msfit-dinobloom-v4-mlp-s3

msfit DINOBloom-base MLP S3

This repository contains an exported msfit checkpoint for 13-class white blood cell classification on WBCBench 2026. Code can be found here: https://github.com/Antony-gitau/msfit

Architecture

• Backbone: dinobloom_base
• Head: mlp
• Image size: 384

Included Files

• model.safetensors: weights-only export for inference
• config.json: model architecture, labels, saved args, and metrics
• README.md: model card and usage examples
• submission_dinobloom_v4_mlp_s3_tta.csv: reference test-set submission CSV associated with this checkpoint

Metrics

• Eval macro-F1: 0.720975839094881
• Eval composite: 0.6182056067568553
• Eval tail mean F1: 0.5154353744188297
• Eval balanced accuracy: 0.7116216524853729
• Leaderboard score: 0.67584

Usage

Python

from msfit.pretrained import load_msfit_model

model, info = load_msfit_model("AntonyG/msfit-dinobloom-v4-mlp-s3", device="cpu")
print(info["config"]["model_args"])

CLI Inference

python msfit/inference.py \
  --checkpoint AntonyG/msfit-dinobloom-v4-mlp-s3 \
  --data-root /path/to/wbcbench-2026-data \
  --predict-split test \
  --tta --tta-views 8 \
  --output-dir ./preds_exported_model

Notes

• dinobloom_base is instantiated through timm; on a fresh machine you may still need network access or cached Hugging Face artifacts to resolve the base DINOBloom architecture.

If this model helps your research, please cite;

@inproceedings{Gitau2026ISBI, author = {Antony Gitau and Martin Paulson and Bjørn-Jostein Singstad and Karl Thomas Hjelmervik and Ola Marius Lysaker and Veralia Gabriela Sanchez}, title = {Multi-Stage Fine-Tuning of Pathology Foundation Models with Head-Diverse Ensembling for White Blood Cell Classification}, booktitle = {2026 IEEE 23rd International Symposium on Biomedical Imaging (ISBI)}, year = {2026}, organization = {IEEE} }