Dataset Viewer (First 5GB)
Auto-converted to Parquet Duplicate
problem
string
source
string
formal_statement
string
lean_header
string
url
string
user_name
string
user_url
string
sft_line_number
int64
messages
list
uuid
string
used_in
list
tools
list
license
string
Prove that there exists a polynomial $ P \in \mathbb{Z}[X] $ such that the number $$ \sqrt[2003]{5 - 2\sqrt{6}} + \sqrt[2003]{5 + 2\sqrt{6}} $$ is a root of $ P $.
aops
theorem problem_568570 : ∃ P : Polynomial ℤ, IsRoot (P.map (Int.castRingHom ℝ)) ((5 - 2 * Real.sqrt (6 : ℝ)) ^ (1 / (2003 : ℝ)) + (5 + 2 * Real.sqrt (6 : ℝ)) ^ (1 / (2003 : ℝ))) := by sorry
import Mathlib import Aesop import Mathlib set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat open Polynomial
null
null
null
null
[]
b2ee7144-44aa-5d7f-8acc-812fae259c90
[]
[]
cc-by-4.0
Prove that the matrix \( \mathbf{V} = (\mathbf{I} - 2t\mathbf{A\Sigma})^{-1}\mathbf{\Sigma} \) is symmetric.
mathstack
theorem problem_67761 {n : Type*} [Fintype n] [DecidableEq n] (t : ℝ) (A Sigma : Matrix n n ℝ) (hA_symm : Aᵀ = A) (hSigma_symm : Sigmaᵀ = Sigma) (h_comm : A * Sigma = Sigma * A) : ((1 - (2 * t) • (A * Sigma))⁻¹ * Sigma)ᵀ = ((1 - (2 * t) • (A * Sigma))⁻¹ * Sigma) := by sorry
import Mathlib import Aesop import Mathlib.Data.Matrix.Basic import Mathlib.LinearAlgebra.Matrix.NonsingularInverse set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat open Matrix
https://math.stackexchange.com/questions/4425561/mgf-of-quadratic-form-mathbf-y-intercal-mathbf-a-mathbf-y-where-y-sim-n-p-mathbf
Soyuz42
https://math.stackexchange.com/users/813572/soyuz42
null
[]
199096ab-5005-570f-8f5d-02320dfafc7e
[]
[]
cc-by-sa-4.0
Let $ A $ be a matrix with integer entries. An operation is defined as adding the same integer to all entries in a single row or a single column. It is given that for infinitely many positive integers $ n $, it is possible to obtain, through a finite number of operations, a matrix in which all entries are divisible by ...
aops
inductive MatrixOp (ι κ : Type*) where | rowAdd (i : ι) (c : ℤ) : MatrixOp ι κ | colAdd (j : κ) (c : ℤ) : MatrixOp ι κ namespace MatrixOp variable {ι κ : Type*} [DecidableEq ι] [DecidableEq κ] def apply_op (M : Matrix ι κ ℤ) : MatrixOp ι κ → Matrix ι κ ℤ | rowAdd i c => fun i' j' => if h : i' = i then M i' j' + c...
import Mathlib import Aesop import Mathlib.Data.Matrix.Basic import Mathlib.Data.Fintype.Basic import Mathlib.Data.Set.Finite import Mathlib.Tactic set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat open Matrix
null
null
null
null
[]
ff9b3b6c-ff59-52eb-a5d6-b7b9b412d467
[]
[]
cc-by-4.0
Let \( V \) and \( W \) be finite-dimensional vector spaces, and let \( P: V \rightarrow V \) and \( Q: W \rightarrow W \) be isomorphisms. Prove that \( \dim(\operatorname{Im}(T)) = \dim(\operatorname{Im}(QTP)) \).
mathstack
universe u v w namespace Problem546038 class MyDivisionRing (K : Type u) : Prop where dummy : True class MyAddCommGroup (V : Type v) : Prop where dummy : True class MyModule (K : Type u) (V : Type v) [MyDivisionRing K] [MyAddCommGroup V] : Prop where dummy : True class MyFiniteDimensional (K : Type u) (V : Type v...
import Mathlib import Aesop set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat
https://math.stackexchange.com/questions/1702380/how-to-prove-isomorphism-maintains-dimension-of-kernel-space
q.Then
https://math.stackexchange.com/users/222237/q-then
null
[]
5200837a-f140-5831-803a-b039ebdfdfb6
[]
[]
cc-by-sa-4.0
Verify that $11^{λ(m)+1} ≡ 11 \pmod m$ for $m = 41 \cdot 11$. Given that $λ(m) = 40$, show that $11^{41} ≡ 11 \pmod{451}$. Specifically, demonstrate that $11^{41} ≡ 11 \pmod{41}$ and $11^{41} ≡ 11 \pmod{11}$.
mathstack
theorem problem_158563 : (11 ^ 41 ≡ 11 [MOD 451]) ∧ (11 ^ 41 ≡ 11 [MOD 41]) ∧ (11 ^ 41 ≡ 11 [MOD 11]) := by sorry
import Mathlib import Aesop import Mathlib.Data.Nat.ModEq set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat
https://math.stackexchange.com/questions/716063/verifying-that-11-m-1-11-pmod-m-for-m-41-cdot-11
mmm
https://math.stackexchange.com/users/127207/mmm
null
[]
fe1bb771-16c5-5464-b69c-26f5541606db
[]
[]
cc-by-sa-4.0
Given \( r > 0 \) and a function \( f \) satisfying \( g(x) \leq F(x) \) for all \( x \in [a,b] \) and \( g(x) = F(x) \) for some \( x \), show that the open ball \( B(f, r) \) contains a point in \( S \), i.e., there exists a continuous function \( g \) such that \( g(x) < F(x) \) for all \( x \in [a,b] \) and \( d(f,...
mathstack
def ClosedInterval (a b : ℝ) : Type := {x : ℝ // a ≤ x ∧ x ≤ b} theorem problem_539785 {a b : ℝ} (hab : a ≤ b) (F f : ClosedInterval a b → ℝ) (r : ℝ) (hr : 0 < r) (h_le : ∀ x, f x ≤ F x) (h_eq : ∃ x, f x = F x) : ∃ g : ClosedInterval a b → ℝ, (∀ x, g x < F x) ∧ (∀ x, |g x - f x| < r) := by sorry
import Mathlib import Aesop set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat
https://math.stackexchange.com/questions/3088857/metric-space-on-a-family-of-functions
mmmmo
https://math.stackexchange.com/users/590072/mmmmo
null
[]
c74f78d0-3d05-5b38-ae42-7c78afc67841
[]
[]
cc-by-sa-4.0
Is a single point topologically connected?
mathstack
def MyIsPreconnected {α : Type*} [TopologicalSpace α] (s : Set α) : Prop := ∀ (U V : Set α), IsOpen U → IsOpen V → s ⊆ U ∪ V → (U ∩ s).Nonempty → (V ∩ s).Nonempty → (U ∩ V ∩ s).Nonempty def MyIsConnected {α : Type*} [TopologicalSpace α] (s : Set α) : Prop := s.Nonempty ∧ MyIsPreconnected s theorem probl...
import Mathlib import Aesop import Mathlib.Topology.Basic set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat open Set
https://math.stackexchange.com/questions/1577525/is-a-single-point-topologically-connected
PurdueCHE
https://math.stackexchange.com/users/251332/purdueche
175,604
[ { "role": "user", "content": "Complete the following Lean 4 code.\n\n```lean4\nimport Mathlib\nimport Aesop\nimport Mathlib.Topology.Basic\nset_option maxHeartbeats 0\nopen BigOperators Real Nat Topology Rat\nopen Set\n/-- Is a single point topologically connected? -/\ndef MyIsPreconnected {α : Type*} [Topo...
d7207623-6048-5d38-a9d9-62666b1668b2
[]
[]
cc-by-sa-4.0
Is a single point topologically connected?
mathstack
def MyIsPreconnected {α : Type*} [TopologicalSpace α] (s : Set α) : Prop := ∀ (U V : Set α), IsOpen U → IsOpen V → s ⊆ U ∪ V → (U ∩ s).Nonempty → (V ∩ s).Nonempty → (U ∩ V ∩ s).Nonempty def MyIsConnected {α : Type*} [TopologicalSpace α] (s : Set α) : Prop := s.Nonempty ∧ MyIsPreconnected s theorem probl...
import Mathlib import Aesop import Mathlib.Topology.Basic set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat open Set
https://math.stackexchange.com/questions/1577525/is-a-single-point-topologically-connected
PurdueCHE
https://math.stackexchange.com/users/251332/purdueche
201,063
[ { "role": "user", "content": "Here is a proof attempt for the following theorem in Lean4.\n\n```lean4\nimport Mathlib\nimport Aesop\nimport Mathlib.Topology.Basic\nset_option maxHeartbeats 0\nopen BigOperators Real Nat Topology Rat\nopen Set\ndef MyIsPreconnected {α : Type*} [TopologicalSpace α] (s : Set α)...
efb4703b-6b0a-5bf5-8fcc-c732eee41064
[]
[]
cc-by-sa-4.0
Is a single point topologically connected?
mathstack
def MyIsPreconnected {α : Type*} [TopologicalSpace α] (s : Set α) : Prop := ∀ (U V : Set α), IsOpen U → IsOpen V → s ⊆ U ∪ V → (U ∩ s).Nonempty → (V ∩ s).Nonempty → (U ∩ V ∩ s).Nonempty def MyIsConnected {α : Type*} [TopologicalSpace α] (s : Set α) : Prop := s.Nonempty ∧ MyIsPreconnected s theorem probl...
import Mathlib import Aesop import Mathlib.Topology.Basic set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat open Set
https://math.stackexchange.com/questions/1577525/is-a-single-point-topologically-connected
PurdueCHE
https://math.stackexchange.com/users/251332/purdueche
243,190
[ { "role": "user", "content": "Complete the following Lean 4 code.\n\n```lean4\nimport Mathlib\nimport Aesop\nimport Mathlib.Topology.Basic\nset_option maxHeartbeats 0\nopen BigOperators Real Nat Topology Rat\nopen Set\n/-- Is a single point topologically connected? -/\ndef MyIsPreconnected {α : Type*} [Topo...
4cdc2c96-df05-5829-9655-cf19f485cf48
[]
[]
cc-by-sa-4.0
Is a single point topologically connected?
mathstack
def MyIsPreconnected {α : Type*} [TopologicalSpace α] (s : Set α) : Prop := ∀ (U V : Set α), IsOpen U → IsOpen V → s ⊆ U ∪ V → (U ∩ s).Nonempty → (V ∩ s).Nonempty → (U ∩ V ∩ s).Nonempty def MyIsConnected {α : Type*} [TopologicalSpace α] (s : Set α) : Prop := s.Nonempty ∧ MyIsPreconnected s theorem probl...
import Mathlib import Aesop import Mathlib.Topology.Basic set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat open Set
https://math.stackexchange.com/questions/1577525/is-a-single-point-topologically-connected
PurdueCHE
https://math.stackexchange.com/users/251332/purdueche
296,508
[ { "role": "user", "content": "Complete the following Lean 4 code.\n\n```lean4\nimport Mathlib\nimport Aesop\nimport Mathlib.Topology.Basic\nset_option maxHeartbeats 0\nopen BigOperators Real Nat Topology Rat\nopen Set\n/-- Is a single point topologically connected? -/\ndef MyIsPreconnected {α : Type*} [Topo...
349bf5d0-06dd-590e-9440-aed0ea746a4b
[]
[]
cc-by-sa-4.0
Is a single point topologically connected?
mathstack
def MyIsPreconnected {α : Type*} [TopologicalSpace α] (s : Set α) : Prop := ∀ (U V : Set α), IsOpen U → IsOpen V → s ⊆ U ∪ V → (U ∩ s).Nonempty → (V ∩ s).Nonempty → (U ∩ V ∩ s).Nonempty def MyIsConnected {α : Type*} [TopologicalSpace α] (s : Set α) : Prop := s.Nonempty ∧ MyIsPreconnected s theorem probl...
import Mathlib import Aesop import Mathlib.Topology.Basic set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat open Set
https://math.stackexchange.com/questions/1577525/is-a-single-point-topologically-connected
PurdueCHE
https://math.stackexchange.com/users/251332/purdueche
527,073
[ { "role": "user", "content": "Complete the following Lean 4 code.\n\n```lean4\nimport Mathlib\nimport Aesop\nimport Mathlib.Topology.Basic\nset_option maxHeartbeats 0\nopen BigOperators Real Nat Topology Rat\nopen Set\n/-- Is a single point topologically connected? -/\ndef MyIsPreconnected {α : Type*} [Topo...
dee3807b-43d5-5e9b-9885-37b3abbf0935
[]
[]
cc-by-sa-4.0
Show that the fundamental group of the space $\mathbb{R}^n$, for $n \geq 1$, is trivial.
mathstack
def MyFundamentalGroup (X : Type*) [TopologicalSpace X] (x₀ : X) : Type* := PUnit theorem problem_237994 {n : ℕ} (hn : 0 < n) : Subsingleton (MyFundamentalGroup (Fin n → ℝ) (fun _ => (0 : ℝ))) := by sorry
import Mathlib import Aesop import Mathlib.Data.Real.Basic import Mathlib.Topology.Instances.Real set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat open Topology
https://math.stackexchange.com/questions/348968/algebraic-topology
aleksandar buskovic
https://math.stackexchange.com/users/70573/aleksandar-buskovic
null
[]
947b5e6c-7f73-505f-a23c-aea28c8833a4
[]
[]
cc-by-sa-4.0
Show that if \( A \) is a symmetrical matrix, then \( B^T A B \) is symmetrical for every matrix \( B \) for which the product is defined.
mathstack
theorem problem_238214 {n m : Type*} [Fintype n] [Fintype m] [DecidableEq n] [DecidableEq m] {α : Type*} [Semiring α] (A : Matrix n n α) (hA : Matrix.transpose A = A) (B : Matrix n m α) : Matrix.transpose ((Matrix.transpose B) * A * B) = (Matrix.transpose B) * A * B := by sorry
import Mathlib import Aesop import Mathlib.Data.Matrix.Basic import Mathlib.Data.Matrix.Notation set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat open Matrix open scoped Matrix
https://math.stackexchange.com/questions/3615548/if-a-is-symmetrical-then-b-tab-is-symmetrical-for-every-matrix-b-for-which-the-p
Carl
https://math.stackexchange.com/users/769882/carl
null
[]
551201ec-eef4-5bef-a1cd-3173b435414e
[]
[]
cc-by-sa-4.0
Show that the series $\sum_{n=1}^\infty \frac{4^n}{n! + 3^n}$ converges.
mathstack
theorem problem_208384 : Summable (fun n : ℕ => (4 : ℝ) ^ n / ((Nat.factorial n : ℝ) + (3 : ℝ) ^ n)) := by sorry
import Mathlib import Aesop import Mathlib set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat open Real
https://math.stackexchange.com/questions/1175554/beginner-real-analysis-help
Bipod
https://math.stackexchange.com/users/189872/bipod
214,585
[ { "role": "user", "content": "Complete the following Lean 4 code.\n\n```lean4\nimport Mathlib\nimport Aesop\nimport Mathlib\nset_option maxHeartbeats 0\nopen BigOperators Real Nat Topology Rat\nopen Real\n/-- Show that the series $\\sum_{n=1}^\\infty \\frac{4^n}{n! + 3^n}$ converges. -/\ntheorem problem_208...
8f483628-8f92-4d67-98d8-04a41e4346ca
[ "nano_v3" ]
[]
cc-by-sa-4.0
Show that the series $\sum_{n=1}^\infty \frac{4^n}{n! + 3^n}$ converges.
mathstack
theorem problem_208384 : Summable (fun n : ℕ => (4 : ℝ) ^ n / ((Nat.factorial n : ℝ) + (3 : ℝ) ^ n)) := by sorry
import Mathlib import Aesop import Mathlib set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat open Real
https://math.stackexchange.com/questions/1175554/beginner-real-analysis-help
Bipod
https://math.stackexchange.com/users/189872/bipod
798,688
[ { "role": "user", "content": "Complete the following Lean 4 code.\n\n```lean4\nimport Mathlib\nimport Aesop\nimport Mathlib\nset_option maxHeartbeats 0\nopen BigOperators Real Nat Topology Rat\nopen Real\n/-- Show that the series $\\sum_{n=1}^\\infty \\frac{4^n}{n! + 3^n}$ converges. -/\ntheorem problem_208...
f5e235d8-4d9b-501f-96c0-201db7614ac8
[]
[]
cc-by-sa-4.0
Show that the series $\sum_{n=1}^\infty \frac{4^n}{n! + 3^n}$ converges.
mathstack
theorem problem_208384 : Summable (fun n : ℕ => (4 : ℝ) ^ n / ((Nat.factorial n : ℝ) + (3 : ℝ) ^ n)) := by sorry
import Mathlib import Aesop import Mathlib set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat open Real
https://math.stackexchange.com/questions/1175554/beginner-real-analysis-help
Bipod
https://math.stackexchange.com/users/189872/bipod
823,670
[ { "role": "user", "content": "Complete the following Lean 4 code.\n\n```lean4\nimport Mathlib\nimport Aesop\nimport Mathlib\nset_option maxHeartbeats 0\nopen BigOperators Real Nat Topology Rat\nopen Real\n/-- Show that the series $\\sum_{n=1}^\\infty \\frac{4^n}{n! + 3^n}$ converges. -/\ntheorem problem_208...
657aa92d-b353-5872-8954-bf70b0b7bb64
[]
[]
cc-by-sa-4.0
Show that the series $\sum_{n=1}^\infty \frac{4^n}{n! + 3^n}$ converges.
mathstack
theorem problem_208384 : Summable (fun n : ℕ => (4 : ℝ) ^ n / ((Nat.factorial n : ℝ) + (3 : ℝ) ^ n)) := by sorry
import Mathlib import Aesop import Mathlib set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat open Real
https://math.stackexchange.com/questions/1175554/beginner-real-analysis-help
Bipod
https://math.stackexchange.com/users/189872/bipod
856,846
[ { "role": "user", "content": "Complete the following Lean 4 code.\n\n```lean4\nimport Mathlib\nimport Aesop\nimport Mathlib\nset_option maxHeartbeats 0\nopen BigOperators Real Nat Topology Rat\nopen Real\n/-- Show that the series $\\sum_{n=1}^\\infty \\frac{4^n}{n! + 3^n}$ converges. -/\ntheorem problem_208...
f4fec2f1-310c-5780-9188-f2854af662ad
[]
[]
cc-by-sa-4.0
If \( x^2 + x + 1 \) is composite, does there always exist an integer \( n \neq x \) such that \( (n^2 + n + 1) \) divides \( (x^2 + x + 1) \)?
mathstack
def Composite (n : Nat) : Prop := ∃ a b : Nat, 1 < a ∧ 1 < b ∧ a * b = n theorem problem_365244 (x : Nat) (hcomp : Composite (x ^ 2 + x + 1)) : ∃ n : Nat, n ≠ x ∧ (n ^ 2 + n + 1) ∣ (x ^ 2 + x + 1) := by sorry
import Mathlib import Aesop set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat
https://math.stackexchange.com/questions/4700472/if-x-2-x-1-is-composite-does-always-exist-a-n-such-that-n-2-n-1-x-2-x-1
rgvalenciaalbornoz
https://math.stackexchange.com/users/949242/rgvalenciaalbornoz
77,996
[ { "role": "user", "content": "Complete the following Lean 4 code.\n\n```lean4\nimport Mathlib\nimport Aesop\nset_option maxHeartbeats 0\nopen BigOperators Real Nat Topology Rat\n/-- If \\( x^2 + x + 1 \\) is composite, does there always exist an integer \\( n \\neq x \\) such that \\( (n^2 + n + 1) \\) divi...
df358044-45e9-4c7e-bb7a-d2be7e95f547
[ "nano_v3" ]
[]
cc-by-sa-4.0
If \( x^2 + x + 1 \) is composite, does there always exist an integer \( n \neq x \) such that \( (n^2 + n + 1) \) divides \( (x^2 + x + 1) \)?
mathstack
def Composite (n : Nat) : Prop := ∃ a b : Nat, 1 < a ∧ 1 < b ∧ a * b = n theorem problem_365244 (x : Nat) (hcomp : Composite (x ^ 2 + x + 1)) : ∃ n : Nat, n ≠ x ∧ (n ^ 2 + n + 1) ∣ (x ^ 2 + x + 1) := by sorry
import Mathlib import Aesop set_option maxHeartbeats 0 open BigOperators Real Nat Topology Rat
https://math.stackexchange.com/questions/4700472/if-x-2-x-1-is-composite-does-always-exist-a-n-such-that-n-2-n-1-x-2-x-1
rgvalenciaalbornoz
https://math.stackexchange.com/users/949242/rgvalenciaalbornoz
156,384
[ { "role": "user", "content": "Complete the following Lean 4 code.\n\n```lean4\nimport Mathlib\nimport Aesop\nset_option maxHeartbeats 0\nopen BigOperators Real Nat Topology Rat\n/-- If \\( x^2 + x + 1 \\) is composite, does there always exist an integer \\( n \\neq x \\) such that \\( (n^2 + n + 1) \\) divi...
d7be9df4-9511-4f56-82e3-15304e8c9771
[ "nano_v3" ]
[]
cc-by-sa-4.0
End of preview. Expand in Data Studio

nvidia/Nemotron-Math-Proofs-v1(複製)

データの内容

数学の証明問題(自然言語)と、その Lean 4 の定理文、推論と証明の軌跡からなる数学推論のデータ。1 行が 1 件の問題と、その推論・解答の会話で、問題文は problem、形式化された定理文は formal_statement に入る。問題は AoPS、Math StackExchange、MathOverflow から集め、gpt-oss-120b で Lean 4 に形式化し、Goedel-Prover-V2-32B が証明を試みて Lean 4 のコンパイラで検証した。ライセンスは CC BY-SA 4.0。

データの概要

項目 値
行数 約 924,942(1 構成・1 分割の合計)
列数 13
容量 29.53 GB(ファイル 3 本)
ファイルの形式 jsonl 1
言語 en
公開 public

構成と分割

構成(config) 分割(split)ごとの行数 列数
default lean 約 924,942 13

列(構成 default・分割 lean)

列 型 先頭の行の値
problem string Prove that there exists a polynomial $ P \in \ma…
source string aops
formal_statement string theorem problem_568570 : ∃ P : Polynomial ℤ,…
lean_header string import Mathlib import Aesop import Mathlib set_o…
url string null
user_name string null
user_url string null
sft_line_number int64 null
messages list<struct{role: string, content: string, reasoning_content: string}> [0 件]
uuid string b2ee7144-44aa-5d7f-8acc-812fae259c90
used_in list<string> [0 件]
tools list<null> [0 件]
license string cc-by-4.0

行数と列は HF のデータセットビューアの集計(元のデータセットの現在の版で数えたもの)。全データセットの一覧は HYUKI5/dataset-catalog にある。

複製について

nvidia/Nemotron-Math-Proofs-v1 をリビジョン 97229c590831adfe96202f5cd071d444d535bf91 の時点でそのまま複製したもの。中身は変えていない。判断モデル Laya(convaiinnovations/laya)の微調整に使う候補として、調査メモ(Laya 学習用データセット調査)に載せたデータを集めている。

出典

論文

アノテーションの方法

データセットの作成日:

作成日: 2025 年 12 月 3 日
最終更新日: 2025 年 12 月 15 日

自然言語の問題の収集元

問題は AoPS コミュニティ、Math StackExchange、MathOverflow から集めた。証明を問う問題と、定理として定式化できる問題に絞り込んだ。問題の収集元と、自然言語の問題を抽出して整える手順の詳細は 論文 を参照。全問題について意味的な重複を取り除き、広く使われるベンチマークとの重複(汚染)も取り除いた。これにより元のデータセットの約 40% が取り除かれ、~580k 件の自然言語の定理が残った。

微調整に使うときの注意

  • 調査メモの台帳での区分: B(継承条件・gated・モデル由来の条件・上流の規約などが付く)。区分は記載の文言で機械的に付けたもので、法的な判定ではない。学習に入れる前に元のカードとライセンスの原文で確かめる。
  • Laya の入力は 2,048 token まで(ModernBERT-large / mmBERT)。長い本文は先頭から黙って切られるので、切れる行は事前に除くか、要約・切り出しをしてから入れる。
  • 分割は記事・文書単位で行い、同じ本文から作った別の問いを学習側と較正側に分けない。温度と閾値は学習に使っていない分割で別々に当てる。

調査メモの記述

メモ 節 ライセンス(記載) 確認 注記
nvidia_nemotron.md 付録. 取得した nvidia 組織のデータセット全件の簡易一覧 (L293) cc-by-sa-4.0 確認済 確認印なし(各ファイルの凡例により確認済扱い) / 原文注記: × 数学。pass_rate 等は難易度ラベルだが金融と無関係

元のカード(日本語訳)

以下は元のデータセットのカードの本文を日本語に訳したもの(機械翻訳)。訳に疑問があるときは、末尾の原文を確かめる。


Nemotron-Math-Proofs-v1

論文: Nemotron-Math: Efficient Long-Context Distillation of Mathematical Reasoning from Multi-Mode Supervision
コード: https://github.com/NVIDIA/NeMo-Skills
ドキュメント: Nemotron-MathProofs-v1 のドキュメント

データセットの説明:

Nemotron-Math-Proofs-v1 は大規模な数学の推論データセットであり、580k 件の自然言語の証明問題、550k 件の Lean 4 の定理文への形式化、~900k 件のモデル生成の推論の軌跡を含む。その軌跡は Lean 4 の証明に至る。このデータセットは、人手で書かれた問題と、体系的に生成された形式化および解法の軌跡を統合している。

各自然言語の問題は、gpt-oss-120b によって、証明の部分をプレースホルダーとした Lean 4 の定理文に形式化される。推論の軌跡と証明は Goedel-Prover-V2-32B が試みる。証明は Lean 4 のコンパイラで検証され、エラーは返されて自己修正が試みられる。

このデータセットは商用利用が可能である。

データセットの所有者:

NVIDIA Corporation

データセットの作成日:

作成日: 2025 年 12 月 3 日
最終更新日: 2025 年 12 月 15 日

ライセンス/利用規約:

このデータセットは Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0) に従う。

想定される用途:

このデータセットは次の用途を想定している:

  • 構造化された数学の推論、特に Lean 4 の形式的な証明の生成を行う LLM の学習
  • 長い文脈または複数の軌跡を扱う推論システムの構築
  • 推論のモード、誤りのパターン、検証パイプラインに関する研究

データセットの特徴づけ

データセットの構成と生成

自然言語の問題の収集元

問題は AoPS コミュニティ、Math StackExchange、MathOverflow から集めた。証明を問う問題と、定理として定式化できる問題に絞り込んだ。問題の収集元と、自然言語の問題を抽出して整える手順の詳細は 論文 を参照。全問題について意味的な重複を取り除き、広く使われるベンチマークとの重複(汚染)も取り除いた。これにより元のデータセットの約 40% が取り除かれ、~580k 件の自然言語の定理が残った。

自動形式化

自然言語の定理は、gpt-oss-120b を用いた自動形式化パイプラインにかけられ、Lean の形式的な文を得た。自動形式化パイプラインは次の段階から成る:

  • 初期の形式化: LLM に、自然言語の定理の文を Lean 4 の形式コードへ訳すよう指示する。生成されたコードは抽出され、整えられる(コメントの削除、import の整理、sorry のプレースホルダーの処理)。
  • Lean 4 のコンパイル確認: 抽出されたコードは Lean 4 のサンドボックスに送られてコンパイルされる。これにより構文の正しさと型検査が確かめられる。
  • 逆翻訳と意味の検証: コンパイルが通ったら、形式コードを 2 回目の LLM の呼び出しで自然言語に逆翻訳する。次に判定用のプロンプトで、この逆翻訳を元の問題文と比べ、形式化が意図した意味を保っているかを確かめる。
  • 反復的な修正のループ: いずれかの段階が失敗すると、パイプラインはエラーに応じた修正用のプロンプトを組み立てる:
  • 解析エラー: コードの抽出の問題を直すよう指示するプロンプト
  • コンパイルエラー: Lean のエラーメッセージを含むプロンプト
  • 意味の不整合: 判定者の理由を含むプロンプト
  • 次に LLM がコードの修正を試み、この過程を繰り返す(最大 8 回)。
  • 適応的な推論: 文脈長の上限で生成が失敗したとき、推論の努力レベル(high → medium → low)を自動で下げる。これによりパイプラインは、token の予算の枯渇から立ち直れる。
  • 忠実性フィルタリング: 最後に、もう 1 段の忠実性フィルタリングを行う。LLM を判定者として、形式化が元の自然言語の文に忠実かを評価させる。3 回の判定のいずれかで不正確とされた文は取り除いた。

これにより ~550k 件の形式化された定理文が残った。

自動定理証明

定理文を入力として、証明器として Goedel-Prover-V2-32B を用いた大規模な推論を行い、推論の軌跡と、形式的に検査された解を表すコンパイル可能な Lean のコードのブロックを得た。言語モデルは、定理について推論し、Lean の書き方を計画し、コンパイル可能な Lean のコードを生成するほか、コンパイラのフィードバックに応答して誤った試みを自己修正することもできる。証明器のパイプラインは次の段階から成る:

  • 初期の証明の生成: LLM は形式的な文(ヘッダー + 非形式的な前置き + 形式的な文)を受け取り、証明の試みを生成する。コードは抽出され、元の文と結合される。
  • Lean 4 による検証: 証明は Lean 4 のサンドボックスでコンパイルされる。成功と見なすには、コンパイルが完了し、sorry のプレースホルダーがなく、失敗のメッセージもないことが必要である。
  • エラー情報を加えたフィードバック: 失敗すると、エラーは解析され、位置を示すマーカー(<error></error> タグ)を付けて整形される。これによりコードのどこで誤りが起きたかを正確に示す。この構造化されたフィードバックがモデルに返される。
  • 反復的な修正のループ: 成功するか、ターン数の上限に達するまで、エラーのフィードバックを付けて対話を続ける。
  • Pass@N のサンプリング: 修正のループ全体を最初からやり直すことを、最大 N 回まで行える。これにより独立した証明の試みを複数得られる。

証明の生成では、誤ったターンを削除する、複数ターンにわたる誤り訂正・修正の戦略を用いる:

  • 思考連鎖(chain-of-thought)の除去: 生成のたびにモデルの思考連鎖の推論は取り除き、lean4 のブロックの内側にある形式的な証明のコードだけを残す。
  • 誤ったターンの削除: 対話の履歴を積み上げる代わりに、失敗した試みは捨てる。修正の各ターンは、元のプロンプトと直近の証明の試みだけを見る。これにより、以前の失敗が文脈を汚すのを防ぐ。
  • 構造化されたエラーのフィードバック: コンパイラのエラーには、ソースコード中の位置を示す <error></error> タグが付けられ、モデルが問題の箇所を特定するのを助ける。
  • 修正を伴う Pass@N: 各問題は n_pass=4 で試みられ、各試みは最大 8 回の修正のターンを許す。

この方法は、失敗した推論の軌跡を積み上げないことで文脈の効率を保ちつつ、自己修正のために構造化されたエラーのフィードバックを与えることで反復的な改善を図る。

限界

以下に、データセットに残る制約のうち、今後の版で改善を目指すものを挙げる:

  • 難易度の釣り合い: 証明の問題に絞る以外に、データセット内の問題を正規化したり釣り合いを取ったりする明示的な試みはしていない。各段階(自動形式化の成否、証明の試みの成否など)から生じる暗黙の選択の偏りがある。
  • トークン数・長さの正規化: 最終的な検証は、Lean 4 のコンパイラでエラーがないことだけに依っていた。含まれる解の中には、未使用の仮定のような余分に生成された行に関する警告が多数あるものがある。
  • プレースホルダーや無内容な定義: 自動形式化器が mathlib の代わりに、自明なプレースホルダーの定義を使う例を確認した。例: def MyContinuous... Prop := True. これは、証明が基礎となる 3 つの公理にだけ依存するかを確かめる最後の検査で検出できる。

学習の結果

Qwen3-8B を SFT したモデルを、Lean の定理証明のために特別に学習させた Goedel-Prover-V2-8B と比較する:

モデル pass@32(自己修正なし) pass@32(自己修正あり)
Goedel-Prover-V2-8B 84.6% 86.7%
Nemotron-MathProofs-v1 で SFT した Qwen3-8B 85.3% 90.2%

この性能は、Lean の定理証明のために特別に学習させたモデル(DeepSeek-Prover-V2-7B、Kimina-Prover-8B、Goedel-Prover-V2-8B など)と同程度である。

このデータセットは Nemotron-Nano-3 の SFT データの一部として使われた。最終的なモデルは miniF2F で次の結果を得た:

モデル pass@32(自己修正なし) pass@32(自己修正あり)
Nemotron-Nano-3 79.92% 86.89%
gpt-oss-20b 43.03% 59.42%
Qwen3-30B-A3B-Thinking 16.80% 32.37%

データセットのフィールド

フィールド 型 説明
uuid str 一意の識別子
problem str 自然言語の問題文(常に存在する)
source str 問題の収集元: "aops" または "mathstack"(常に存在する)
formal_statement str | null Lean 4 の定理のコード(約 4% のエントリでは null)
lean_header str | null Lean の import 文と設定
messages list 検証済みの証明の試みを対話として並べたもの。各エントリは正しい Lean 4 のコードを生む(なければ空のリスト)
url str | null 元の投稿の URL(mathstack のみ)
user_name str | null 元の投稿者のユーザー名(mathstack のみ)
user_url str | null 元の投稿者のプロフィール URL(mathstack のみ)
used_in list 学習時にこのサンプルを含めた Nemotron のモデルを示す参照
tools list ツールの定義

再現

このデータセットは NeMo-Skills を用いて作られた。Nemotron-Nano-3 の評価の数値を再現するための詳しい手順、プロンプト、コマンドは、Nemotron-MathProofs-v1 のドキュメント を参照。

データの収集方法
ハイブリッド: 自動、合成

データセットの形式

モダリティ: テキスト
形式: JSONL
構造: テキスト + メタデータ

データセットの量

サブセット サンプル数
train 1,376,666

ディスク上の合計サイズ: ~28 GB

参考文献:

  • Agarwal, Sandhini, et al. gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925 (2025).

  • Du, Wei, et al. Nemotron-Math: Efficient Long-Context Distillation of Mathematical Reasoning from Multi-Mode Supervision. arXiv preprint arXiv:2512.15489 (2025)

  • Lin, Yong, et al. Goedel-prover-v2: Scaling formal theorem proving with scaffolded data synthesis and self-correction. arXiv preprint arXiv:2508.03613 (2025).

倫理的な配慮:

NVIDIA は、信頼できる AI は共有の責任であると考えており、幅広い AI の応用の開発を可能にするための方針と実践を定めている。利用規約に従ってダウンロードまたは使用される場合、開発者は、関連する業界と用途の要件をこのモデルが満たし、想定外の製品の誤用に対処できるよう、社内のモデルのチームと協力すべきである。
品質、リスク、セキュリティの脆弱性、または NVIDIA の AI に関する懸念は こちら から報告してほしい。

原文(英語)

Nemotron-Math-Proofs-v1

Paper: Nemotron-Math: Efficient Long-Context Distillation of Mathematical Reasoning from Multi-Mode Supervision
Code: https://github.com/NVIDIA/NeMo-Skills
Documentation: Nemotron-MathProofs-v1 documentation

Dataset Description:

Nemotron-Math-Proofs-v1 is a large-scale mathematical reasoning dataset containing ~580k natural language proof problems, ~550k formalizations into theorem statements in Lean 4, and ~900k model-generated reasoning trajectories culminating in Lean 4 proofs. The dataset integrates human-authored problems with systematically generated formalizations and solution traces.

Each natural language problem is formalized by gpt-oss-120b into Lean 4 theorem statements with placeholders for the proof. Reasoning traces and proofs are attempted by Goedel-Prover-V2-32B. Proofs are verified by the Lean 4 compiler, with errors passed back and self-correction attempted.

This dataset is ready for commercial use.

Dataset Owner(s):

NVIDIA Corporation

Dataset Creation Date:

Created on: Dec 3, 2025
Last Modified on: Dec 15, 2025

License/Terms of Use:

This dataset is governed by the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0).

Intended Usage:

This dataset is intended for:

  • Training LLMs to perform structured mathematical reasoning, in particular generation of Lean 4 formal proofs.

  • Building long-context or multi-trajectory reasoning systems

  • Research on reasoning modes, error patterns, and verification pipelines

Dataset Characterization

Dataset Composition and Generation

Natural language problem sourcing

We collected problems from the AoPS community, Math StackExchange, and MathOverflow, filtering for proof based problems and those that can be formulated as theorems. See our paper for more details on the problem sources and pipeline to extract and prepare natural language problems. We semantically deduplicated all problems and decontaminated them with popular benchmarks. This removed ~40% of the original dataset and resulted in ~580k natural language theorems.

Autoformalization

The natural language theorems were put into an autoformalization pipeline powered by gpt-oss-120b to obtain formalized statements in Lean. The autoformalization pipeline consists of the following stages:

  • Initial Formalization: The LLM is prompted to translate a natural language theorem statement into Lean 4 formal code. The generated code is extracted and cleaned (removing comments, organizing imports, handling sorry placeholders).
  • Lean 4 Compilation Check: The extracted code is sent to a Lean 4 sandbox for compilation. This verifies syntactic correctness and type-checking.
  • Backtranslation & Semantic Verification: If compilation succeeds, the formal code is backtranslated to natural language by a second LLM call. A judge prompt then compares this backtranslation to the original problem statement to verify the formalization preserves the intended meaning.
  • Iterative Refinement Loop: When any step fails, the pipeline constructs an error-specific refinement prompt:
  • Parsing errors: prompt to fix code extraction issues
  • Compilation errors: prompt including the Lean error message
  • Semantic inconsistency: prompt including the judge's reasoning
  • The LLM then attempts to fix the code, and the process repeats (up to 8 iterations).
  • Adaptive Reasoning: Automatically reduces the reasoning effort level (high → medium → low) when generation fails due to context length limits, allowing the pipeline to recover from token budget exhaustion.
  • Faithfulness Filtering: An additional round of final Faithfulness Filtering is applied, where an LLM-as-judge is prompted to assess the faithfulness of the formalization to the original natural language statement. A statement judged innacurate in any of three passes was removed.

This resulted in ~550k formalized theorem statements.

Automated theorem proving

Given the theorem statements, we ran large scale inference using Goedel-Prover-V2-32B as a prover, in order to obtain reasoning traces and compilable blocks of Lean code representing formally checked solutions. Along with reasoning through the theorems, planning Lean writing, and generating compilable Lean code, the language model is also able to respond to compiler feedback and self-correct erroneous attempts. The prover pipeline consists of the following stages:

  • Initial Proof Generation: The LLM receives a formal statement (header + informal prefix + formal statement) and generates a proof attempt. The code is extracted and merged with the original statement.
  • Lean 4 Verification: The proof is compiled in the Lean 4 sandbox. Success requires: completed compilation, no sorry placeholders, and no failure messages.
  • Error-Enriched Feedback: On failure, errors are parsed and formatted with positional markers (<error></error> tags) to indicate exactly where in the code the error occurred. This structured feedback is passed back to the model.
  • Iterative Refinement Loop: The conversation continues with error feedback until success or the maximum number of turns is reached.
  • Pass@N Sampling: The entire refinement loop can be restarted from scratch up to N times, providing multiple independent proving attempts.

During proof generation, we employ a multi-turn error correction/refinement strategy with wrong turn deletion:

  • Chain-of-thought removal: After each generation, the model's chain-of-thought reasoning is stripped, keeping only the formal proof code within lean4 blocks
  • Wrong turn deletion: Rather than accumulating conversation history, failed attempts are discarded. Each refinement turn sees only the base prompt plus the most recent proof attempt, preventing context pollution from earlier failures
  • Structured error feedback: Compiler errors are annotated with <error></error> tags marking their location in the source code, helping the model localize issues
  • Pass@N with refinement: Each problem is attempted with n_pass=4 where each attempt allows up to 8 refinement turns

This approach balances context efficiency (by not accumulating failed reasoning traces) with iterative improvement (by providing structured error feedback for self-correction).

Limitations

Below we flag some limitations remaining in the dataset that we aim to improve in future iterations:

  • Difficulty balance: beyond filtering for proof problems there was no explicit attempt to normalize or balance the problems present in the dataset. There are implicit selection biases resulting from various stages (autoformalization success, proof attempt success, etc.).
  • Token count/length normalization: final verification was only dependent on Lean 4 compiler containing no errors. Some included solutions contain numerous warnings pertaining to extraneous generated lines such as unused hypotheses.
  • Placeholder/vacuous definition: we observed instances of the autoformalizer using placeholder trivial definitions instead of mathlib, e.g. def MyContinuous... Prop := True. This can be detected with a final check that the proof only depends on the three foundational axioms.

Training Results

We compare a Qwen3-8B SFT against Goedel-Prover-V2-8B, a model specifically trained for Lean theorem proving:

Model pass@32 (no self-correction) pass@32 (with self-correction)
Goedel-Prover-V2-8B 84.6% 86.7%
Qwen3-8B SFT-ed on Nemotron-MathProofs-v1 85.3% 90.2%

This performance is comparable to models specifically trained for Lean theorem proving, such as DeepSeek-Prover-V2-7B, Kimina-Prover-8B, and Goedel-Prover-V2-8B.

This dataset was used as part of the SFT data for Nemotron-Nano-3. The final model achieves the following results on miniF2F:

Model pass@32 (no self-correction) pass@32 (with self-correction)
Nemotron-Nano-3 79.92% 86.89%
gpt-oss-20b 43.03% 59.42%
Qwen3-30B-A3B-Thinking 16.80% 32.37%

Dataset Fields

Field Type Description
uuid str Unique identifier
problem str Natural language problem statement (always present)
source str Problem source: "aops" or "mathstack" (always present)
formal_statement str | null Lean 4 theorem code (null for ~4% of entries)
lean_header str | null Lean import statements and setup
messages list Verified proof attempts as chat conversations; each entry produces correct Lean 4 code (empty list if none)
url str | null Original post URL (mathstack only)
user_name str | null Original poster's username (mathstack only)
user_url str | null Original poster's profile URL (mathstack only)
used_in list Reference for which Nemotron models included this sample during training
tools list Tool definition

Reproduction

This dataset was created using NeMo-Skills. For detailed instructions, prompts, and commands to reproduce evaluation numbers for Nemotron-Nano-3, see the Nemotron-MathProofs-v1 documentation.

Data Collection Method
Hybrid: Automated, Synthetic

Dataset Format

Modality: Text
Format: JSONL
Structure: Text + Metadata

Dataset Quantification

Subset Samples
train 1,376,666

Total Disk Size: ~28 GB

Reference(s):

  • Agarwal, Sandhini, et al. gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925 (2025).

  • Du, Wei, et al. Nemotron-Math: Efficient Long-Context Distillation of Mathematical Reasoning from Multi-Mode Supervision. arXiv preprint arXiv:2512.15489 (2025)

  • Lin, Yong, et al. Goedel-prover-v2: Scaling formal theorem proving with scaffolded data synthesis and self-correction. arXiv preprint arXiv:2508.03613 (2025).

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here

Downloads last month
38

Papers for HYUKI5/nvidia__Nemotron-Math-Proofs-v1