Awesome

WebLINX

🤗Dataset	📄Paper	🌐Website	📓Colab
🤖Models	💻Explorer	🐦Tweets	🏆Leaderboard

<br>

WebLINX: Real-World Website Navigation with Multi-Turn Dialogue
Xing Han Lù*, Zdeněk Kasner*, Siva Reddy
*Equal contribution
ICML 2024 (Spotlight)

[!IMPORTANT] WebLINX is now available as a benchmark through BrowserGym, allowing you to access demonstration steps in the same way you would access a web agent environment like WebArena or MiniWoB. This also allows you to run agents from the Agentlab library, including agents that achieve SOTA performance through Claude-3.5-Sonnet. To enable this integration, we are releasing the weblinx-browsergym extension for BrowserGym on PyPi, as well as a new dataset, WebLINX 1.1, derived from WebLINX on Huggingface. In WebLINX 1.1, a small number of demonstrations were removed after processing, but no new demonstration was added. There are substantial changes to the steps being evaluated, with the inclusion of tab actions. Please report your results as "WebLINX-1.1", "WebLINX-BrowserGym" or "WebLINX-BG" in your work, to differentiate from the initial release of weblinx (1.0).

Intro

Welcome to WebLINX's official repository! In addition to providing code used to train the models reported in our WebLINX paper, we also provide a comprehensive Python library (aka API) to help you work with the WebLINX dataset.

If you want to get started with weblinx, please check out the following places:


🌐	Website	If you want a quick overview of the project, this is the best place to start.
📓	Colab	Eager to try it out? Start by running this colab notebook!
🗄️	Docs	You can find quickstart instructions, the official user guide, and all relevant API specifications in the docs.
📄	Paper	If you want to get more in-depth, please read our paper, which provides comprehensive description of the project and report relevant results.
🤗	Dataset	The official dataset page, you can download preprocessed dataset and follow instructions to get started.

If you want to learn more about the codebase itself, please keep on reading!

Installation

# Install the base package
pip install weblinx

# Install all dependencies
pip install weblinx[all]

# Install specific dependencies for...
# ...processing HTML 🖥️
pip install weblinx[processing]
# ...video processing 📽️
pip install weblinx[video]
# ...evaluating models 🔬
pip install weblinx[eval]
# ...development of this library 🛠️
pip install weblinx[dev]

Structure

This repository is structured in the following way:

Module	Description
`weblinx`	The `__init__.py` provides many useful abstractions to provide a Pythonic experience when working with the dataset. For example, you can use `weblinx.Demonstration` to manipulate a demonstration at a high-level, `weblinx.Replay` to focus on more finegrained details of the demonstration, including iterating over turns, or `weblinx.Turn` to focus on a specific turn. All relevant information is included in the documentations!
`weblinx.eval`	Code for evaluating action models trained with WebLINX, it has both `import`able functions/metrics, but can also be accessed via command line
`weblinx.processing`	Code for processing various inputs or outputs used by the models, it is extensively used in the models' processing code
`weblinx.utils`	Miscellaneous utility functions used across the codebase.

Modeling

Our modeling/ repo-level directory has code for processing, training and evaluating the models reported in the paper (DMR, LLaMA, MindAct, Pix2Act, Flan-T5). It is separate from the weblinx library, which focuses on data processing and evaluation. You can use it by cloning this repository, and it is recommended to edit the files in modeling/ directly for your own needs. Our modeling code is separate from the weblinx library, but requires it as a dependency. You can install the modeling code by running:

# First, install the base package
pip install weblinx

# Then, clone this repo
git clone https://github.com/McGill-NLP/weblinx
cd weblinx/modeling

For the rest of the instructions, please take a look at the modeling README.

Evaluation

To install packages necessary for evaluation, run:

pip install weblinx[eval]

You can now access the evaluation module by importing in Python:

import weblinx.eval

Use weblinx.eval.metrics for evaluation metrics, weblinx.eval.__init__ for useful evaluation-related functions. You may also find it useful to take a look at weblinx.processing.outputs to get an idea of how to use the outputs of the model for evaluation.

To run the automatic evaluation, you can use the following command:

python -m weblinx.eval --help

For more examples on how to use weblinx.eval, take a look at the modeling README.

Note: We are still working on the code for weblinx.eval and weblinx.processing.outputs. If you have any questions or would like to contribute docs, please feel free to open an issue or a pull request.

Citations

If you use this library, please cite our work using the following:

@misc{lù2024weblinx,
      title={WebLINX: Real-World Website Navigation with Multi-Turn Dialogue}, 
      author={Xing Han Lù and Zdeněk Kasner and Siva Reddy},
      year={2024},
      eprint={2402.05930},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

License

This project's license can be found at LICENSE. Please note that the license of the data in tests/data follow the license from the official dataset, not the license of this repository. The official dataset's license can be found in the official dataset page. The license of the models trained using this repo might also differ - please find them in the respective model cards.