Datasets:

summerAIO
/

LongBook-math-1M

license: mit
language:
  - ja
  - en
tags:
  - mathematics
  - synthetic
  - japanese
  - retrieval
task_categories:
  - question-answering
pretty_name: LongBook Math 1M (JA synthetic)
size_categories:
  - 100K<n<1M

LongBook Math 1M (Japanese synthetic mathematics dataset)

1,000,000 deterministic synthetic Japanese mathematics documents for retrieval evaluation and generic-model RAG, built for LongBook.

Format: JSONL, one document per line: {id, namespace, text, metadata{domain, difficulty, lang, kind, index}}
id = math-{domain}-{index:07d}, unique and sortable.
text structure (Japanese): 【分野】/【問題 id】/【解法】/【答え】/【キーワード】/【確認】
Answers are computed in Python by the generator, not hallucinated.
8 domains × 3 difficulties: arithmetic / algebra / calculus / linear_algebra / probability / geometry / number_theory / logic_sets
Generator: dataset/build_math_dataset.py (--seed 42), FTS5 ingest via dataset/bulk_ingest_math_dataset.py.
Reading guide for sub-agents: dataset/MATH_SUBAGENT.md (model-agnostic, default Qwen/Qwen3.5-9B).

Usage (datasets)

from datasets import load_dataset
ds = load_dataset("summerMC/LongBook-math-1M", split="train", data_files="math-*.jsonl")
print(ds[0])

Usage (LongBook retrieval)

python3 dataset/ingest_math_dataset.py --input <downloaded dir> --namespace math --db ./data/longbook.sqlite3
curl http://127.0.0.1:30000/v1/longbook/search \
  -H 'Content-Type: application/json' \
  -d '{"query":"二次方程式 因数分解","namespace":"math","top_k":5}'