README.md

metadata

language:
  - en
task_categories:
  - text-generation
size_categories:
  - 1K<n<10K
license: mit
pretty_name: FlutterFT

FlutterFT

FlutterFT is a 5,000-record English-language instruction-tuning corpus for Flutter/Dart code generation, spanning seven task types: question-to-code, test generation, bug fixing, refactoring, completion, code explanation, and API usage. Each record is a chat-format (system/user/assistant) example.

Important notice on source provenance and licensing

Source repository, commit, and license information for the underlying code snippets was not recorded during data collection and has not been verified per record. The code snippets in this corpus were collected from public GitHub repositories, but public visibility of a repository does not by itself grant permission to redistribute its code under a new license. This dataset's own packaging (this card, the JSONL structure, associated tooling) is released under the MIT license, but that does not and cannot override the license terms of the original source code, which are unknown here.

Use this dataset at your own risk and responsibility. If you are the author of any code that appears in this corpus and object to its inclusion, please open a discussion on this repository and it will be addressed.

Dataset structure

Task type	Records
question_to_code	2,845
code_to_explanation	513
test_generation	513
bug_fix	394
refactor	393
completion	312
api_usage	30

Each line in train.jsonl is a JSON object:

{
  "id": "seed_000001",
  "messages": [
    {"role": "system", "content": "You are a Flutter and Dart coding assistant."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "task_type": "question_to_code",
  "source": {},
  "metadata": {},
  "split": "train",
  "split_origin": "..."
}

Dataset creation

The pipeline started from 4,400 raw English question/code pairs, applied deterministic quality rules plus full-context LLM-assisted review, derived additional task variants (test generation, bug fixing, refactoring, completion, explanation) from accepted roots, and checked for exact duplicates and near-duplicate leakage against the held-out benchmark (see the companion flutterbench dataset).

Known limitations

• Task distribution is dominated by question_to_code (~57%); api_usage is a small subset (30 records).
• Quality varies: some records are large or contain formatting artifacts from automated generation.
• Only one downstream model family (Qwen2.5-Coder-3B-Instruct) has been evaluated against this data; results should not be assumed to generalize.
• No functional/compilation verification was performed on these training examples themselves (functional verification exists for a subset of the companion benchmark, see flutterbench).

Citation

If you use this dataset, 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.