---
library_name: open_clip
tags:
  - breast-ultrasound
  - biomedclip
  - image-text
  - segmentation
  - bus-cot
---

# BUS-COT upstream encoder checkpoints

Five upstream vision-language encoders trained or evaluated under the fixed
BUS-COT segmentation protocol (seed 12). These are encoder checkpoints only;
this repository contains no ultrasound images, masks, captions, or LLM/VQA
artifacts.

| Directory | Checkpoint | Method |
| --- | --- | --- |
| `baseline/` | `open_clip_pytorch_model.bin` | Untouched BiomedCLIP base checkpoint |
| `clip/` | `epoch_3.pt` | CLIP contrastive post-training |
| `cliprefine/` | `epoch_3.pt` | CLIPRefine post-training |
| `global_ind_gmpo/` | `epoch_3.pt` | Global-IND 4-way GMPO |
| `sample_ind_gmpo/` | `epoch_3.pt` | Sample-IND 4-way GMPO |

All models use the embedded OpenCLIP backend and
`hf-hub:microsoft/BiomedCLIP-PubMedBERT_256-vit_base_patch16_224` backbone.
Training used 4,197 paired examples, resolution 224, batch size 32, three
epochs, and seed 12. GMPO runs use `beta_dpo=10.0`; their automatic IND
diagnostics are in `diagnostics/` and run-specific settings are in each
directory's `protocol.json`.

The checkpoints are intended for the fixed BUS-COT upstream saliency →
k-means/filter → box-prompted SAM ViT-H evaluation pipeline. They do not
include a SAM checkpoint.