PathXAI refined D1 inference adaptation
This model repository contains one compact refined D1 adaptation checkpoint. It does not contain SAM3 base weights. manifest.json records the adapter file's size and SHA-256, source checkpoint identity, and the pinned external SAM3 dependency.
Files and inference dependencies
| File | Role | Size | SHA-256 |
|---|---|---|---|
weights/pathxai_d1_adapter.pth |
D1 fine-tuned LoRA, prompt embeddings, organ bias, prompt mask, adapter, and retained auxiliary trained states | 25,259,126 bytes | 35316dc06ca3ecf994e281901297e58c8a8147255ff0dcf6f9adf74ef14c337b |
Inference also requires the separate pinned SAM3 base model facebook/sam3, revision 3c879f39826c281e95690f02c7821c4de09afae7, file sam3.pt (3,450,062,241 bytes; SHA-256 9999e2341ceef5e136daa386eecb55cb414446a00ac2b55eb2dfd2f7c3cf8c9e). It is not hosted here. Users must request and receive access from the official SAM3 model page, then download it directly from Meta's repository.
The adapter was derived from D1 checkpoint best_model.pth (1,484,446,605 bytes; SHA-256 4d8ac978814f66ba5a716e26c8c44ada50f3f6c3f1bc987a97b24238babe9e5a) from refined training run medsam3_d1_refined/d1_final/run_20260804_140335_stage3_tissue. Before export, all 295 prompt_enc.base_text_encoder.* tensors were compared against the mapped tensors in the pinned SAM3 checkpoint and were exactly equal. The derived file omits those unchanged tensors, retains the D1 LoRA, prompt, adapter, and auxiliary trained states, and excludes optimizer/scheduler/scaler/resume-only state. The original checkpoint was not modified.
Use the PathXAI tutorial on GitHub for the local Jupyter workflow, model configuration, SO1/SO4 patient-level inference, evaluation, and recorded HTML examples. Follow its installation guide to fetch the pinned upstream MedSAM3 source. CONCH weights and patient/case data are not included in this model repository.
Access
Access to this repository does not grant access to the separate upstream SAM3 repository. This repository uses manual gated access; users request approval with their own Hugging Face account and download with an approved account token.
This research software is not a medical device and is not intended to diagnose, treat, cure, or prevent disease. Users are responsible for validating it for their own data and setting.
License and notices
HF metadata is license: other with the SAM License linked as LICENSE. This is the SAM License text published in the official SAM3 source and dated November 19, 2025. It covers SAM Materials including trained weights and derivatives and requires a copy to accompany distributed SAM materials. The licensing status of the D1 adaptation and applicability of this agreement to the pinned upstream revision must be confirmed by the publisher before distribution. This card does not claim blanket rights clearance.
Integrity
Verify the adapter against manifest.json before loading. The manifest also identifies the separate SAM3 dependency and upstream checksum. Do not substitute a newer SAM3 base without checking compatibility.