README.md

metadata

license: apache-2.0
datasets:
  - TIGER-Lab/BrowserAgent-Data
language:
  - en
base_model:
  - Qwen/Qwen2.5-7B-Instruct
metrics:
  - success_rate
  - trajectory_f1
tags:
  - agent
  - browser
  - web
  - sft

Model

We release the SFT (Supervised Fine-Tuned) model used in BrowserAgent, based on Qwen/Qwen2.5-7B-Instruct.
This model learns structured web-browsing behaviors—such as click, type, scroll, read, submit—from human-style demonstrations and produces schema-constrained action sequences for browser environments.

• Base: Qwen2.5-7B-Instruct
• Objective: Next-token prediction on normalized, schema-validated browsing trajectories
• Format: JSON-like structured actions (compatible with BrowserAgent runtime)

Data

The SFT data includes:

• Human and assisted browsing demonstrations
• Canonicalization under a unified action schema
• Filtering and de-duplication to ensure validity and safety

Code

https://github.com/TIGER-AI-Lab/BrowserAgent

Sample Usage

hf download TIGER-Lab/BrowserAgent-SFT --local-dir ./models/browseragent-sft --repo model

Citation

@misc{yu2025browseragentbuildingwebagents,
      title={BrowserAgent: Building Web Agents with Human-Inspired Web Browsing Actions}, 
      author={Tao Yu and Zhengbo Zhang and Zhiheng Lyu and Junhao Gong and Hongzhu Yi and Xinming Wang and Yuxuan Zhou and Jiabing Yang and Ping Nie and Yan Huang and Wenhu Chen},
      year={2025},
      eprint={2510.10666},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.10666}, 
}