siglip-lora-osteosarcoma-base-v3

lfniederauer/siglip-lora-osteosarcoma-base-v3

siglip-lora-osteosarcoma-base-v3 is a machine learning model from lfniederauer. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for jax. The card lists the license as apache-2.0.

Weighted FedAvg of trainable params (LoRA adapters + head) from basev2 (roots, n=1707) and pkghpobest (PKG, n=544).

Downloads · 30 days

21% of all-time downloads

All-time downloads

Public

Repo size

1.6 MB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.png1.9 MB · 98%

At a glance

Library: jax
License: apache-2.0
Model type: siglip_lora_classifier
Access: Public
Created: Jul 13, 2026
Updated: Jul 13, 2026
SHA: 12e67026

Datasets

tatsuyaryu/OsteosarcomaHistopathologyClassification

Model card

From the Hugging Face model README

SigLIP-LoRA Osteosarcoma — base_v3 (federated merge)

Weighted FedAvg of trainable params (LoRA adapters + head) from base_v2 (roots, n=1707) and pkg_hpo_best (PKG, n=544).

Current federated checkpoint for the clean-contract research path. Supersedes historical fedavg-v1 for this story.

Model family

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fedavg-v1	Historical FL round (superseded)	siglip-lora-osteosarcoma-fedavg-v1

Holdout results (roots clean test, 410 patches)

Metric	base_v2	base_v3	Delta
Overall accuracy	91.22%	90.24%	−0.98 pp
Tumor sensitivity	0.973	0.973	0
Tumor → Non-Tumor misses	6	6	0 (preserved)
Non-Tumor → Tumor false alarms	26	31	+5 (more conservative)

PKG external proxy (137 decontaminated test patches)

Metric	base_v2	base_v3
Tumor sensitivity	0.955	0.985
Critical tumor→NT misses	5	1

Error gallery (research use / clinically sensitive)

The montage below embeds histopathology patches for failure-mode review. Research use only — not for clinical decision-making.

Notebooks: notebooks/base_v3_qualitative_research.ipynb, notebooks/base_client_pkg_qualitative_research.ipynb.

Model description

Federated Orbax bundle of trainable params only (LoRA + head) on frozen google/siglip-so400m-patch14-384. Not a transformers checkpoint.

Property	Value
Framework	JAX / Flax
Base model	`google/siglip-so400m-patch14-384`
Merge	`fedavg_weighted_by_n`
Contributors	clinic_00 (0.758), clinic_pkg (0.242)
n_total	2251
Image size / norm	384 / SigLIP
Roots holdout accuracy	90.24%
Roots tumor sensitivity	0.973

Classes

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

Reproduce the round

# PKG private HPO from roots, then weighted FedAvg
make pkg-round-from-roots ROUND_OUT=checkpoints/base_v3
# or stepwise:
make pkg-hpo-best
make merge CHECKPOINT_A=checkpoints/base_v2 CHECKPOINT_B=checkpoints/pkg_hpo_best OUT=checkpoints/base_v3
make eval-roots EVAL_CHECKPOINT=checkpoints/base_v3

Deploy / usage

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/base_v3 --image patch.jpg
python scripts/infer.py --hf-repo lfniederauer/siglip-lora-osteosarcoma-base-v3 --image patch.jpg

Publish:

make sync-hf-assets
make publish-hf-base-v3

Files in this repo

File	Description
`orbax_checkpoint/`	Merged trainable weights (LoRA + head)
`config.json`	Model + federated provenance
`preprocessor_config.json`	Size + norm
`training_metadata.json` / `round_metadata.json`	Merge weights + contributors
`assets/`	Confusion, safety, UMAP, error gallery, hygiene

Datasets

ID	Reference
Roots (clean)	Deduplicated `tatsuyaryu/OsteosarcomaHistopathologyClassification` as `client_00_clean`
PKG (clean)	TCIA Osteosarcoma Tumor Assessment as `client_pkg` (contamination-filtered)

Limitations

Research use only. Error gallery contains clinically sensitive histopathology pixels.
Slight roots accuracy drop vs base_v2 is driven by more Non-Tumor→Tumor false alarms, not by increased tumor misses.
Requires JAX + big_vision.

Citation

@misc{siglip_lora_osteosarcoma_base_v3,
  title={SigLIP-LoRA base_v3 Federated Osteosarcoma Histopathology Classifier},
  author={lfniederauer},
  year={2026},
  howpublished={\url{https://huggingface.co/lfniederauer/siglip-lora-osteosarcoma-base-v3}}
}

@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/}
}

Library: jax
Type: siglip_lora_classifier
License: apache-2.0
Created: Jul 13, 2026
Updated: Jul 13, 2026