STEM Reasoning · facebook
principia-collection
A STEM reasoning dataset for training language models to derive mathematical objects and numerical answers.
What principia-collection is
Principia Collection is a large-scale STEM reasoning dataset developed by Meta. It is designed to train and evaluate the capacity of language models to deduce mathematical objects and numerical solutions from scientific problem statements spanning mathematics, physics, and chemistry.
The dataset addresses the challenge of eliciting structured, multi-step scientific reasoning in artificial intelligence models. Rather than limiting evaluations to simple scalar values, it supplies tasks categorised under standardised classification schemes, specifically Mathematical Subject Classification (MSC 2020) and Physics Subject Headings (PhySH).
How it was made
Each entry in the collection provides a problem statement, a ground truth label, an assigned answer type, and a topic classification. Both the problem statements and ground truth solutions are produced via a synthetic data generation pipeline using GPT-OSS-120B as the proposer model.
The data is structured into two main splits.
Key facts
- Proposer model
- GPT-OSS-120B
- Taxonomies
- MSC 2020 and PhySH
- Data format
- Parquet
- Language
- English
What you can use it for
- Mathematical object derivation
- Fine-tuning language models to output structured mathematical entities such as matrices, sets, intervals, piecewise functions, and algebraic systems.
- Scientific benchmark enhancement
- Training reasoning models to improve performance on advanced scientific evaluation suites such as SuperGPQA, GPQA-Diamond, and AIME.
- Unit-aware numerical reasoning
- Teaching language models to compute exact numeric outputs across integer, decimal, and fractional formats while accounting for scientific units.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from facebook/principia-collection at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 248,743 rows of default/mathematical_object
answer
string_text- min
- 1
- median
- 134
- mean
- 263.32
- max
- 16,481
answer_type
string_label- Set19%
- Interval17%
- Equation17%
problem_statement
string_text- min
- 29
- median
- 1,292
- mean
- 1355.61
- max
- 5,998
topic
string_text- min
- 35
- median
- 113
- mean
- 116.84
- max
- 295
Strengths and limits
Good at
- Problems are indexed directly to formal scientific classification schemes from MSC 2020 and PhySH.
- The dataset tests both structured non-scalar mathematical objects and unit-dependent numerical calculations.
- Training on the numerical subset provides consistent performance gains across benchmarks including SuperGPQA, GPQA-Diamond, and AIME.
Watch out for
- All problem statements and answers are synthetically generated by a proposer model rather than sourced from human-authored scientific literature.
- Full technical details of the generation pipeline and empirical benchmark results depend on a forthcoming paper.
Questions
Can I look inside principia-collection without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is principia-collection under?
Its authors declare cc-by-nc-sa-4.0. Read the terms on the original page before reusing the data.
What answer types are present in the mathematical object subset?
The mathematical object subset includes equations, inequalities, intervals, sets, matrices, and piecewise functions, with a median answer length of 73 tokens.
What answer types are included in the numerical subset?
The numerical split contains integers, decimals, and fractions, balanced across variants that either include or omit measurement units.
How are the scientific topics organised?
Topics are mapped to subject codes from Physics Subject Headings (PhySH) and Mathematical Subject Classification (MSC 2020).
How was the dataset generated?
All problems and ground truth answers were produced synthetically using GPT-OSS-120B as the proposer model.
Written from principia-collection's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.