VPTracker (Qwen3.5-2B)
This repository contains the VPTracker full-parameter fine-tune from experiment 069. It uses Qwen3.5-2B as the base model and location-aware visual prompts for global vision-language tracking.
The checkpoint was trained with seed 42 and the ms-swift 4.4.0 development environment. The current code records upstream ms-swift v4.5.3 for inference and future training; verify the loading smoke test after upgrading runtime dependencies.
Loading
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration
model_id = "jcwang0602/VPTracker"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3_5ForConditionalGeneration.from_pretrained(model_id, device_map="auto")
The model card describes the 069 checkpoint only. Dataset preparation, visual-prompt rendering, evaluation, and reward plugins are available in the VPTracker code repository.