---
pretty_name: ScaleWoB VERL
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.parquet
      - split: validation
        path: validation.parquet
  - config_name: easy
    data_files:
      - split: train
        path: train_easy.parquet
---

# ScaleWoB VERL

Private dataset package for browser-agent reinforcement learning with VERL, GiGPO, and the ScaleWoB runtime.

## Contents

| Config | Split | File | Rows | Environments |
|---|---|---:|---:|---:|
| `default` | train | `train.parquet` | 1,459 | 85 |
| `default` | validation | `validation.parquet` | 162 | 66 |
| `easy` | train | `train_easy.parquet` | 1,621 | 85 |

The default train and validation splits have no overlapping `(env_id, task_id)` pairs. `train_easy.parquet` is an alternate all-task training file; unlike the default split, its records do not contain `split` or `has_params` metadata.

## Schema

Each row contains:

- `data_source`: always `scalewob`.
- `prompt`: chat-format user prompt consumed by VERL.
- `extra_info.index`: source row index.
- `extra_info.split`: `train` or `val` when present.
- `extra_info.has_params`: task stratum when present.
- `extra_info.scalewob.env_id`: environment identifier.
- `extra_info.scalewob.task_id`: task identifier within the environment.
- `extra_info.scalewob.description`: natural-language task instruction.

See `metadata.json` for machine-readable row counts, checksums, and compatibility information.

## Runtime dependency

These Parquet files contain task metadata, not the browser assets. Use the matching private environment repository and pin both repositories to immutable tags or commits.

Expected runtime versions for this package:

- `scalewob==0.12.1`
- `playwright==1.62.0`
- Chromium `151.0.7922.34`
- VERL `0.7.1`

## Loading

```python
from datasets import load_dataset

dataset = load_dataset("<namespace>/scalewob-verl", name="default", token=True)
train = dataset["train"]
validation = dataset["validation"]
```

For training containers, downloading once on the host and mounting the resulting directory read-only is preferred over downloading independently in every worker.

## Intended use and limitations

Intended for private research on browser-agent reinforcement learning. Tasks operate on local simulated sites rather than live services. Task text may contain synthetic names, account identifiers, addresses, payment-like values, or other placeholders; do not interpret them as instructions to interact with real services.

No public redistribution license is asserted by this package. Keep the repository private unless provenance, privacy, trademark, and redistribution rights have been reviewed.