---
library_name: transformers
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3-VL-8B-Instruct
tags:
- medical-imaging
- endoscopy
- vision-language-model
- qwen3-vl
- siglip2
- research
---

# GI Endoscopy Grounded VLM Checkpoints

Research checkpoints for the GI Endoscopy Grounded VLM project.

## Recommended checkpoint

`qwen3-vl-8b-lora/sqrt-balanced-seed42/checkpoint-400`

Base model: `Qwen/Qwen3-VL-8B-Instruct`

HyperKvasir validation results:

- Accuracy: 0.4620
- Macro F1: 0.2441
- Balanced accuracy: 0.2684

## SO400M bridge ablations

`ablations/so400m-direct-bridge/checkpoint-500`

- Full-validation accuracy: 0.2655
- Macro F1: 0.1560
- Balanced accuracy: 0.1750

`ablations/so400m-fixed64-distillation/final`

- Validation-100 accuracy: 0.0500
- Macro F1: 0.0041
- Outcome: single-class collapse
- This checkpoint must not be used for deployment.

## External checkpoints

SO400M classifier:

`Sahibnoor1/gi-siglip2-dino-hyperkvasir-checkpoints`

Kvasir-SEG segmentation:

`Sahibnoor1/kvasir-siglip2-segmentation-checkpoints`

## Warning

Research demonstration only. These models are not validated for diagnosis,
treatment, or patient management. Do not upload identifiable patient information.