INSAIT-Institute/SaberMath-documents

---

SABER-Math Documents

Document corpus for the SABER-Math mathematical information reranking benchmark. Each of the 71,117 entries is a competition-style math problem together with its solution, serving as a retrievable document. The candidates field in SABER-Math Queries indexes into this corpus.

Dataset structure

Each example contains:

Field	Type	Description
id	string	Unique document identifier
problem	string	Problem statement
solution	string	Full worked solution to the problem
answer	string	Final answer
idea	string	Short description of the key idea / proof technique
tags	list[string]	Topic tags for the problem
domains	list[string]	Mathematical domains (e.g. algebra, number theory)
metadata	struct	Provenance: competition, source, originating dataset, and the raw forum posts (BBCode/HTML content, timestamps, authors, attachments) the problem was extracted from
original_index	int64	Index of the document in the source dataset

Document indices referenced by the candidates lists in SABER-Math Queries correspond to row positions in this corpus.

Inspection

from datasets import load_dataset

docs = load_dataset("INSAIT-Institute/SaberMath-documents", split="train")
queries = load_dataset("INSAIT-Institute/SaberMath-queries", split="train")

q = queries[0]
print(q["problem"])
for idx, rel in zip(q["candidates"], q["relevance_scores"]):
    print(rel, docs[idx]["problem"][:80])

Follow the code in the GitHub repository to run the end to end evaluation.

Licensing

The CC BY-SA 4.0 license applies to the benchmark material authored by us (annotations, judgments, curation, and derived fields). Redistributed source text remains subject to the terms of its original source:

• NuminaMath: retains the conditions of its original Apache 2.0 license.
• AoPS: not available for model training and subject to the applicable AoPS restrictions, which require attribution when the resources are used as input to an AI system. These entries are explicitly marked in the dataset.
• Official competition problems: typically in the public domain or released under usage conditions similar to the above, forbidding the use of the data for training. Users are responsible for complying with the original source terms for each entry. See the metadata fields (dataset, source, competition) to identify the provenance of individual examples.

Citation

You can check out the full paper on ArXiv.

@inproceedings{
  georgiev2026sabermath,
  title={{SABER}-Math: An Automated Reranking Benchmark for Mathematical Information Retrieval},
  author={Anonymous},
  booktitle={The 2026 Conference on Empirical Methods in Natural Language Processing},
  year={2026},
  url={https://openreview.net/forum?id=Tb2EKLtAS0}
}