github.devotd.
NewEvals

servicenow/browsergym v0.14.3 — a Gym environment for web task automation

servicenow/browsergym is a Python framework that bundles web agent benchmarks such as MiniWoB, WebArena, VisualWebArena, and WorkArena under a common Gym API to run and evaluate them, and the latest publicly released version is v0.14.3.

Repositoryservicenow/browsergym ↗

servicenow/browsergym is a Python framework that bundles web agent benchmarks such as MiniWoB, WebArena, VisualWebArena, and WorkArena under a common Gym API to run and evaluate them, and the latest publicly released version is v0.14.3.

GitHub description: 🌎💪 BrowserGym, a Gym environment for web task automation

stars
★ 1.4K
Forks
207
Language
Python
License
NOASSERTION
Latest release
v0.14.3
Read original source ↗View on GitHub ↗

servicenow/browsergym is a Gym environment that automates tasks performed in a web browser. A Gym is an interaction specification in which an agent receives observations from the environment and sends actions. The repository is written in Python, and the latest publicly released version is v0.14.3. This record contains no change list for v0.14.3, so this post covers the version and the publicly available scope.

The README describes this repository as an open-source framework intended to promote web agent research. It provides several benchmarks, such as MiniWoB, WebArena, VisualWebArena, and WorkArena, as its foundation, and its core is running and evaluating them under a common Gym API by grouping the real tasks they cover. It is not a consumer product but a tool for research and evaluation, and the repository, created in February 2024, has a public star count of 1408.

There are two limitations. First, the license is recorded only as NOASSERTION, so the usage and distribution conditions have to be confirmed by checking the original repository text directly. Second, this repository is not a place for creating new benchmarks but rather the side that gathers already-published benchmarks into a single environment framework and runs them. Actual agent scores or performance comparison figures are not included in this repository's record. It can be used for work that runs multiple benchmarks under the same specification, but judging which performs better is premature based on this record alone.