Skip to content

Dataset · facebook

NaturalReasoning: reasoning tasks

facebook/natural_reasoning

A large set of complex questions with reference answers and generated solutions.

What's inside

The NaturalReasoning dataset is a large-scale general reasoning task set. Questions are back-translated from the DCLM and FineMath pre-training corpora, deduplicated, and cleaned of overlaps with MATH, GPQA, MMLU-Pro, and MMLU-STEM. For each question where possible, there is a reference final answer from the source document and a model-generated answer; language is English, size is 1M to 10M examples.

What you can do with it

  • Fine-tune LLM reasoning on SFT
  • Test model robustness to complex questions
  • Create LoRA for general reasoning
Tasks
Text generation
Languages
English
Size
1M–10M rows
Downloads this month
9.6K
Likes
600

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("facebook/natural_reasoning")