metadata
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
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.