Datasets:
File size: 2,914 Bytes
6a63a06 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 | ---
license: mit
task_categories:
- text-generation
tags:
- code-generation
- unit-tests
- fuzzing
- icml2026
- reproduction
pretty_name: FuzzEval unit tests for HumanEval-f / MBPP-f
size_categories:
- 100K<n<1M
---
# 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](https://arxiv.org/abs/2505.13553),
official repo [trustml-lab/selective-code-generation](https://github.com/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](https://huggingface.co/datasets/openai/openai_humaneval)
and [google-research-datasets/mbpp](https://huggingface.co/datasets/google-research-datasets/mbpp) (sanitized).
## How it was built
For each problem we take the *canonical solution* as the fuzzing target:
1. **Seed corpus** — argument tuples parsed out of the benchmark's own asserts.
2. **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.
3. **Recording** — every input that executes without raising, twice with the
same result (determinism check) and whose `repr` round-trips, yields a unit
test `(u, v)` with `v = 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
```bash
python fuzz_tests.py --target 600 --max-execs 60000 --out data/fuzz_tests.jsonl
```
Code: see the Workspace tab of the reproduction logbook.
|