Vicen-te's picture
Super-squash branch 'main' using huggingface_hub
00d3f74
|
Raw
History Blame Contribute Delete
1.21 kB
---
language:
- en
license: cc-by-4.0
size_categories:
- n<1K
task_categories:
- text-generation
- table-question-answering
pretty_name: SQL Create Context (mini)
tags:
- sql
- text-to-sql
- fine-tuning
- qwen3.5
---
# SQL Create Context (mini) — 300 train / 200 eval
A curated, deduplicated 500-row split of
[`b-mc2/sql-create-context`](https://huggingface.co/datasets/b-mc2/sql-create-context)
used to LoRA-fine-tune Qwen3.5-2B for Text-to-SQL.
## Schema
Each row has three string fields:
| field | description |
|------------|------------|
| `schema` | A `CREATE TABLE ...` statement (SQLite-flavoured) |
| `question` | Natural-language question |
| `answer` | Gold SQL query |
## Splits
| split | rows |
|-------|------|
| train | 300 |
| eval | 200 |
## Provenance
Filtered to schemas ≤ 1500 chars and deduplicated on the full triple. Sampled
with `seed=42` from the source train split. No examples leak between splits.
## Intended use
This split is intentionally small (500 rows) for parameter-efficient
fine-tuning on a single GPU. It is not a capability benchmark; use the
upstream Spider/BIRD datasets for that.
## Licensing
Inherits the source dataset's CC-BY-4.0 license.