ARCS

ARCS (Ambiguity Resolution Corpus for SQL) is a text-to-SQL benchmark featuring naturally occurring, unconstrained ambiguities over real-world databases, with complete annotations of valid ambiguity points, interpretations, and SQL queries.

As text-to-SQL systems move beyond demonstrations toward real-world deployment, ambiguity in user questions becomes a primary source of errors. These ambiguities are often subtle, domain- or data-specific, and can silently cause system outputs to deviate from the user's true intent.

ARCS contains 101 unique questions across six SQLite databases. The test split expands them into 311 end-to-end instances with an intended resolution, while tasks/base_tasks.json preserves the original questions and all valid interpretations.

Website: https://megagonlabs.github.io/tabulaflow/research/arcs/

Code: https://github.com/megagonlabs/tabulaflow/blob/main/tabulaflow/research/ARCS.md

Sample Task

{
  "qid": "001",
  "db": "retails",
  "question": "Report the total revenue for each nation in 1995.",
  "gold_ambiguity_points": [
    {
      "phrase": "total revenue",
      "interpretations": [
        "before discount (Gross revenue)",
        "after discount (Net revenue)"
      ]
    },
    {
      "phrase": "total revenue",
      "interpretations": [
        "include returned items",
        "exclude returned items"
      ]
    },
    {
      "phrase": "for each nation",
      "interpretations": [
        "for each customer nation",
        "for each supplier nation"
      ]
    },
    {
      "phrase": "in 1995",
      "interpretations": [
        "order date in 1995",
        "commitment date in 1995",
        "shipment date in 1995",
        "receipt date in 1995"
      ]
    }
  ]
}

🚀 Quickstart with TabulaFlow

TabulaFlow provides the supported ARCS loader, structured-disambiguation agents, and evaluation metrics. It downloads the pinned release, opens the SQLite databases locally, and records predictions, scores, trajectories, token usage, and latency.

These instructions are also available in the benchmark guide.

First, install uv, then install TabulaFlow and download ARCS (approximately 9 GB):

uv tool install tabulaflow
tabulaflow benchmark download arcs

Run five test instances with a configured model provider:

export OPENAI_API_KEY="your-api-key"

tabulaflow benchmark run arcs \
  --split test \
  --llm openai:gpt-6-luna \
  --sample-size 5

ARCS defaults to TabulaFlow's structured ambiguity agent. The agent identifies ambiguity points, resolves them against the intended user response, generates SQL, and evaluates both disambiguation and execution quality.

Files

• tasks.jsonl — canonical test split, with one row per intended resolution
• tasks/base_tasks.json — 101 original questions before resolution expansion
• tasks/intended_query_ids.json — intended query IDs used for expansion
• tasks/readable/ — human-readable tasks, gold SQL, and expected results
• databases/sqlite/*.sqlite — six referenced SQLite databases
• databases/column_meanings.json — column-level descriptions

Embedded execution-result DataFrames use TabulaFlow's parquet_v1 serialization and are loaded by its ARCS benchmark loader.