---
license: mit
task_categories:
  - translation
language:
  - en
tags:
  - code
  - python
  - rust
  - transpilation
  - compiler
pretty_name: Depyler CITL Corpus
size_categories:
  - n<1K
---

# Depyler CITL Corpus

Python→Rust transpilation pairs for Compiler-in-the-Loop training.

## Dataset Description

606 Python CLI examples with corresponding Rust translations (where available), designed for training transpiler ML models.

| Split | Examples | With Rust | Size |
|-------|----------|-----------|------|
| train | 606 | 439 (72.4%) | 957 KB |

## Schema

```
- example_name: str           # Directory name (e.g., "example_fibonacci")
- python_file: str            # Python filename
- python_code: str            # Full Python source
- rust_code: str              # Corresponding Rust (empty if not transpiled)
- has_rust: bool              # Whether Rust translation exists
- category: str               # Extracted category
- python_lines: int           # Line count
- rust_lines: int             # Line count
- blocking_features: [str]    # Detected Python features (v2)
- suspiciousness: float       # Tarantula score 0-1 (v2)
- error: str                  # Transpilation error if failed (v2)
```

## Tarantula Fault Localization

The `corpus_insights.json` file contains fault localization analysis using the Tarantula algorithm from [entrenar](https://github.com/paiml/entrenar) CITL.

### Priority Features (by suspiciousness score)

| Feature | Score | Categories Affected | Priority |
|---------|-------|---------------------|----------|
| async_await | 0.946 | 4 | P0 |
| generator | 0.927 | 14 | P0 |
| walrus_operator | 0.850 | 1 | P1 |
| lambda | 0.783 | 29 | P1 |
| context_manager | 0.652 | 93 | P2 |

Higher suspiciousness = more correlated with transpilation failures.

### Insights File Structure

```json
{
  "summary": { "total_pairs": 606, "success_rate": 71.9 },
  "tarantula_fault_localization": { "scores": {...} },
  "priority_features_to_implement": [...],
  "zero_success_categories": [...],
  "category_insights": {...}
}
```

Regenerate with: `python3 scripts/generate_insights.py`

## Usage

```python
from datasets import load_dataset

ds = load_dataset("paiml/depyler-citl")

# Filter to pairs with Rust translations
pairs = ds["train"].filter(lambda x: x["has_rust"])

for row in pairs:
    print(f"Python: {row['python_lines']} lines → Rust: {row['rust_lines']} lines")
```

## Related Projects

- [depyler](https://github.com/paiml/depyler) - Python→Rust transpiler
- [alimentar](https://github.com/paiml/alimentar) - Dataset loading library
- [entrenar](https://github.com/paiml/entrenar) - ML training with CITL

## License

MIT