README.md

metadata

license: cc-by-4.0
language:
  - en
pretty_name: Unsolved Math Clean
tags:
  - mathematics
  - open-problems
  - research
  - benchmark
  - reasoning
  - latex
  - clean
task_categories:
  - question-answering
  - text-generation
size_categories:
  - 1K<n<10K
configs:
  - config_name: problems
    default: true
    data_files:
      - split: train
        path: data/problems/train.parquet
      - split: validation
        path: data/problems/validation.parquet
      - split: test
        path: data/problems/test.parquet
  - config_name: open_problems_eval
    data_files:
      - split: train
        path: data/open_problems_eval/train.parquet
  - config_name: research_prompts
    data_files:
      - split: train
        path: data/research_prompts/train.parquet
      - split: validation
        path: data/research_prompts/validation.parquet
      - split: test
        path: data/research_prompts/test.parquet

🧠 Unsolved Math — Clean

8,626 curated open research problems in mathematics and CS — including 122 Millennium Prize Problems — deduplicated, schema-flattened, and packaged as proper parquet configs with an eval-only benchmark view.

A reasoning frontier dataset: every problem here is actually unsolved or partially solved — ideal for honest capability probing instead of contaminated benchmarks.

Clean derivative of ulamai/UnsolvedMath (8,785 problems). License unchanged: CC-BY-4.0.

🧹 Quality Pipeline

Step	Removed	Reason
Raw input	8,785	Nested JSON
Duplicate titles	153	Same problem listed twice
Empty statements	6	No problem body
Final clean	8,626	Flattened schema, deterministic splits

📊 Composition

By status:

Status	Problems
Open	4,354
Partially solved	3,580
Solved	692

By difficulty:

Level	Problems
L1: Tractable	908
L2: Intermediate	516
L3: Advanced	6,143
L4: Expert	937
L5: Millennium Prize	122

📦 Configs

Config	Rows	Description
problems (default)	8,626	Full metadata: statement, background, status, difficulty, category, provenance
open_problems_eval	4,354	Open problems only — benchmark/eval view
research_prompts	8,626	Instruction-formatted: "analyze this open problem, discuss partial results and obstacles"

🎯 Usage

from datasets import load_dataset

ds = load_dataset("saidutta69/unsolved-math-clean", "problems", split="train")

# Honest eval: only truly open problems
ev = load_dataset("saidutta69/unsolved-math-clean", "open_problems_eval")

# Research-style SFT prompts
rp = load_dataset("saidutta69/unsolved-math-clean", "research_prompts")
print(rp["train"][0]["instruction"][:300])

⚠️ Notes

• These are unsolved problems — there are no reference answers. Use for capability probing, calibration of hedging behavior, and research discussion; not for accuracy scoring.
• LaTeX notation preserved verbatim.
• Millennium-level problems are intentionally near-impossible; expect models to fail honestly.

📜 Citation

@misc{unsolved-math-clean,
  author = {Sai Dutta Abhishek Dash (clean); original by ulamai},
  title = {Unsolved Math — Clean},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/saidutta69/unsolved-math-clean}}
}