sanaX3065/aegis-qwen2vl-3b

Aegis - Browser QA Vision Agent

Fine-tuned from Qwen2.5-VL-3B-Instruct using 10-phase progressive LoRA training on a 20k browser screenshot + DOM dataset.

Model details

Field	Value
Base model	Qwen2.5-VL-3B-Instruct
Training phases	10 (progressive curriculum)
Final loss	0.0536
LoRA rank	r=16, alpha=16
Max seq length	8192
Merged	Yes - no adapter needed

Quick start

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
import torch

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "sanaX3065/aegis-qwen2vl-3b",
    torch_dtype=torch.float16,
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("sanaX3065/aegis-qwen2vl-3b")