xprilion/Qwen3.5-0.8b-browser-agent-lora

Qwen-3.5-0.8B Browser Agent (LoRA)

This is a fine-tuned version of Qwen/Qwen3.5-0.8B designed to act as a lightweight Browser Navigation Agent.

It has been trained using LoRA to process a text-based "Accessibility Tree" of a webpage and output structured commands like CLICK(id), TYPE(id, "text"), and GOTO("url").

🚀 Capabilities

• Lightweight: Runs comfortably on GPUs with 8GB VRAM (uses ~950MB in 4-bit).
• Structured Output: Emits clear thoughts and actions, avoiding the typical repetition of small base models.
• Agentic reasoning: Can handle multi-step goals like "Navigate to the setup page and find the GitHub repo."

🛠️ Usage

This repository contains the LoRA adapter. You should load it using PEFT on top of the base Qwen model.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

model_id = "Qwen/Qwen3.5-0.8B"
adapter_id = "xprilion/Qwen3.5-0.8b-browser-agent-lora"

# 1. Load with 4-bit quantization
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
)

tokenizer = AutoTokenizer.from_pretrained(model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=quantization_config,
    device_map="auto",
)

# 2. Load the LoRA adapter
model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

# 3. Input Format
# Goal: [YOUR GOAL]
# Accessibility Tree:
# [1] link "Home"
# [2] button "Search"
# ...
# Thought:

📊 Performance

The model was fine-tuned on custom "Browser Traces" to ensure it stops after a single action and correctly maps dynamic instructions to DOM elements.