---
language:
  - en
task_categories:
  - text-generation
size_categories:
  - n<1K
license: mit
pretty_name: FlutterBench
---

# FlutterBench

FlutterBench is a 600-item, English-language, held-out evaluation benchmark
for Flutter/Dart code generation, disjoint from the companion `flutterft`
training corpus. This repository contains only the **public** portion of the
benchmark: prompts and scoring rubrics. Reference (private) answers are
intentionally **not** included here, to preserve the benchmark as a genuine
held-out evaluation set for anyone who wants to test a model against it
without seeing the expected output in advance.

## Important notice on source provenance and licensing

**Source repository, commit, and license information for the underlying code
referenced by these prompts was not recorded during data collection and has
not been verified per record.** Public visibility of a source repository does
not by itself grant permission to redistribute derived material under a new
license. This dataset's own packaging (this card, the JSONL structure) is
released under the MIT license, but that does not and cannot override the
license terms of any original source code the prompts may reference, which
are unknown here. **Use this dataset at your own risk and responsibility.**
If you are the author of any code referenced in this benchmark and object to
its inclusion, please open a discussion on this repository and it will be
addressed.

## Contents

- `prompts.jsonl` — 600 public prompts with task type, difficulty, and
  metadata (chat-format, no assistant answer).
- `rubrics.jsonl` — 600 five-criterion, task-dependent scoring rubrics
  (0-5 anchors) for evaluating generated responses.

| Task type | Items |
| --- | ---: |
| question_to_code | 282 |
| test_generation | 200 |
| refactor | 26 |
| code_to_explanation | 26 |
| bug_fix | 25 |
| completion | 25 |
| api_usage | 16 |

Difficulty distribution: 42 easy, 313 medium, 245 hard.

## Evaluation policy

Text-similarity metrics (e.g. n-gram overlap against a private reference) are
a diagnostic signal only, not proof of functional correctness. A subset of
120 items has an associated executable check (FlutterBench-Exec: `dart
format`, `dart analyze`, `flutter test` in an isolated package); results
should be reported as pass@1 on the specific subset actually exercised, not
extrapolated to the full 600 items. Because rubrics reference exact expected
outputs, treat any score derived without executing the code as a diagnostic,
not a correctness claim.

## Known limitations

- Reference/private answers are not published in this repository by design.
  Researchers who need them for controlled comparison should contact the
  maintainer directly.
- Only 61 of a 120-item executable-check subset have been authored and
  locally validated as of this release; the rest depend on upstream source
  not available in this dataset and are not scored.
- Only one downstream model family (Qwen2.5-Coder-3B-Instruct) has been
  evaluated against this benchmark; results should not be assumed to
  generalize to other models.

## Citation

If you use this benchmark, please cite the repository. A predecessor study on
Turkish Flutter code generation is:

Uluırmak, B. A., & Kurban, R. (2025). Fine-tuning Large Language Models for
Turkish Flutter Code Generation. *Sakarya University Journal of Computer and
Information Sciences*, 8(4), 637-650.