Existing benchmarks for web agent tasks are either offline and static, or operate within a fully reproducible environment with limited Internet dynamics. The WebCanvas project aims to pioneer the online evaluation of web agents. Additionally, we offer a suite of toolkits for scaling and maintaining web agent data to support this endeavor. We welcome any constructive feedback on the project and look forward to partnering with you in developing agents for web tasks!
- [2024, September 9] Support evaluation for OpenAI new o1 models, includes o1-preview and o1-mini. Just set the 'planning_text_model' parameter to 'o1-preview' or 'o1-mini'.
- [2024, August 9] We're excited to announce the release of v0.0.3 of WebCanvas! This update introduces support for evaluation of data operations, such as caching data in process and outputting the final answer. You can now define and evaluate a broader range of web tasks using iMean Builder and WebCanvas. Additionally, we've introduced a new metric: US dollar consumption / key node completion(usd_efficiency_score), detailed in this section. We believe that an agent's efficiency is crucial for online web tasks, and this metric will help quantify that efficiency.
- [2024, July 13] We've released v0.0.2 of WebCanvas. This update brings the ability to call different base model services, including OpenAI, Claude, Gemini, and together.ai. Now, you can choose any of these model services for testing on our platform. Additionally, we've launched a new repository: WebCanvas Showcase, which demonstrates how different agent frameworks can be integrated with the WebCanvas framework for online evaluation. We're kicking things off with the integration of SEEACT1 and WebCanvas. Play with it and explore the possibilities!
- [2024, June 18] Our paper will be presented at agentic markets workshop in ICML 2024 and natural language reasoning and structured explanations workshop in ACL 2024. See you in Vienna and Bangkok!
- [2024, June 18] Our pre-print paper "WebCanvas: Benchmarking Web Agents in Online Environments" is available!
- [2024, June 6] We've released WebCanvas, including Data, Platform, Toolkits, and Web agents(in this repo)!
- Base Agent Framework: Includes a base agent framework with several key modules - Planning, Observation, Memory, Reward, Action Execution, Evaluation, designed to be plug-and-play, enabling developers to easily test and iterate on their own LLM-based web agents.
- Dynamic and Real-time Web Environment Interaction: Utilizes live web environments to provide a realistic assessment and feedback of web agents, thus addressing key challenges for web agents such as error handling, CAPTCHA solving, key-node based evaluation, metrics for agent efficiency etc..
- Key Nodes Annotation: Introduces the concept of "key nodes" to offer in-progress feedback and a granular, phase-based assessment system that rigorously evaluate web agents in the wild.
- Scale Web Agent Evaluation in Live Web Environments: Connected to a comprehensive suite of toolkits with accurate observation capture and rich action space to define demonstration trajectories and intermediate states for real-time, open-ended web tasks, allowing for robust evaluation in dynamic web environments. Check out our How to guide.
- Mind2Web-Live Dataset: Presents a refined version of the original Mind2Web2 static dataset, containing 542 tasks with 2439 intermediate evaluation states, serving as the foundation general purpose benchmark.
- Open Data Access: Raw data of all challenges can be downloaded, including raw html(ready), screenshot(ready), DOM tree(soon), Axtree(soon), captured action(ready), element position(ready), etc., refer to challenge propose and data download. We plan to open the data to the community free for research use.
- Better Modularity and More Flexible Integration: To help easier integration of WebCanvas evaluation, connect offline agents to online web environment.
- Better Observation: Faster in computing, more accurate, and combine more modality(text, code, vision, conversation, etc.)
- Broader Action Space: Add actions like cache in memory, output final answer, code execution etc. to develop a better interface for web agent, which may differ from human's.
- Dynamic Evaluation Function: Provide toolkit for community to define dynamic evaluation functions(for example, model-based evaluation) as supplementary of current static evaluation functions.
- More Dataset Coverage: Introduce more datasets in different domains that address key capabilities in online web tasks.
- Accumulate Knowledge on Agent Experiences: Develop better algorithm to handle error encountered when inference in live environment, also accumulate knowledge on agent experiences in different websites.
- Statistics on Agent Cost other than Performance: Enable calculation of token consumption or GPU consumption of agent framework or agent model to serve as another optimization goal for truly practical web agent.
- Cloud Version to Mitigate Environment Discrepancy: We are working on a cloud version for more reliable evaluation.
- Support more base model calling(Claude, Gemini, Open-source Models from together.ai, etc.). (Done)
- Add information extraction related actions and relative evaluation metrics. (Done)
- Enable token consumption calculation. (Done)
- Add more brilliant web agent benchmarking data as showcase: webarena3, GAIA4, workarena5, etc. (in progress)
- Better modularity to ease integration. (in progress)
- Add vision as an extra observation and implement various grounding strategies. (in progress)
- Keep updating error handling module.
- Develop up-to-date visualizations of current live websites agent performance.
- Enable script-based actions and evaluation.
- Enable multi-modality input other other language-based instruction like GAIA4.
- Add tool calling in the reasoning framework.
First, ensure your environment is ready by installing the necessary dependencies:
conda create -n webcanvas python=3.11
conda activate webcanvas
pip install -r requirements.txtBefore running the repos, you need to set up the required API keys as using features dependent on external APIs. Please refer to this docs.
From our experiments, the experimental environment plays a crucial role in agent performance. We recommend experimenting on a Windows server using Chrome or Firefox browser engines, preferably on servers located in the United States. Below is the experiment results on Mind2Web-Live test set.
| Planning Model | IP Region | System | Browser | Completion Rate | Task Success Rate | Efficiency Score |
|---|---|---|---|---|---|---|
| gpt-3.5-turbo-0125 | United States | Windows | Chrome | 40.2% | 16.5% | 3.03 |
| gpt-3.5-turbo-0125 | United States | Windows | Firefox | 42.1% | 20.2% | 2.79 |
| gpt-3.5-turbo-0125 | United States | Linux | Chrome | 36.5% | 15.4% | 3.33 |
| gpt-3.5-turbo-0125 | United Kingdom | Windows | Chrome | 23.6% | 8.65% | 7.78 |
| gpt-3.5-turbo-0125 | Singapore | Windows | Chrome | 42.3% | 21.2% | 2.95 |
Register on the platform here.
First, ensure your environment variables are correctly set so that the code can access the necessary credentials and URL.
export GRAPHQL_USERNAME=your_username
export GRAPHQL_PASSWORD=your_password
If you registered using your Google account, please setup a password in the profile page on iMean Builder.
To download a file, use the following command:
python data/dataset_io.py download \
--challenge-id your_challenge_id \
--save-path /path/to/save/fileyour_challenge_id: The ID of the challenge for the download. Obtain this ID on the url link of the challenge for now. For example, the ID of Mind2Web-Live Test is "WjVIjPfpa-psiltU3oD2W"./path/to/save/file: The path where the downloaded file will be saved.
The raw data contain rich information on step level to inspire future research. However, it's not for our evaluation.
To process the raw data, run the follow command:
python data/raw_data_processor.py \
--input-file path/to/input/file \
--output-file path/to/output/file
You can run the repos with the following command:
python evaluate.py \
--global_reward_mode dom_reward \
--index -1 \
--single_task_name "Find Dota 2 game and add all DLC to cart in steam." \
--planning_text_model gpt-4o-mini \
--global_reward_text_model gpt-4o-miniThis command runs the script with DOM-based self-reward, processing the default task "Find Dota 2 game and add all DLC to cart in steam" or using the default data index -1. It also uses the GPT-3.5 Turbo model for both observation and global reward processing. The evaluation mode is controlled by the task_mode parameter in configs/setting.toml, allowing you to choose between batch mode and single mode(without automatic evaluation). Remember to specify your path to the test file in configs/setting.toml.
This program supports several command-line arguments to customize its behavior:
-
--global_reward_mode: Selects the method for getting global rewards.- Options:
dom_vision_reward,dom_reward,vision_reward,no_global_reward - Default:
dom_reward - Description: Define how rewards are got based on the interaction mode:
dom_vision_reward: Rewards are calculated using both DOM and vision data. Currently only support GPT4v as vision model.dom_reward: Rewards are based solely on DOM interactions. You can specify the language model you want to use for reward reasoning by parameter global_reward_text_model.vision_reward: Rewards are derived from vision-based interactions only. Currently only support GPT4v as vision model.no_global_reward: No global rewards are calculated.
- Options:
-
--index: Decide which data index to start with.- Type: String
- Default:
-1 - Description: Use this parameter to specify a range or specific index for data processing. For example,
0,5will process data from index 0 to 5.
-
--single_task_name: Defines the task name of the single task to execute.- Type: String
- Default:
"Find Dota 2 game and add all DLC to cart in steam."
-
--planning_text_model: Specifies the model used for planning module.- Type: String
- Default:
gpt-4o-mini
-
--global_reward_text_model: Specifies the model used for global reward reasoning.- Type: String
- Default:
gpt-4o-mini
Evaluating web agents in an online environment can sometimes be painful due to issues like network problems or bot tests on certain websites. Adopting an evaluation method that accommodates these issues allows for an accurate assessment of an agent's performance under specific current conditions. Additionally, we provide a more flexible interaction mode, enabling users to manually solve environmental issues and get the optimized performance of their web agents. You can simply set the interaction_mode parameter in configs/setting.toml to enable this feature. We will accumulate our implementation on error handling in online agent inference, and try to minimize human efforts by triggering only when exceptions occur in the following version.
IMPORTANT: You should upload the generated out.json file to participate a challenge. To upload your result, use the following command:
python data/dataset_io.py upload \
--file-path /path/to/your/file \
--challenge-id your_challenge_id \
--name your_agent_name \
--base-model your_agent_base_modelReplace the placeholders with your actual values:
/path/to/your/file: The path to the result you want to upload.your_challenge_id: The ID of the challenge you want to participate.your_agent_name: The agent name for the upload.your_agent_base_model: The agent base model information for the upload.
You can also submit through our platform. We will conduct an official check on your submission to prevent cheating.
We provide a token consumption calculation functionality for evaluating the efficiency of your agent, and it is enabled automatically.
The token consumption is calculated based on the number of tokens consumed by planning module and global reward reasoning module(if applicable) during the evaluation process.
The token consumption calculation results of each experiment will be saved in the token_results folder in JSON format.
We use the tiktoken package to calculate the consumption of tokens. For those models whose encodings cannot be obtained, the default encoding "cl100k_base" is used. Therefore, for non-OPENAI models, the calculated tokens may have certain deviations.
The amount spent on tokens is only available when the model name is provided in the 'token_pricing' under setting.toml; otherwise, only the quantity of tokens will be counted. If you want to calculate the monetary expenditure of models not listed in 'token_pricing', you should first add the full name of the model (such as "gpt-4o-2024-05-13") to the 'pricing_models' list. Then, add the unit price of input and output for this model below the list, such as "gpt-4o-2024-05-13_input_price = 0.000005" and "gpt-4o-2024-05-13_output_price = 0.000015".
Few example results on Mind2Web-Live test set:
| Planning Model | Completion Score | Task Success Rate | USD Efficiency Score |
| GPT-4o-2024-05-13 | 51.4% | 28.8% | 0.142 |
| Llama-3.1-405B-Instruct-Turbo | 47.8% | 24.0% | 0.174 |
| Llama-3.1-70B-Instruct-Turbo | 44.8% | 20.2% | 0.031 |
| GPT-4o-mini-2024-07-18 | 42.9% | 21.2% | 0.004 |
| GPT-3.5-turbo-0125 | 42.5% | 17.3% | 0.092 |
You can follow instructions on this documentation about how to create your own challenging benchmark for web agents.
Currently, you need to set up a challenge(you can keep it private at first) on WebCanvas Platform to download raw data of your dataset.
If your are thinking of scaling your Web trajectory data for training and evaluation, you can contact us directly for some technical assistance.
Demo video:
We welcome contributions to WebCanvas!
Thank you for your interest in improving WebCanvas. Your contributions are greatly appreciated and essential to the growth and success of our project. Please refer to the roadmap and TODOs for promising directions.
We are building a vibrant and inclusive community around WebCanvas! Join our community to stay up-to-date with the latest developments and to contribute to the project:
We value your feedback and suggestions!
- Talk to Founder, we welcome any discussion and feedback on the future of live agent evaluation!
If you use this project in your research, please cite our paper:
@article{pan2024webcanvas,
title={WebCanvas: Benchmarking Web Agents in Online Environments},
author={Pan, Yichen and Kong, Dehan and Zhou, Sida and Cui, Cheng and Leng, Yifei and Jiang, Bing and Liu, Hangyu and Shang, Yanyi and Zhou, Shuyan and Wu, Tongshuang and others},
journal={arXiv preprint arXiv:2406.12373},
year={2024}
}
Footnotes
-
Zheng, Boyuan, et al. "Gpt-4v (ision) is a generalist web agent, if grounded." arXiv preprint arXiv:2401.01614 (2024). ↩
-
Deng, Xiang, et al. "Mind2web: Towards a generalist agent for the web." Advances in Neural Information Processing Systems 36 (2024). ↩
-
Zhou, Shuyan, et al. "Webarena: A realistic web environment for building autonomous agents." arXiv preprint arXiv:2307.13854 (2023). ↩
-
Mialon, Grégoire, et al. "Gaia: a benchmark for general ai assistants." arXiv preprint arXiv:2311.12983 (2023). ↩ ↩2
-
Drouin, Alexandre, et al. "WorkArena: How Capable are Web Agents at Solving Common Knowledge Work Tasks?." arXiv preprint arXiv:2403.07718 (2024). ↩