Datasets:

kiaw73b5o7a
/

arcs-bench

license: cc-by-4.0
language:
  - en
task_categories:
  - text-generation
  - question-answering
tags:
  - text-to-sql
  - ambiguity
configs:
  - config_name: default
    data_files: tasks.jsonl
size_categories:
  - n<1K

ARCS

Ambiguity Resolution Corpus for SQL. 311 task instances across 6 databases.

Files

tasks.jsonl — one row per intended interpretation.
databases/sqlite/*.sqlite — databases referenced by the db field.
databases/column_meanings.json — column-level descriptions.

Loading

from datasets import load_dataset
ds = load_dataset("kiaw73b5o7a/arcs-bench", split="train")

Most rows have the gold SQL string (gold_queries[*].query) masked to null to avoid label leakage; a small sample retains it. Other gold_queries fields (e.g. exec_result) are kept on every row.