---
title: livesqlbench-data
canonical_url: "https://www.modelscope.cn/datasets/yangsaisai/livesqlbench-data"
md_url: "https://www.modelscope.cn/datasets/yangsaisai/livesqlbench-data.md"
repository: yangsaisai/livesqlbench-data
last_updated: 2026-10-27
license: mit
storage_size: "2.6 GB"
downloads: 7083
stars: 0
---

# livesqlbench-data

> livesqlbench-data - yangsaisai 在 ModelScope 开源的数据集。LiveSQLBench tasks, separate evaluation references, and database assets; 1620 tasks across four variants. No generated SFT trajectories.

yangsaisai/livesqlbench-data 是 ModelScope 魔搭社区上的数据集，存储大小 2.6 GB，采用 mit 许可。

- **Repository**: yangsaisai/livesqlbench-data
- **License**: mit
- **Storage size**: 2.6 GB
- **Downloads**: 7083
- **Stars**: 0
- **Last updated**: 2026-10-27

Source: https://www.modelscope.cn/datasets/yangsaisai/livesqlbench-data

---

# LiveSQLBench data

This repository packages task inputs, separate evaluation references, and database assets extracted from [bird-bench/livesqlbench](https://github.com/bird-bench/livesqlbench) for the TableGPT-V3 LiveSQLBench integration. It is a derived data package, not an official upstream release. It contains no generated agent trajectories or completed SFT examples.

## Contents

| Dataset | Tasks | Database format |
| --- | ---: | --- |
| livesqlbench-base-lite | 270 | PostgreSQL SQL dumps |
| livesqlbench-base-full-v1 | 600 | PostgreSQL SQL dumps |
| livesqlbench-large-v1 | 480 | PostgreSQL SQL dumps |
| livesqlbench-base-lite-SQLite | 270 | SQLite files |
| Total | 1620 | |

- `tasks.jsonl`: task inputs, identified by `(dataset, instance_id)`. Asset paths are relative to the repository root.
- `references.jsonl`: corresponding reference SQL, tests, preprocessing and cleanup SQL, and other source fields. Keep reference answers and tests out of model inputs and agent workspaces during evaluation.
- `assets/`: schema text, column descriptions, knowledge files, PostgreSQL SQL dumps, and SQLite databases, organized by dataset and database.
- `manifest.json`: upstream source, task counts, and package status. The original machine-specific source directory is replaced by the upstream repository URL in this publication copy.
- `task_annotations.jsonl`: one local review annotation for `livesqlbench-base-lite/credit_2`, documenting an ambiguity between the query and reference SQL. It preserves the original evaluation verdict and does not automatically alter task execution. Its evidence path points to a local review artifact that is not included in this package.
- `SHA256SUMS`: SHA-256 checksums of the published files, excluding this checksum file itself.
- `LICENSE`: original MIT license, copyright (c) 2024 bird_sql.

Task records, reference records, database assets, and the original license are copied without changes. The upstream commit used for the original extraction was not recorded in the extraction manifest; this package does not claim an exact upstream revision. These resources have not all been independently validated task by task. PostgreSQL restoration and evaluation require the corresponding runtime and evaluator; downloading the data alone does not configure them.

## Download

Install the ModelScope SDK (`pip install modelscope`), then run:

```python
from modelscope import dataset_snapshot_download

dataset_snapshot_download(
    dataset_id="yangsaisai/livesqlbench-data",
    local_dir="sft_generation/livesqlbench/data",
)
```

Use a new or empty destination directory. The downloaded directory is the data root; do not add another enclosing `data/` level. To verify the downloaded files on Linux, run `sha256sum -c SHA256SUMS` from that directory; on macOS, use `shasum -a 256 -c SHA256SUMS`.

## License and attribution

The source repository provides the MIT license preserved in `LICENSE`. Retain its copyright and permission notice when redistributing. Refer to the upstream project for benchmark methodology and citation details.