Qwen3-0.6B-BrowserAgent

A compact browser agent. Given a task goal and a BrowserGym page observation, it outputs only the next action, for example click('37').

It is Qwen/Qwen3-0.6B fine-tuned with LoRA (merged into the weights) on saital/browser-agent-phase1-sft-action-only, a set of successful MiniWoB steps generated by a larger teacher model.

Your browser does not support the video tag.

Results

Our fine-tuned Qwen3-0.6B Browser Agent model demonstrates a substantial performance improvement over the base Qwen3-0.6B model across all evaluation metrics.

Most notably, exact match accuracy surged from 0.83% to 89.58%, representing an 88.75 percentage point increase (approximately 108× improvement).

The model achieved perfect 100% action type matching (up from 84.17%) and maintained 100% valid format compliance (up from 98.33%).

Additionally, target element matching improved dramatically from 4.17% to 92.5% (+88.33 pp), while fill text matching increased from 20.0% to 97.5% (+77.5 pp).

These results confirm that our fine-tuning approach effectively enhanced the model's capability to perform precise browser automation tasks with high reliability and accuracy.

Usage

Use non-thinking mode (enable_thinking=False) with the BrowserGym system prompt from the dataset.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

repo = "Rumiii/Qwen3-0.6B-BrowserAgent"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype=torch.float16, device_map="auto")

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": USER_PROMPT},  # task, goal, URL, history, observation
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
inputs = tok(text, return_tensors="pt", add_special_tokens=False).to(model.device)
out = model.generate(**inputs, max_new_tokens=64, do_sample=False)
print(tok.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Training

• LoRA SFT (r=64, alpha=128) on all attention and MLP projections, then merged
• Loss on action tokens only
• 3 epochs, learning rate 1e-4, batch size 16, single T4 GPU
• Best checkpoint by validation loss

Limitations

• Trained on about 30 synthetic MiniWoB task types; little transfer to real-world websites is expected.
• Trained on successful trajectories only, so it has not learned to recover from mistakes.
• Weakest on tasks that require choosing among many similar elements or copying text exactly.