SAT Multi-Step Math Evidence Retrieval Benchmark

#retrieval #evidence ranking #multi-step reasoning #SAT math solution retrieval #mathematical evidence verification #multi-step problem solving

5000
0.3GB
JSON
CATL
MOBIUSI INC

Updated:2026-10-19

AI Analysis & Value Prop

This benchmark contains SAT math problems requiring multi-step derivations, a candidate evidence library of concepts and formulas, and annotations identifying the evidence needed at each step along with complete evidence chains. It assesses whether retrieval-augmented generation systems retrieve knowledge that is both complete across the solution steps and applicable to the problem, rather than merely returning one related formula. It supports SAT math knowledge retrieval, evidence-chain evaluation, and the development of retrieval-augmented generation systems.

Dataset Insights

Sample Examples

{}

Technical Specifications

Field	Type	Description
file_name	string	File name
chain_summary	object	Calculated summary statistics for the multi-step evidence chain, including its size, dependency depth, and completeness.
source_record	object	The original record supplied during dataset creation, containing the problem, candidate evidence library, and complete step-by-step annotations.
evidence_chain	array	An ordered evidence chain derived from the original step annotations, representing the complete and applicable knowledge required at each reasoning step.
retrieval_evaluation_target	object	A RAG retrieval evaluation target derived from the reference evidence chain, emphasizing evidence-set completeness and applicability.

Compliance Statement

Authorization Type	Proprietary - Commercial AI Training License (No Redistribution)
Commercial Use	Requires exclusive subscription or authorization contract (monthly or per-invocation charging)
Privacy and Anonymization	No PII, no real company names, simulated scenarios follow industry standards
Compliance System	Compliant with China's Data Security Law / EU GDPR / supports enterprise data access logs

Frequently Asked Questions

What does the dataset contain?: It includes SAT math problems that require multi-step derivations, a candidate library of concepts and formulas, annotations for the evidence needed at each step, normalized evidence chains, and retrieval evaluation targets.

What capability does it evaluate?: It focuses on whether a RAG system retrieves knowledge that is complete and applicable across the steps of a solution, rather than merely returning a superficially related formula or concept.

Which use cases is it designed for?: It can support SAT math knowledge retrieval, evaluation of retrieval-augmented generation systems, multi-step math question-answering research, and analysis of evidence-chain completeness and knowledge grounding.

How can I obtain the dataset and verify its license?: Check the dataset product page or contact the provider for access options, license scope, and usage restrictions. Confirm the applicable terms before commercial deployment, redistribution, or public release of evaluation results.

Can't find the data you need?

Post a request and let data providers reach out to you.

Get this Dataset

Verified for Enterprise Use

Cite this Work

@dataset{Mobiusi2026,
  title={SAT Multi-Step Math Evidence Retrieval Benchmark},
  author={MOBIUSI INC},
  year={2026},
  url={https://www.mobiusi.com/datasets/97a95b238987a2d02fde77652e27e4e5?dataset_task_cate_id=10},
  urldate={2026-10-14},
  keywords={SAT math dataset,multi-step math problems,evidence retrieval benchmark,RAG evaluation data,math formula retrieval,knowledge grounding,evidence chain annotations,SAT preparation data},
  version={1.0}
}

Using this in research? Please cite us.

placeholder

Products

Scene

Domain

Modality

Task