OA-CXR — default paper model
Code, installation, training and evaluation: hekaiwang/OA-CXR.
This is the original development-selected OA-CXR model: seed 17, image epoch 12, direct linear L1 readout epoch 27. These are not newly selected release-test weights.
Contents
main/image_state.pt: 29,573,522 bytes; pure tensor image-model state.main/readout_state.pt: 2,358,434 bytes; selected readout with fit-only normalization.main/config.json: architecture, tensor SHA256, implementation bindings and provenance.main/training_history.json: original aggregate epoch histories and development metrics.reproduction/kermany-v7-masks.tar.gz: optional CC BY 4.0 derived annotation PNGs required to reconstruct the exact Kermany training/evaluation labels. No radiographs.
Install the linked GitHub project and run python scripts/public/download.py --output weights.
Then python scripts/public/predict.py --weights weights/main --image /path/image.png --output runs/example.
The repository pins an immutable revision and checks hashes; no Hugging Face login is
required to use this public repository.
Training and intended interpretation
The image model was initialized from XRV's CheXpert-only DenseNet121 and trained on 19,086 COVIDQU/Kermany source images for 13 epochs (975 updates). Three readout candidates were independently trained for 30 epochs (10,920 updates each). Selection used development data, not release test results. The encoder is frozen in readout training. Full reconstruction/training commands and frozen split manifests are in GitHub. Training archives and source images are obtained from their providers.
Scores estimate relative retention of basal, peripheral and whole-lung annotation regions for effusion, pneumothorax and consolidation queries. They do not diagnose those conditions. An uncropped source is defined to have retention one even if its clinical acquisition was incomplete. The model is not clinically validated.
All original experiments used seed 17; reported bootstrap intervals do not measure training-seed variation. Some sources were historically observed; patient independence and foundation-pretraining non-overlap are not fully established. NIH evaluation uses pseudo masks. CUDA BF16 and CPU FP32 may produce small numerical differences.
Attribution and terms
See LICENSE and THIRD_PARTY_NOTICES.md. Author-owned contributions and third-party pretrained material have distinct scope. XRV's main library declares Apache-2.0; a separate checkpoint-specific license was not supplied upstream. No blanket MIT claim for all upstream weights/data is made. V7/CloudFactory annotation derivatives retain CC BY 4.0 and are not relicensed as MIT. MAIRA-2 and other restricted third-party weights are not redistributed.