Towards Functional Correctness of Large Code Models with Selective Generation
Paper • 2505.13553 • Published
Datasets:
tests listlengths 0 600 | status stringclasses 7
values | n_execs int64 0 60k | arcs int64 0 41 | n_seeds int64 0 26 | task_id stringlengths 6 13 | dataset stringclasses 2
values | entry_point stringlengths 1 31 | secs float64 0.04 647 |
|---|---|---|---|---|---|---|---|---|
[
{
"input": "('',)",
"output": "0",
"seed": true
},
{
"input": "('abcde',)",
"output": "5",
"seed": true
},
{
"input": "('aaaaAAAAaaaa',)",
"output": "1",
"seed": true
},
{
"input": "('Jerry jERRY JeRRRY',)",
"output": "5",
"seed": true
},
{
"input"... | ok | 748 | 1 | 4 | HumanEval/16 | humaneval | count_distinct_characters | 0.05 |
[
{
"input": "(3, 7)",
"output": "1",
"seed": true
},
{
"input": "(10, 15)",
"output": "5",
"seed": true
},
{
"input": "(49, 14)",
"output": "7",
"seed": true
},
{
"input": "(144, 60)",
"output": "12",
"seed": true
},
{
"input": "(146, 60)",
"out... | ok | 982 | 4 | 4 | HumanEval/13 | humaneval | greatest_common_divisor | 0.05 |
[
{
"input": "(3.5,)",
"output": "0.5",
"seed": true
},
{
"input": "(1.33,)",
"output": "0.33000000000000007",
"seed": true
},
{
"input": "(123.456,)",
"output": "0.45600000000000307",
"seed": true
},
{
"input": "(2.0070718025263945,)",
"output": "0.007071802526... | ok | 896 | 1 | 3 | HumanEval/2 | humaneval | truncate_number | 0.06 |
[
{
"input": "('',)",
"output": "[]",
"seed": true
},
{
"input": "('asdfgh',)",
"output": "['a', 'as', 'asd', 'asdf', 'asdfg', 'asdfgh']",
"seed": true
},
{
"input": "('WWW',)",
"output": "['W', 'WW', 'WWW']",
"seed": true
},
{
"input": "('asdfdh',)",
"output": ... | ok | 878 | 5 | 3 | HumanEval/14 | humaneval | all_prefixes | 0.09 |
[
{
"input": "('', 'x')",
"output": "0",
"seed": true
},
{
"input": "('xyxyxyx', 'x')",
"output": "4",
"seed": true
},
{
"input": "('cacacacac', 'cac')",
"output": "4",
"seed": true
},
{
"input": "('john doe', 'john')",
"output": "1",
"seed": true
},
{
... | ok | 868 | 7 | 4 | HumanEval/18 | humaneval | how_many_times | 0.09 |
[{"input":"('',)","output":"''","seed":true},{"input":"('x',)","output":"'x'","seed":true},{"input":(...TRUNCATED) | ok | 1,014 | 8 | 5 | HumanEval/10 | humaneval | make_palindrome | 0.1 |
[{"input":"('111000', '101010')","output":"'010010'","seed":true},{"input":"('1', '1')","output":"'0(...TRUNCATED) | ok | 1,037 | 6 | 3 | HumanEval/11 | humaneval | string_xor | 0.11 |
[{"input":"([],)","output":"False","seed":true},{"input":"([1, 2, -3, 1, 2, -3],)","output":"False",(...TRUNCATED) | ok | 802 | 7 | 6 | HumanEval/3 | humaneval | below_zero | 0.11 |
[{"input":"([1.0, 2.0, 3.0],)","output":"0.6666666666666666","seed":true},{"input":"([1.0, 2.0, 3.0,(...TRUNCATED) | ok | 911 | 3 | 3 | HumanEval/4 | humaneval | mean_absolute_deviation | 0.11 |
[{"input":"([], 'john')","output":"[]","seed":true},{"input":"(['xxx', 'asd', 'xxy', 'john doe', 'xx(...TRUNCATED) | ok | 981 | 2 | 4 | HumanEval/7 | humaneval | filter_by_substring | 0.12 |
End of preview. Expand in Data Studio
FuzzEval unit tests for HumanEval-f and MBPP-f
Automatically generated unit tests for a reproduction of the ICML 2026 paper "Towards Functional Correctness of Large Code Models with Selective Generation" (Jeong, Kim & Park — arXiv:2505.13553, official repo trustml-lab/selective-code-generation).
The paper's FuzzEval paradigm replaces a benchmark's handful of hand-written unit tests with hundreds of unit tests obtained by fuzzing the reference solution. This dataset is our reconstruction of that step for openai/openai_humaneval and google-research-datasets/mbpp (sanitized).
How it was built
For each problem we take the canonical solution as the fuzzing target:
- Seed corpus — argument tuples parsed out of the benchmark's own asserts.
- Coverage-guided loop — inputs are mutated with type-aware operators
(integers, floats, strings, lists/tuples/sets, dicts, recursively);
a mutant that reaches a new
(line -> line)arc of the reference is added to the corpus, exactly as a coverage-guided fuzzer such as Atheris would. - Recording — every input that executes without raising, twice with the
same result (determinism check) and whose
reprround-trips, yields a unit test(u, v)withv = reference(u).
Budget per problem: 600 unique tests, 60k executions, 90 s.
Contents
fuzz_tests.jsonl — one row per problem:
| field | meaning |
|---|---|
task_id |
HumanEval/i or MBPP/i |
dataset |
humaneval / mbpp |
entry_point |
function under test |
tests |
list of {input: repr(args tuple), output: repr(value), seed: bool} |
arcs |
distinct control-flow arcs covered in the reference |
n_execs, status, secs |
fuzzing budget diagnostics |
seed: true marks a test whose input came from the benchmark's own asserts —
those are the "manual unit tests" used by the paper's SCG-small baseline.
Stats
- 591 problems attempted, 562 with a full 600-test suite (153 HumanEval, 409 MBPP)
- 337,200 unit tests total
- 26 problems excluded: 3 have no literal seed inputs, 1 has no valid seed execution, and 22 hit the time budget (reference solutions that blow up on mutated inputs — the paper handles these with manual input-constraint post-processing, which we did not do).
Reproduce
python fuzz_tests.py --target 600 --max-execs 60000 --out data/fuzz_tests.jsonl
Code: see the Workspace tab of the reproduction logbook.
- Downloads last month
- -