SAM3 medical pure-2D — 259800-step training checkpoint
Private MedicalSAM3 research checkpoint at 259,800 optimizer updates. The model processes single 2D images or slices at 1008 × 1008 input resolution. This is an intermediate checkpoint, not a completed-training release. The training run used the project’s pure-2D SAM3 image model without LoRA or tracker/memory training.
Contents
| File | Description |
|---|---|
checkpoint.pt |
Full training checkpoint with model and continuation state |
checkpoint_receipt.json |
Recorded optimizer step, image cursor and checkpoint size |
status.json |
Recorded allocation-end state |
training.yaml |
Training configuration; absolute Anvil paths are environment-specific |
checkpoint_manifest.json |
Source identity and integrity metadata |
SHA256SUMS |
Source checkpoint SHA-256 |
LICENSE |
Upstream SAM License |
checkpoint.pt is 10,079,253,062 bytes. Its SHA-256 is:
c4223c36d46cf0c8214d0c4d1af6a1e520babd64bac4166078bff0bc9edbe0ec
The source receipt records optimizer_updates=259800, global_image_cursor=937200, epoch=0, and four saved rank RNG states. The run status is checkpointed_for_continuation as of 2026-10-18. This is the full checkpoint for resuming this run; the separate 184k repository contains an inference-only weight export and matching inference source snapshot.
Use
Authenticate with an account authorized for this private repository, then download checkpoint.pt. Restore it with the matching MedicalSAM3 project code and training configuration. This is a PyTorch training checkpoint and should only be loaded from a trusted source. For inference-only use, extract the model state with the matching project code rather than assuming this file has the same format as the 184k model-only export.
No training images or annotations are included. The earlier full-data geometry scan was stopped before covering 176,309 training images; this is a documented data-validation limit, not a property of the checkpoint file.