.17632/vfxf4trtcg.5). Used here through CoffeeLeaf-CO's redistributed/reannotated subset.
- AI Lab Makerere. *Beans*. [Hugging Face dataset](https://huggingface.co/datasets/AI-Lab-Makerere/beans). Non-coffee examples are mapped to unsupported.
- Santa Maria Pinedo, John Clark (2026). *Coffee leaf dataset by phytosanitary class*. [Mendeley Data V2](https://doi.org/10.17632/mfpxg4y65r.2).
- Jonatan Fragoso. *BRACOL for YOLO Detection*. [Reviewed annotations](https://www.kaggle.com/datasets/jonatanfragoso/bracol-for-yolov8-detection); [associated paper](https://www.mdpi.com/2076-3417/15/9/5040). Original photograph credit remains with BRACOL's authors.
- Aroca-Trujillo, Jorge Luis; Perez-Ruiz, Alexander (2025). *Colombian coffee tree leaves multispectral images dataset*, Data in Brief 59, 111421. [AgML mirror](https://huggingface.co/datasets/Project-AgML/coffee_rust_multispec_classification). External evaluation only; RGB bands used here.
- cchery2001 / PG26038 thesis project, EARTH University, Costa Rica. [Coffee leaf sample bank](https://huggingface.co/datasets/cchery2001/pg26038-coffee-leaf-samples). External evaluation only.

All prepared crops/label mappings, deduplication and split assignments are changes made for this project; counts and manifests are supplied separately from the publishers' original collection counts. The six-image external example sheet is deliberately selected from accepted successes and does not represent overall external performance.

## Architecture and library citations

```bibtex
@inproceedings{tan2019efficientnet,
  title={EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks},
  author={Tan, Mingxing and Le, Quoc V.},
  booktitle={International Conference on Machine Learning},
  pages={6105--6114}, year={2019}, publisher={PMLR},
  url={https://proceedings.mlr.press/v97/tan19a.html}
}
@misc{wightman2019timm,
  author={Ross Wightman}, title={PyTorch Image Models}, year={2019},
  doi={10.5281/zenodo.4414861},
  url={https://github.com/huggingface/pytorch-image-models}
}
```

## Inference code provenance

The portable `predict.py` implements the original project's canonical preprocessing and decision helpers: `training/export_mobile.py`, `training/calibrate_brightness.py`, `training/evaluate_additional.py` and `training/compare_model_runtime.py`. [Original project](https://github.com/Imhaohao/global-ai-hackathon). It supports the shipped exports and the selected PyTorch checkpoints; no reinforcement learning or new training is performed by this package.