Ahmed46S4/webgym_tasks

---
license: cdla-permissive-2.0
task_categories:
- reinforcement-learning
language:
- en
tags:
- web-navigation
- web-agents
- task-planning
size_categories:
- 100K<n<1M
configs:
- config_name: default
data_files:
- split: train
path: train.jsonl
- split: test
path: test.jsonl
---

# WebGym Tasks Dataset

## Dataset Description

This dataset contains web navigation tasks for training and evaluating autonomous web agents. Each task consists of a natural language instruction that describes an action to be performed on a specific website, along with evaluation criteria and metadata.

### Dataset Summary

- **Total Training Tasks**: 292,092
- **Total Test Tasks**: 1,167
- **Domains**: Multiple domains including Lifestyle & Leisure, Sports & Fitness, and more
- **Source Benchmarks**: Includes tasks from mind2web-live and other web navigation benchmarks

### Data Format

The dataset is provided in JSONL (JSON Lines) format with two splits:
- `train.jsonl`: Training set
- `test.jsonl`: Test set

### Data Fields

Each task contains the following fields:

- `benchmark_name` (string): Source benchmark (e.g., "mind2web-live")
- `task_name` (string): Natural language description of the task
- `domain` (string): High-level domain category
- `subdomain` (string): Specific subdomain category
- `website` (string): Target website URL
- `definite_answer` (string): Expected answer if applicable
- `task_id` (string): Unique task identifier. Numeric strings for base tasks (e.g., "0", "1"), strings with suffix for decomposed tasks (e.g., "100002_d1")
- `difficulty` (integer): Task difficulty level
- `evaluator_reference` (list): Evaluation criteria with descriptions and facts
- `task_id_decomposed_from` (string/null): Parent task ID if this is a decomposed subtask. Loaded as string to match task_id format

### Example

```json
{
"benchmark_name": "mind2web-live",
"task_name": "Find the score of the 2020 Super Bowl in nfl.com",
"domain": "Lifestyle & Leisure",
"subdomain": "Sports & Fitness",
"website": "https://nfl.com",
"definite_answer": "",
"task_id": 0,
"difficulty": 2,
"evaluator_reference": [
{
"id": 1,
"description": "find score information for the 2020 Super Bowl",
"facts": [
"score of the 2020 Super Bowl",
"information found on nfl.com"
]
}
],
"task_id_decomposed_from": null
}
```

### Dataset Notes

**Synthetic Data Disclosure**: This dataset contains synthetically generated tasks and may include synthetic components in task descriptions, evaluation criteria, and other fields.

**Task ID Fields**: Both `task_id` and `task_id_decomposed_from` are stored as strings for consistency. The `task_id` field contains numeric strings for base tasks (e.g., "0", "1") and strings with suffixes for decomposed tasks (e.g., "100002_d1"). The dataset includes 33,497 decomposed tasks representing about 11.5% of the training data.

### Usage

Load the dataset using the Hugging Face datasets library:

```python
from datasets import load_dataset

# Load from HuggingFace Hub
dataset = load_dataset("your-username/webgym-tasks")

# Or load from local files
dataset = load_dataset("json", data_files={
"train": "train.jsonl",
"test": "test.jsonl"
})

# Access examples
for task in dataset["train"]:
print(f"Task {task['task_id']}: {task['task_name']}")
```

### Citation

If you use this dataset, please cite:

```bibtex
@article{bai2026webgym,
title={WebGym: Scaling Training Environments for Visual Web Agents with Realistic Tasks},
author={Bai, Hao and Taymanov, Alexey and Zhang, Tong and Kumar, Aviral and Whitehead, Spencer},
journal={arXiv preprint arXiv:2601.02439},
year={2026}
}
```

### License

This dataset is released under the [Community Data License Agreement - Permissive - Version 2.0 (CDLA-Permissive-2.0)](https://cdla.dev/permissive-2-0/).