tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
We create miniCodeProps by translating programs from Tons of Inductive Programs TIP 14
httpsgithubcomtiporgbenchmarks from Haskell into Lean 4 We manually translated code from
three files in TIP that contain a mix of function definitions propositions describing the results of
calling those functions and termination lemmas used in the function definitions In total this yields
177 Lean 4 theorem statements
31 Collection Process
The properties in TIP describe induction in both programming and purely mathematical contexts We
first filtered out files that contained primarily mathematical induction then translated the remaining
Haskell code into Lean 4 We then associated each property in the translated files with dependencies
and metadata useful for different evaluation modes of the benchmark
3Numbers Lists Trees Heaps
Medley Functions 0 15 2 0
Medley Properties 21 64 1 0
TerminationSorting Functions 3 49 2 5
Termination Properties 3 22 0 3
Sorting Properties 0 63 0 0
Table 1 miniCodeProps Composition We classified each function and property in the Lean source
files by their primary subject matter We also separate these numbers by the splits of miniCodeProps
Medley Termination lemmas and Sorting algorithms Termination and Sorting algorithm functions
are grouped because the properties in both splits describe the same set of functions
32 Source Code Composition
321 Functions
Most of the translated functions operate on linked lists while the rest involve natural numbers and
binary trees With a few notable exceptions the functions perform conceptually simple operations
such as filtering returning the last element and counting the elements of a list The more complicated
functions are increasingly esoteric sorting functions In total miniCodeProps is derived from 76
TIP functions A breakdown of the primary subject of each function is provided in Table 1
322 Properties
The properties to be proven in miniCodeProps express intuitively correct properties of the functions
being described See Section 34 for examples and Table 1 for a breakdown by subject type
323 Termination Lemmas
In Lean 4 recursive functions must be paired with a proof of termination While Lean automatically
infers the proof in simple cases more complicated recursive calls require the user to explicitly prove
termination As the vast majority of our function definitions are recursive our benchmark includes 28
lemmas that support 4 termination proofs of nonstandard sorting algorithms from TIP These lemmas
themselves represent practically useful and highly nontrivial properties of code
33 Dataset Splits
To improve understanding of the precise capabilities of theorem proving agents when evaluated on
miniCodeProps we section the 177 properties into 3 data splits Medley Termination lemmas and
Sorting algorithms Medley properties are named prop_n for n186 and tend to involve simple
functions on numbers and lists Sorting algorithm properties are those remaining theorems named
beginning with prop and the remainder are Termination lemma properties
The benchmark is formatted as a jsonlines file with the information in Table 2 per line The full_name
property is used to determine the split the property belongs to
34 Qualitative Examples
We present several examples from the benchmark and describe what makes them interesting and
difficult
341 Zip over Concatenation
We begin with a reimplementation of the standard zip function that combines two lists into a single
list of paired elements prop_84 states that zipping a list xswith the concatenation of two other
listsys zs is equivalent to separately zipping appropriate sections of xswith each of ysandzs
then concatenating the results To the best of our knowledge Mathlib does not contain an equivalent
lemma for the standard definition of the zip function
4Table 2 Example code property and associated data
Key Description Examples
full_name Unique name of the property
in the benchmarkprop_60
prop_defn The text of the code property
to be provenlemma prop_60 xs List Nat ys
List Nat not null ys last
xs ys last ys by sorry
prop_loc The file name and line num
ber where the property defi
nition beginsLeanSrcLeanSrcPropertieslean257
score An integer in the range 15
used as a difficulty heuristic5
deps Text containing all the depen
dencies functions and lem
mas that the property re
quires to be fully definedimport Mathlib
def last List Nat Nat
0
x x
_xxs last xs
def null List αBool
True
_ False
proof_state The initial proof state of the
propertyxs ys List N
null ys true last xs ys
last ys
file_locs A list of the last line in each
file needed to define the prop
erty Appending each file up
to this line before the prop
erty fully defines the prop
ertyLeanSrcLeanSrcDefinitionslean 249
LeanSrcLeanSrcPropertieslean 258
def zip List αList βList αβ
_
_
xxs yys x y zip xs ys
lemma prop_84 xs List α ys List β zs List β
zip xs ys zs
zip Listtake Listlength ys xs ys
zip Listdrop Listlength ys xs zs by
This property is intuitively correct but requires some intuition about proofs of recursive structures
to correctly prove in Lean The following is an incorrect proof generated by GPT4 It contains one
semantic error failing to generalize xsin the inductive hypothesis and one misuse of Lean internals
the cases keyword should be used instead of match

We create miniCodeProps by translating programs from Tons of Inductive Programs TIP 14

httpsgithubcomtiporgbenchmarks from Haskell into Lean 4 We manually translated code from

three files in TIP that contain a mix of function definitions propositions describing the results of

calling those functions and termination lemmas used in the function definitions In total this yields

177 Lean 4 theorem statements

31 Collection Process

The properties in TIP describe induction in both programming and purely mathematical contexts We

first filtered out files that contained primarily mathematical induction then translated the remaining

Haskell code into Lean 4 We then associated each property in the translated files with dependencies

and metadata useful for different evaluation modes of the benchmark

3Numbers Lists Trees Heaps

Medley Functions 0 15 2 0

Medley Properties 21 64 1 0

TerminationSorting Functions 3 49 2 5

Termination Properties 3 22 0 3

Sorting Properties 0 63 0 0

Table 1 miniCodeProps Composition We classified each function and property in the Lean source

files by their primary subject matter We also separate these numbers by the splits of miniCodeProps

Medley Termination lemmas and Sorting algorithms Termination and Sorting algorithm functions

are grouped because the properties in both splits describe the same set of functions

32 Source Code Composition

321 Functions

Most of the translated functions operate on linked lists while the rest involve natural numbers and

binary trees With a few notable exceptions the functions perform conceptually simple operations

such as filtering returning the last element and counting the elements of a list The more complicated

functions are increasingly esoteric sorting functions In total miniCodeProps is derived from 76

TIP functions A breakdown of the primary subject of each function is provided in Table 1

322 Properties

The properties to be proven in miniCodeProps express intuitively correct properties of the functions

being described See Section 34 for examples and Table 1 for a breakdown by subject type

323 Termination Lemmas

In Lean 4 recursive functions must be paired with a proof of termination While Lean automatically

infers the proof in simple cases more complicated recursive calls require the user to explicitly prove

termination As the vast majority of our function definitions are recursive our benchmark includes 28

lemmas that support 4 termination proofs of nonstandard sorting algorithms from TIP These lemmas

themselves represent practically useful and highly nontrivial properties of code

33 Dataset Splits

To improve understanding of the precise capabilities of theorem proving agents when evaluated on

miniCodeProps we section the 177 properties into 3 data splits Medley Termination lemmas and

Sorting algorithms Medley properties are named prop_n for n186 and tend to involve simple

functions on numbers and lists Sorting algorithm properties are those remaining theorems named

beginning with prop and the remainder are Termination lemma properties

The benchmark is formatted as a jsonlines file with the information in Table 2 per line The full_name

property is used to determine the split the property belongs to

34 Qualitative Examples

We present several examples from the benchmark and describe what makes them interesting and

difficult

341 Zip over Concatenation

We begin with a reimplementation of the standard zip function that combines two lists into a single

list of paired elements prop_84 states that zipping a list xswith the concatenation of two other

listsys zs is equivalent to separately zipping appropriate sections of xswith each of ysandzs

then concatenating the results To the best of our knowledge Mathlib does not contain an equivalent

lemma for the standard definition of the zip function

4Table 2 Example code property and associated data

Key Description Examples

full_name Unique name of the property

in the benchmarkprop_60

prop_defn The text of the code property

to be provenlemma prop_60 xs List Nat ys

List Nat not null ys last

xs ys last ys by sorry

prop_loc The file name and line num

ber where the property defi

nition beginsLeanSrcLeanSrcPropertieslean257

score An integer in the range 15

used as a difficulty heuristic5

deps Text containing all the depen

dencies functions and lem

mas that the property re

quires to be fully definedimport Mathlib

def last List Nat Nat

0

x x

_xxs last xs

def null List αBool

True

_ False

proof_state The initial proof state of the

propertyxs ys List N

null ys true last xs ys

last ys

file_locs A list of the last line in each

file needed to define the prop

erty Appending each file up

to this line before the prop

erty fully defines the prop

ertyLeanSrcLeanSrcDefinitionslean 249

LeanSrcLeanSrcPropertieslean 258

def zip List αList βList αβ

_

xxs yys x y zip xs ys

lemma prop_84 xs List α ys List β zs List β

zip xs ys zs

zip Listtake Listlength ys xs ys

zip Listdrop Listlength ys xs zs by

This property is intuitively correct but requires some intuition about proofs of recursive structures

to correctly prove in Lean The following is an incorrect proof generated by GPT4 It contains one

semantic error failing to generalize xsin the inductive hypothesis and one misuse of Lean internals

the cases keyword should be used instead of match