---
license: apache-2.0
language:
- en
task_categories:
- text-generation
- question-answering
tags:
- agents
- browser-use
- betterwright
- playwright
- tool-use
- synthetic-data
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.jsonl.gz
  - split: validation
    path: data/validation-*.jsonl.gz
  - split: test
    path: data/test-*.jsonl.gz
---

# BetterWright Agentic Browser — 6,093-row stopped checkpoint

This is the public checkpoint of a generation run originally planned for 50,000 rows. Generation was stopped at the account owner's request and the exact 6,093 accepted rows were packaged. It is synthetic training data, not live browser recordings.

## Contents

- 5,971 BetterWright demonstrations and 122 Playwright demonstrations.
- 32 task domains and 21 browser feature categories.
- Harness-shaped conversations for Codex CLI, OpenCode, Pi, and Hermes.
- Compact `OBS; NEXT; CHECK; FALLBACK` decision rationale.
- Train/validation/test split: 5,851/121/121.
- Generator and critic: `qwen3.8-27b`. Retained rows scored at least 8/10 per automated category and averaged at least 8.6/10.

Each row includes `id`, task metadata, `user_request`, `start_url`, a uniform `messages` list, automated `quality_scores`, `quality_issues`, and generation metadata. Tool arguments are JSON strings.

## Important limitations

The trajectories and web states are simulated on reserved `.example` hosts; they were not replayed against live sites. The generator and critic used the same model, so the scores are not an independent evaluation. A checkpoint audit found that many concise simulated tool results do not expose full BetterWright UI/control directories, and some later actions use CSS/test-id selectors not explicitly present in prior observations. This can teach selector hallucination. Treat the corpus as experimental synthetic SFT data, mix it with execution-verified trajectories, and evaluate downstream agents in a sandbox before consequential browser use.

## Validation

The packaged checkpoint has 6,093 unique IDs and 6,093 unique normalized user requests. Message/tool-call pairing, JSON tool results, reasoning format, score thresholds, and split row counts were checked before upload. See `validation_report.json` and `generation_config.json`.

## Source grounding

Tool and harness shapes were developed against pinned public source revisions:

```json
{
  "betterwright": "5ba586ff1b26353774d81d6b8f8241d0969251c5",
  "codex": "eb10d91e48ccbd0930427461fb392337addb1ac0",
  "hermes-agent": "df4b3733baa5534bec1e6e888e5ba86a0a6f9c3b",
  "opencode": "69c172e8a7c0086887b1f93ed5a162f14b6aa0c5",
  "pi-mono": "96317e50b8d6e7f6d0e47fd29122baf1461c00f5"
}
```

BetterWright is an MIT-licensed project from The BetterWright Project and contributors. This independently generated dataset is not an official BetterWright release: https://github.com/BetterWright/betterwright