code-smell-detection v1.0 — 103 hand-authored examples, 20 smell types, 6 languages
#1
by SalutefaceAAC - opened
code-smell-detection v1.0 — now live!
Excited to share Code Smell Detection, a curated synthetic dataset of 103 hand-authored code snippets labelled with 20 classic software engineering code smells across 6 languages.
What's inside
- 20 smell types from Fowler's Refactoring taxonomy: Long Method, God Object, Feature Envy, Data Clumps, Message Chains, Refused Bequest, Parallel Inheritance, and 13 more
- 6 languages: Python, JavaScript, TypeScript, Go, Java, Rust
- 3 severity levels (low / medium / high) + a clean negative class (~6%) to prevent trivial "always-smelly" classifiers
- 3 difficulty levels (beginner / intermediate / advanced)
- Each example includes the raw code snippet, a plain-English explanation of why it's a smell, and a concrete refactoring suggestion
- Splits: train (72) / val (15) / test (16) in both JSONL and Parquet
Design principles
- Authenticity: snippets look like real production code, not toy examples
- Language-idiomatic: Go interfaces, Rust newtypes, Java Spring patterns, TypeScript generics are all respected
- Negative class included: 6 clean examples prevent models from learning trivial shortcuts
- Severity calibrated:
high= structural problem affecting testability;medium= noticeable but contained;low= minor maintenance debt
Use cases
- Fine-tuning or evaluating LLM code-review capability
- Training lightweight classifiers (smell type or severity)
- Few-shot prompting for automated PR review
- Teaching material for software engineering courses
Quick start
from datasets import load_dataset
ds = load_dataset("SalutefaceAAC/code-smell-detection")
print(ds["train"][0])
Related datasets in this series
Feedback, issues, and PRs very welcome — especially suggestions for smell types or language coverage to prioritize in v1.5!