---
license: cc-by-4.0
language:
- en
pretty_name: CUA-Bench
size_categories:
- n<1K
task_categories:
- reinforcement-learning
tags:
- browser-agents
- computer-use-agents
- evaluation
- online-mind2web
configs:
- config_name: default
  default: true
  data_files:
  - split: navigation
    path: data/navigation.jsonl
  - split: single_search
    path: data/single_search.jsonl
  - split: multi_step_search
    path: data/multi_step_search.jsonl
---

# CUA-Bench

CUA-Bench contains 90 browser-agent evaluation tasks selected from
[Online-Mind2Web](https://huggingface.co/datasets/osunlp/Online-Mind2Web).
The original task IDs, instructions, and initial website URLs are preserved.

## Splits

| Split | Task type | Rows |
|---|---|---:|
| `navigation` | Navigate to a requested page or website state | 30 |
| `single_search` | Complete one direct search or lookup | 30 |
| `multi_step_search` | Complete a search requiring multiple steps or constraints | 30 |
| **Total** | | **90** |


## Row schema

| Field | Meaning |
|---|---|
| `task_id` | Original stable Online-Mind2Web task identifier |
| `category` | Behavioral task category and corresponding split |
| `instruction` | Instruction presented to the browser agent |
| `initial_url` | Live website URL where the task begins |
| `stopping_condition` | Condition describing when execution should stop |
| `constraints` | Safety and scope restrictions for the task |

Every row contains exactly these six fields. Categories, stopping conditions,
and constraints are evaluation metadata added for this task catalog.

## Load the dataset

```python
from datasets import load_dataset

dataset = load_dataset("WootzappLab/cua-bench")
navigation_tasks = dataset["navigation"]
```

To load one category directly:

```python
tasks = load_dataset("WootzappLab/cua-bench", split="single_search")
```

## Execution notes

The tasks use live website URLs from Online-Mind2Web. Website content and UI
behavior can change over time, so successful execution may depend on current
site availability, authentication, geography, and anti-automation controls.
The three current dataset splits provide task definitions only; the evaluation
harness and verifier are maintained separately. Any retained fixture or
verifier files in older repository revisions are not inputs for these splits.

## Attribution

The task IDs, instructions, and URLs originate from
[Online-Mind2Web](https://huggingface.co/datasets/osunlp/Online-Mind2Web).
Refer to the source dataset for its collection methodology and licensing terms.