---
license: apache-2.0
tags:
- medical
- segmentation
- vision-language
- sa2va
- vessel-segmentation
- dpo
---

# Sa2VA 26B DPO v8 (Step 100)

Checkpoint at step 100 of Sa2VA 26B model trained with Direct Preference Optimization (DPO) strategy.

## Model Details

- **Model Type**: Sa2VA (Segment Anything 2 Vision Assistant)
- **Base Model**: InternLM2-26B
- **Training Strategy**: Direct Preference Optimization (DPO)
- **Checkpoint**: Step 100
- **Resolution**: 512x512
- **Task**: Medical vessel segmentation (coronary arteries)

## Usage

```python
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("ly17/sa2va-26b-dpo-v8-step100", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("ly17/sa2va-26b-dpo-v8-step100", trust_remote_code=True)
```

## Training Details

This checkpoint uses DPO to align the model with preferred segmentation outputs, improving quality through preference learning.