ihaveadog/qwen25-vl-7b-browser-agent-lora

Qwen2.5-VL-7B - Browser Agent (SFT Baseline)

This is a LoRA adapter for Qwen2.5-VL-7B-Instruct trained to act as a visual web browser agent. It was trained using Imitation Learning (Behavioral Cloning) on a dataset of expert human trajectories capturing screenshots, DOM elements, and browser actions.

🔬 Methodology

• Training Strategy: Supervised Fine-Tuning (SFT) over 15+ epochs.
• Dataset: Constructed via DAgger (Dataset Aggregation). We used live thought injection to guarantee reasoning-rich traces.
• Goal: Serve as the stable baseline for subsequent Reinforcement Learning (GRPO) experiments.

🚀 Performance

This model successfully navigates standardized benchmark tasks and outputs actions in the strict JSON format required by the browser-use library. However, it may struggle with negative constraints compared to the RL-refined version.

💻 Usage

from peft import PeftModel
from transformers import Qwen2_5_VLForConditionalGeneration

model = Qwen2_5_VLForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct", device_map="auto")
model = PeftModel.from_pretrained(model, "ihaveadog/qwen25-vl-7b-browser-agent-lora")