---
language:
- en
- py
license: mit
tags:
- code-generation
- t5
- text2text-generation
- python
pipeline_tag: text2text-generation
widget:
- text: "translate English to Python: create a list of numbers from 1 to 10"
- text: "translate English to Python: write a function that calculates factorial"
---

# NL2Py v2: English-to-Python Neural Code Translator

[![GitHub Repository](https://img.shields.io/badge/GitHub-abdullah00ashraf-181717.svg?logo=github)](https://github.com/abdullah00ashraf/abdullah00ashraf)
[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](https://opensource.org/licenses/MIT)

> **Parent GitHub Repository**: [https://github.com/abdullah00ashraf/abdullah00ashraf](https://github.com/abdullah00ashraf/abdullah00ashraf)  
> **Ecosystem Documentation**: [Axiyon / Producttolaunch Catalog](https://github.com/abdullah00ashraf/sentinel-hufp-v7)


Part of the **Axiyon Research Ecosystem**.  
**Author:** Abdullah Ashraf ([@abdullahashraf122](https://huggingface.co/abdullahashraf122))  
**GitHub Profile:** [https://github.com/abdullah00ashraf](https://github.com/abdullah00ashraf)

## Model Overview
NL2Py v2 is a specialized seq2seq transformer fine-tuned on the `iamthedan/code-instructions-25k-python` dataset with structural prompt prefixes (`translate English to Python: `). It generates syntax-valid Python code snippets directly from natural language specifications.

### Architecture Highlights
* **Base Architecture**: T5-small (60M parameters)
* **Tokenizer**: SentencePiece with custom prefix masking
* **Max Length**: 128 input tokens $\to$ 512 target tokens
* **Loss Function**: Cross-Entropy with label smoothing

### Interactive Demo
Run the included `app.py` or execute the training pipeline via `NL2Py_v2_Colab_Pipeline.ipynb`.