README.md

metadata

title: Salesforce CodeT5 Large Demo
emoji: ⚡
colorFrom: indigo
colorTo: gray
sdk: gradio
sdk_version: 5.24.0
app_file: app.py
pinned: false
license: apache-2.0
datasets:
  - CodeSearchNet/codesearchnet_python
  - bigcode/the-stack-dedup
  - codeparrot/codeparrot-clean
  - openai_humaneval
  - google/mbpp
  - nvidia/OpenCodeReasoning
hf_oauth: true
hf_oauth_scopes:
  - inference-api
short_description: Using the powerful Salesforce CodeT5-large model

⚡ Salesforce CodeT5-large Demo ⚡

Welcome! This repository/Hugging Face Space hosts a demonstration application for the powerful Salesforce CodeT5-large model. It showcases the model's capabilities in various code intelligence tasks using a Gradio interface.

About CodeT5-large

CodeT5 is an advanced encoder-decoder transformer model pre-trained on a vast collection of source code from multiple programming languages alongside natural language text. The codet5-large variant excels at tasks such as:

• Code Generation: Creating code snippets from natural language descriptions (e.g., comments, docstrings).
• Code Summarization: Generating concise natural language summaries for given code blocks.
• Code Translation: Translating code from one programming language to another.
• Code Refinement: Improving code quality, fixing bugs, or optimizing code.

Using the Demo (Hugging Face Space)

This application is built with Gradio, providing an interactive web UI.

1. Access the Space: Navigate to the Hugging Face Space hosting this demo.
2. Interact: Use the input fields provided by the Gradio interface (app.py) to interact with the model.
   • (Example: You might enter a Python docstring in one box to get the generated function body in another, or input code to get a summary. Please update this section with specific instructions based on your app.py functionality!)
3. Observe: See the results generated by the CodeT5-large model in the output fields.

Running Locally (GitHub / Manual Setup)

If you prefer to run this demo on your local machine:

1. Clone the Repository:

   git clone <repository_url> # Replace with HF Space or GitHub repo URL
   cd <repository_directory>
   

2. Set up Environment: (Optional but recommended) Create and activate a virtual environment:

   python -m venv venv
   source venv/bin/activate # Linux/macOS
   # venv\Scripts\activate # Windows
   

3. Install Dependencies: Ensure you have Python 3 installed. You'll need Gradio and the necessary libraries for CodeT5 (like transformers and torch). Create a requirements.txt file if one doesn't exist:

   # requirements.txt
   gradio==5.23.3
   transformers
   torch
   # Add any other specific libraries your app.py needs
   

   Then install:

   pip install -r requirements.txt
   

4. Run the Application:

   python app.py
   

5. Access Locally: Open your web browser and navigate to the URL provided (typically http://127.0.0.1:7860).

Fine-tuning Datasets for Python & Logic

The CodeT5 model's performance on specific Python tasks or logical reasoning can be enhanced through fine-tuning. Here are some recommended datasets included in the metadata:

• CodeSearchNet (Python): Excellent for tasks involving matching natural language queries to relevant Python code snippets.
• The Stack (Deduped): A massive, permissively licensed dataset. Filter for Python files (lang:python) for broad fine-tuning on diverse Python code.
• CodeParrot (Clean): A high-quality dataset specifically curated for Python code generation tasks.
• HumanEval: A benchmark dataset consisting of Python function programming problems defined by docstrings, ideal for fine-tuning code generation based on specifications and evaluating functional correctness.
• MBPP (Mostly Basic Python Problems): Contains around 1,000 crowd-sourced Python programming problems focused on basic concepts, useful for improving generation from descriptions and simple logical problem-solving.

License

This project and the underlying CodeT5 model are distributed under the terms of the Apache License 2.0. Please refer to the LICENSE file for details.