Browser Agent Tasks
WootzappLab/browser-agent-tasks on Hugging Face
WootzApp Research
100 multi-step public-web tasks, each with its paired recordings: the screenshot run and the W8 run of the same task, with final answers and agent logs. Each task states the instruction, a public starting page, when to stop, and what the agent must leave untouched.
This is the data behind W8 DOM-diff Bench.
- Tasks
- 100
- Distinct sites
- 90
- Screenshots (screenshot runs)
- 465
- DOM-model states (W8 runs)
- 465
Layout
Task definitions, plus two recordings of every run.
tasks.jsonl holds one task per line. Under trajectories/, each task has one folder per recording.
| Path | Contents |
|---|---|
| tasks.jsonl | 100 task definitions |
| trajectories/data-ss/task_NN | screenshotN.png, final_answer.json, task_data.json, web_surfer.log |
| trajectories/data-dom/task_NN | dom_modelN.txt, final_answer.json, task_data.json, web_surfer.log |
Task schema
Six fields per task.
| Field | Type | Description |
|---|---|---|
| task_id | Stable identifier | |
| category | Task family | |
| instruction | Complete instruction given to the browser agent | |
| start_url | Suggested public starting page | |
| stopping_condition | When the agent must stop | |
| constraints | Safety and scope restrictions |
Example · browser_task_019
One task, both recordings.
This is the “exact release date” example in Table 4 of the W8 DOM-diff Bench report. The screenshot run shows a relative date; the W8 run keeps the exact timestamp.
{
"task_id": "browser_task_019",
"category": "github release research",
"instruction": "Find the latest release of pytorch/pytorch. Report the release tag, release date, and the first three highlight bullets from the release notes in their displayed order.",
"start_url": "https://github.com/pytorch/pytorch/releases",
"stopping_condition": "Stop after the latest release metadata and first three highlights have been recorded.",
"constraints": [
"Do not sign in.",
"Do not create issues or modify repository content."
]
}
Load the tasks
from datasets import load_dataset
tasks = load_dataset(
"WootzappLab/browser-agent-tasks",
split="train",
)Scope
Every task runs on the public web. Agents do not sign in, purchase, submit irreversible forms, or start navigation.