REUSE-QA

REUSE-QA evaluates how tool-using agents reuse work across related questions. This ungated repository contains 7,479 questions in 2,488 families from FinanceBench and SQL-API-Bench. The full benchmark contains 7,797 questions in 2,627 families, including 318 GAIA questions available in the gaia companion repository. Each family links an original question to variants that change its wording, inputs, or conditions. Researchers can use these streams to compare answer caching, plan reuse, and other reuse policies on answer accuracy and model cost.

Contents

The default configuration contains two source splits, financebench and sql_api_bench, with 7,479 questions: 2,488 originals and 4,991 variants (2,488 P, 604 ER, and 1,899 CA).

The SQL databases are downloaded separately with the preparation script below.

The following table describes the full three-source benchmark. Its GAIA row is distributed in the companion repository; this repository contains the FinanceBench and SQL-API-Bench rows.

Source	Questions	Originals	P	ER	CA
FinanceBench	453	150	150	101	52
GAIA (Levels 1–2)	318	139	139	17	23
SQL-API-Bench	7,026	2,338	2,338	503	1,847
Total	7,797	2,627	2,627	621	1,922

Variants use these variant_type values:

Paper abbreviation	Stored value	Change
P	paraphrase	Rewords the original while retaining its reference answer.
ER	entity_replacement	Changes a query-specific entity, value, or requested attribute.
CA	condition_augmentation	Adds or revises a constraint, potentially requiring further filtering or checking.

Originals have variant_type="original". IDs may retain legacy variant1, variant2, or variant3 markers for P, ER, and CA, respectively. Other ID patterns also occur; use variant_type and parent_id rather than parsing an ID to determine a record's role.

Load the questions

Install datasets if needed:

pip install datasets

from datasets import load_dataset

stream = load_dataset(
    "RECAP-data/reuse-qa", "default", split="financebench"
)
print(stream[0]["question"])

The files are UTF-8 JSON Lines. From a downloaded copy of the repository:

import json

with open("data/financebench.jsonl", encoding="utf-8") as handle:
    records = [json.loads(line) for line in handle]

by_id = {record["id"]: record for record in records}
variant = next(record for record in records if record["parent_id"] is not None)
original = by_id[variant["parent_id"]]

Record format

Every record has the same eight top-level fields:

Field	Meaning
id	Unique record identifier.
parent_id	Original question's ID for a variant; null for an original. The parent is in the same split.
dataset	Source split name.
variant_type	original, paraphrase, entity_replacement, or condition_augmentation.
question	Question presented to the agent.
answer	Reference answer, stored as a string.
answer_type	text, numeric, or json_rows.
context	Source-specific task inputs and database references.

In the Hugging Face viewer, each question is one row with these eight columns. Select a source using the split dropdown. context is a structured object within one cell; its individual fields are not separate top-level columns. SQL reference rows remain JSON text in the answer cell.

context retains the benchmark's eight-field structure: db_name, company, doc_name, evidence, constraint_type, file_name, file_path, and level. FinanceBench supplies supporting passages in evidence; SQL questions identify their database through db_name. Unused fields are null, except that unused evidence is an empty list.

Keep the reference answer out of the agent's inputs during evaluation.

For SQL records, parse answer with json.loads(record["answer"]) to obtain the reference result rows. Preserve their column names and values when scoring.

Evaluate reuse across a stream

Each original appears before its variants in the JSONL files. Preserve the file order for online reuse evaluation, begin with an empty cache or plan library, and allow reuse only from earlier questions. Record the dataset revision and stream order with results.

Tools and supporting data

FinanceBench uses an oracle-evidence setting: a document tool returns the supporting passages in context.evidence.

SQL-API-Bench combines SQL with scalar tools such as primality checks and geocoding. The questions identify one of 26 Spider databases through context.db_name. To prepare the databases separately, run:

pip install gdown
python scripts/prepare_sql_api_bench.py --work-dir ./sql_api_bench_data

The script downloads and extracts the upstream Spider archive, retains its notices, and reports the database directory for the bundled questions. The complete upstream download is larger than the 26 databases needed here.

If Spider is already available locally, no download is needed:

python scripts/prepare_sql_api_bench.py --spider-dir /path/to/spider

Resolve a database as <database_dir>/<db_name>/<db_name>.sqlite, using the directory printed by the script. The databases retain Spider's source terms. Scalar tools and the agent execution environment are separate dependencies.

To comply with GAIA’s redistribution requirements, its questions, reference answers, and all 31 referenced attachments are distributed separately in the gated companion repository, which provides its loading and tool instructions.

Sources and terms of use

Source materials retain their upstream terms, as detailed in LICENSE. FinanceBench is licensed under CC BY-NC 4.0; SQL-API-Bench specifies CC BY-NC-SA 4.0. The separately downloaded databases come from Spider, whose data are licensed under CC BY-SA 4.0. This compilation includes transformed questions and counterfactual FinanceBench evidence; it does not replace the source terms or grant additional rights in third-party content.

Citation

Please cite the accompanying manuscript, RECAP: Structured Plan Reuse for Efficient Tool-Using Agents, and the source benchmark:

• Islam et al. FinanceBench: A New Benchmark for Financial Question Answering (2023). Paper.
• Mialon et al. GAIA: A Benchmark for General AI Assistants. ICLR (2024). Dataset and citation.
• Khabiri et al. Declarative Techniques for NL Queries over Heterogeneous Data. EMNLP Industry Track (2025). Paper.
• Yu et al. Spider: A Large-Scale Human-Labeled Dataset for Complex and Cross-Domain Semantic Parsing and Text-to-SQL Task. EMNLP (2018). Paper.