README.md

---
license: apache-2.0
task_categories:
  - text-generation
  - text2text-generation
language:
  - en
  - ur
  - am
  - zh
tags:
  - code
  - multilingual
  - legesher
  - transpilation
  - tiny-aya-expedition
  - language-decoded
pretty_name: Language Decoded Data
size_categories:
  - 10K<n<100K
---

# Language Decoded | Multilingual Code Dataset

Multilingual Python code datasets for the **Language Decoded** project (part of Cohere's Tiny Aya Expedition), investigating whether code's reasoning benefit for language models is **language-dependent** or **structure-dependent**.

## Research Question

> Does fine-tuning on non-English code (Python with translated keywords) improve multilingual reasoning as much as English code does?

Prior work ([Aryabumi et al., 2024](https://arxiv.org/abs/2408.10914)) showed English code improves English reasoning by 8.2%, but never tested non-English code. This dataset enables that experiment.

## Dataset Structure

This repo contains multiple experimental conditions as subdirectories:

| Subdirectory | Condition | Description |
|---|---|---|
| `source-python/` | Source | Filtered Python files from The Stack (shared base) |
| `baseline/` | Condition 1 | No code augmentation (control) |
| `english-code/` | Condition 2 | Original English-keyword Python code |
| `multilingual-code-ur/` | Condition 3a | Python transpiled to Urdu keywords via Legesher |
| `multilingual-code-am/` | Condition 3b | Python transpiled to Amharic keywords via Legesher |
| `multilingual-code-zh/` | Condition 3c | Python transpiled to Chinese keywords via Legesher |
| `multilingual-text/` | Condition 4 | Non-code multilingual text (control) |

## Usage

```python
from datasets import load_dataset

# Load a specific condition
ds = load_dataset("Legesher/language-decoded-data", data_dir="multilingual-code-ur")
```

## Transpilation

Code translation is performed using [Legesher](https://github.com/Legesher/legesher), which translates Python reserved words (keywords, builtins, exceptions) into target languages while preserving code structure and semantics.

Example (English → Chinese):

```python
# English
for item in range(10):
    if item > 5:
        print(item)

# Chinese / 中文 (via Legesher)
循环 元素 在 范围(10):
    如果 元素 > 5:
        打印(元素)
```

## Source Data

- **Base**: [The Stack](https://huggingface.co/datasets/bigcode/the-stack-dedup) — permissively licensed Python subset
- **Filtering**: Quality-filtered to 50K-100K files
- **Transpilation tool**: [Legesher v0.6.0+](https://github.com/Legesher/legesher)

## Evaluation Benchmarks

Models fine-tuned on these conditions are evaluated on:

- **XNLI** — Cross-lingual natural language inference (15 languages)
- **XStoryCloze** — Story completion (11 languages)
- **TyDi QA** — Question answering (11 languages)
- **MMLU** — Multilingual knowledge

## Related Resources

- **Models**: [Legesher/language-decoded-lora](https://huggingface.co/Legesher/language-decoded-lora) — LoRA adapters trained on these conditions
- **Community code**: [Legesher/language-decoded-community](https://huggingface.co/datasets/Legesher/language-decoded-community) — Human-written native language code
- **Experiments**: [Legesher/language-decoded-experiments](https://huggingface.co/datasets/Legesher/language-decoded-experiments) — Training logs and eval results
- **Paper**: Coming soon

## Citation

```bibtex
@misc{language-decoded-2026,
  title={Language Decoded: Investigating Language-Dependent vs. Structure-Dependent Reasoning Benefits of Code},
  author={Madison Edgar and Saad Bazaz and Rafay Mustafa and Sarah Jawaid and Rashik Shahjahan and Khojasteh Mirza and Sohaib Bazaz},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/Legesher/language-decoded-data}
}
```

## License

Apache 2.0