---
license: mit
task_categories:
- text-generation
- conversational
- table-question-answering
language:
- en
- ja
tags:
- synthetic
- sft
- react
- agent
- alfworld
- dbbench
size_categories:
- 1K<n<10K
---

# DBBench + ALFWorld SFT Dataset (Merged)

## Overview

This dataset is a **simple concatenation (merge)** of the following two synthetic SFT datasets:

- **ALFWorld Trajectory Dataset**: `u-10bei/sft_alfworld_trajectory_dataset_v5`
  - https://huggingface.co/datasets/u-10bei/sft_alfworld_trajectory_dataset_v5
- **DBBench SFT Dataset (ReAct Format)**: `u-10bei/dbbench_sft_dataset_react_v4`
  - https://huggingface.co/datasets/u-10bei/dbbench_sft_dataset_react_v4

The goal is to provide a single dataset repo that contains both:

- household task-solving trajectories (ALFWorld-compatible)
- SQL / database operation trajectories (AgentBench DBBench-compatible)

## Statistics

- **Total conversations**: 3702
  - `dbbench`: 1200
  - `alfworld`: 2502

## Data Format

To avoid schema mismatches between the two source datasets, we normalize into a common schema:

- `id` (string): unique sample id
- `source` (string): `"dbbench"` or `"alfworld"`
- `messages` (list): list of `{role, content}` objects
- `metadata_json` (string): JSON dump of the original `metadata` object from the source dataset

### Example

```json
{
  "id": "dbbench_sft_0000",
  "source": "dbbench",
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "metadata_json": "{"sql": "...", ...}"
}
```

## Usage

```python
from datasets import load_dataset
import json

ds = load_dataset("moroqq/dbbench_and_alfworld_sft_dataset")
sample = ds["train"][0]

metadata = json.loads(sample["metadata_json"])
messages = sample["messages"]
```

## Notes

- This repo is **derived** from the two datasets linked above. Please refer to the original dataset cards for details on generation, limitations, and intended use.
- License follows the original datasets: **MIT**.

## Created

- 2026-02-17T22:25:19+00:00 (UTC)