Private Optuna HPO on the decontaminated PKG clinic (client_pkg), initialized from base_v2. This is the client update that is later FedAvg-merged into base_v3.

Only LoRA adapters + head leave the site. Raw PKG images never leave the client in the federated story.

Model family

Checkpoint	Role	Hugging Face
base_v2	Roots single-site HPO best (init)	siglip-lora-osteosarcoma-base-v2
pkg_hpo_best (this repo)	PKG private Optuna update	siglip-lora-osteosarcoma-pkg-hpo-best
base_v3	Federated merge (roots + PKG)	siglip-lora-osteosarcoma-base-v3
fedavg-v1	Historical FL round (superseded by base_v3)	siglip-lora-osteosarcoma-fedavg-v1

Results

Eval	Metric	Value
PKG local test (training_metadata)	Accuracy	94.89%
Roots clean holdout (410, transfer)	Accuracy	~90.2%
Roots clean holdout	Tumor sensitivity	0.982

UMAP roots vs PKG embeddings (family context)

Analysis notebooks: notebooks/pkg_hpo_best_qualitative_research.ipynb, notebooks/base_client_pkg_qualitative_research.ipynb.

Model description

Same architecture as base_v2: frozen SigLIP-SO400M + LoRA (r=8, α=16) on attention Q/V + linear head. Not a transformers checkpoint.

Property	Value
Framework	JAX / Flax
Base model	`google/siglip-so400m-patch14-384`
Init checkpoint	`checkpoints/base_v2`
Client	`data/client_pkg` (contamination-filtered vs roots test)
n_train (FedAvg weight)	544
Image size / norm	384 / SigLIP
PKG test accuracy (metadata)	94.89%

Classes

Label	ID
Non-Tumor	0
Non-Viable-Tumor	1
Viable	2

Training (PKG private Optuna)

Init from roots base_v2
Selected: lr≈5.63e-4, epochs=7, color_jitter≈0.019, max_shift=0, scale [0.752, 1.213], hflip=True
Fixed: batch=1 + grad accum 4, LoRA r=8/a=16, image=384
Seed: 1122

bash

make pkg-hpo-best
# or: make pkg-hpo-from-roots

Deploy / usage

bash

git clone https://github.com/lfniederauer/FederatedLoRA-OsteosarcomaClassification
cd FederatedLoRA-OsteosarcomaClassification
pip install -r requirements.txt
export BIG_VISION_ROOT=~/git/big_vision

python scripts/infer.py --checkpoint checkpoints/pkg_hpo_best --image patch.jpg
python scripts/infer.py --hf-repo lfniederauer/siglip-lora-osteosarcoma-pkg-hpo-best --image patch.jpg

Publish:

bash

make sync-hf-assets
make publish-hf-pkg

Merge into the global federated checkpoint:

bash

make merge CHECKPOINT_A=checkpoints/base_v2 CHECKPOINT_B=checkpoints/pkg_hpo_best OUT=checkpoints/base_v3
# published as lfniederauer/siglip-lora-osteosarcoma-base-v3

Files in this repo

File	Description
`orbax_checkpoint/`	Trainable weights (LoRA + head)
`config.json`	Model + HPO provenance
`preprocessor_config.json`	Size + norm
`training_metadata.json`	Hyperparams + test_accuracy
`assets/`	Family figures (UMAP, data hygiene)

Datasets

ID	Reference
PKG (clean)	TCIA Osteosarcoma Tumor Assessment, ingested as `client_pkg` with near-duplicates vs sacred roots test removed (`h≤5`)
DS_PRIMARY	`tatsuyaryu/OsteosarcomaHistopathologyClassification` (roots / init)

Limitations

Research use only.
PKG local accuracy is on the client's own held-out split; roots transfer metrics are for safety auditing, not cherry-picking.
Requires JAX + big_vision.

Citation

bibtex

@misc{siglip_lora_osteosarcoma_pkg_hpo_best,
  title={SigLIP-LoRA PKG Client HPO Update for Osteosarcoma Histopathology},
  author={lfniederauer},
  year={2026},
  howpublished={\url{https://huggingface.co/lfniederauer/siglip-lora-osteosarcoma-pkg-hpo-best}}
}

@dataset{leavey2019osteosarcoma_tcia,
  author    = {Leavey, Patrick and Sengupta, Aniruddha and Rakheja, Dinesh and Daescu, Ovidiu and Arunachalam, Harish Babu and Mishra, Rashika},
  title     = {Osteosarcoma data from {UT Southwestern}/{UT Dallas} for Viable and Necrotic Tumor Assessment ({Osteosarcoma-Tumor-Assessment})},
  year      = {2019},
  publisher = {The Cancer Imaging Archive},
  doi       = {10.7937/tcia.2019.bvhjhdas},
  url       = {https://www.cancerimagingarchive.net/collection/osteosarcoma-tumor-assessment/}
}