| license: mit | |
| task_categories: | |
| - other | |
| tags: | |
| - web-browsing | |
| - agent-benchmark | |
| - ai-safety | |
| - reinforcement-learning | |
| # WebGames Dataset | |
| This dataset contains a collection of challenges designed for evaluating general-purpose web-browsing AI agents. The challenges are designed to be easy for humans but difficult for AI agents, and they are fast and simple to run using client-side state and a single-page JavaScript application. Each task provides a unique password upon successful completion, facilitating easy evaluation. | |
| [Project Page](https://webgames.convergence.ai) | [Paper](https://hf.co/papers/2502.18356) | |
| ## Try it now | |
| 🎮 [webgames.convergence.ai](https://webgames.convergence.ai) | |
| ## Run locally | |
| ```bash | |
| cd webgames | |
| pnpm install | |
| pnpm run dev | |
| ``` | |
| ## Download tasks | |
| Tasks are available as a dataset on Hugging Face. | |
| Alternatively, you can download them from the webgames website: | |
| 1. Go to [webgames.convergence.ai?showDownloads=true](https://webgames.convergence.ai?showDownloads=true) | |
| 2. Click the download buttons in the top-right corner (csv or jsonl available) | |
| 3. Verify your agent solutions using `solution in messages[-1]` or equivalent, or use the Inspect AI eval scaffolding in the eval folder. |