Note: This repository is a fork of the original OpenAI miniF2F repository https://github.com/openai/miniF2F, with additional data and many formal statement fixes.
If you find errors in the statement formalizations, please submit a pull request. We will continue updating the dataset. Please cite the version you used by commit or date for reproducible evaluations.
We will not accept pull requests with additional proofs to reduce test set contamination.
MiniF2F is a formal mathematics benchmark (translated across multiple formal systems) consisting of exercise statements from olympiads (AMC, AIME, IMO) as well as high-school and undergraduate maths classes.
The goal of the project is to provide a shared benchmark to evaluate and directly compare automated theorem proving systems based on the formal systems targeted, initially Lean, Isabelle, and Metamath (targeting also Hol Light).
The benchmark (released under permissive licenses (MIT for Metamath, Apache for Lean)) is a work in progress and contributions are welcome and encouraged through pull requests.
The initial version of the benchmark is described in detail in the following pre-print:
@article{zheng2021minif2f,
title={MiniF2F: a cross-system benchmark for formal Olympiad-level mathematics},
author={Zheng, Kunhao and Han, Jesse Michael and Polu, Stanislas},
journal={arXiv preprint arXiv:2109.00110},
year={2021}
}
The original repo is miniF2F. It has then seen significant fixes and improvements, notably the addition of an informal statement and an informal proof for each problem. The curation of the informal component is described in the following paper. To cite it:
@inproceedings{
2210.12283,
title={Draft, Sketch, and Prove: Guiding Formal Theorem Provers with Informal Proofs},
author={Albert Q. Jiang and Sean Welleck and Jin Peng Zhou and Wenda Li and Jiacheng Liu and Mateja Jamnik and Timothée Lacroix and Yuhuai Wu and Guillaume Lample},
booktitle={Submitted to The Eleventh International Conference on Learning Representations},
year={2022},
url={https://arxiv.org/abs/2210.12283}
}
We decided to start a separate repository, instead of submitting PRs, for better maintainence of the dataset.
Test | Valid | |
---|---|---|
Lean | 244 | 244 |
Metamath | 244 | 244 |
Isabelle | 244 | 244 |
Hol Light | 165 | 165 |
Informal | 244 | 244 |
Solve for
theorem mathd_algebra_17
(a : ℝ)
(h₀ : real.sqrt (4 + real.sqrt (16 + 16 * a)) + real.sqrt (1 + real.sqrt (1 + a)) = 6) :
a = 8 :=
begin
sorry
end
theorem mathd_algebra_17:
fixes a :: real
assumes "1 + a>0"
assumes "sqrt (4 + sqrt (16 + 16 * a))
+ sqrt (1 + sqrt (1 + a)) = 6"
shows "a = 8"
sorry
let mathd-algebra-17 = `!a. sqrt (&4 + sqrt (&16 + &16 * a)) + sqrt (&1 + sqrt (&1 + a)) = &6 /\ &0 <= (&1 + a) ==> a = &8`;;
$(
@{
mathd-algebra-17.0 @e |- ( ph -> A e. RR ) $@
mathd-algebra-17.1 @e |- ( ph -> ( ( sqrt ` ( 4 + ( sqrt ` ( ; 1 6 + ( ; 1 6 x. A ) ) ) ) ) + ( sqrt ` ( 1 + ( sqrt ` ( 1 + A ) ) ) ) ) = 6 ) $@
mathd-algebra-17 @p |- ( ph -> A = 8 ) @=
? @.
@}
$)
Each problem is represented by a unique name and a file for each of the formal systems we target. Each file consists at minima in the problem statement and optionally one or more example proofs associated with it. The benchmark is divided in two splits:
valid
: validation set that can be used while designing automated theorem proving systems (early-stopping, reinforcement learning, data-augmentation, curriculum design, ...).test
: held-out test set reserved for final evaluation.
Naming conventions are still a work in progress. Olympiads problems are generally named after their
competition year and problem number (eg. imo-1990-p3
or aime-1983-p2
). Problems coming from a
particular dataset (eg the MATH dataset) are named to ease their
retrieval (eg. mathd-algebra-125
). Other problems are prefixed by a category hint and a unique
name in the style of Metamath naming conventions (eg. induction-11div10tonmn1ton
).
Each exercise file complies to the following system-specific conventions.
To install the project make sure you have elan installed, then in the directory where you want the project installed run:
git clone https://github.com/openai/miniF2F
cd miniF2F
leanpkg configure
leanproject get-mathlib-cache
leanproject build
Since having one file per statement causes slowness in Lean parsing stage, all Lean statements are
exceptionally aggregated in two files (valid.lean
and test.lean
). These files contain a list of
the problem statements defined as theorem
s. Optionally, proofs for these statements are provided
as well as potential lemmas to support the ground-truth proof.
No theorem
should appear that do not correspond to a problem statement; use lemma
instead.
Please use lean/scripts/lint_style.py
to check all the statements pass the linter. You can also
make use of lean/scripts/simple_formatter.sh
to enforce a few basic formatting rules.
The lean
folder is released under the Apache License (so that it is aligned with Lean's mathlib
license).
Each file contains the problem statement with the same name as the problem unique name. The statement is commented (using Metamath convention) if provided without proof.
The metamath
folder is released under the MIT License.
Each file contains the problem statement defined as a HOL Light term whose name must match the file name.
The hollight
folder is released under the FreeBSD License.
Each file contains the problem statement defined as a theorem whose name must match the file name, optionally with a proof for it as well as the necessary imports.
The isabelle
folder is released under the Apache License.
Each file contains the problem statement and the proof written in natural mathematical language. The data come from the following sources:
- The MATH dataset.
- The AOPS website.
- Manual annotation by Albert Qiaochu Jiang, Timothée Lacroix, Guillaume Lample, Sean Welleck, Jiacheng Liu, and Marie-Anne Lachaux.
MiniF2F is meant to serve as a shared and useful resource for the machine learning community working on formal mathematics.
There is no obligation tied with the use and reporting of a result based on miniF2F. But if you're using it and discovering new proofs (manually or automatically) please contribute them back to the benchmark.
All contributions, such as new statements for later versions, addition of missing statements for existing versions, bug fixes, additional proofs are all welcome.
A version of miniF2F is defined by a frozen set of statements. The goal for each version is to get full coverage on all formal systems for that version even if that might not be the case when the version is frozen.
When reporting a result based on miniF2F please always specify the version you used. The current
version is v2
, frozen as of October 2022, including 488 statements (fully translated to Lean, Isabelle, and Metamath but still WIP in other formal systems).
Each version will live in its own branch to allow later additions of translated statements or fixes
to existing statements as needed. The main
branch remains reserved for active development and
should not be used when reporting results.
- Version
v2
- Freeze date: October 2022
- Branch: v2
- Version:
v1
- Freeze date: August 2021
- Branch: v1