Llama 3.2 3 B appreciation

Introduction

Llama-3.2-3B-appreciation is an AI model designed to assist French secondary school teachers in automatically generating student evaluations. The model facilitates the creation of personalized feedback based on various parameters such as student performance and behavior.

Architecture

The model uses a medium-sized language model with 3 billion parameters. It is built on the Llama 3.2 3B-Instruct base model and utilizes the PEFT (Parameter-Efficient Fine-Tuning) library. The input interface is designed with Gradio and takes into account subjects, student levels, and evaluation metrics like grades and behavior.

Training

The training process is divided into two phases:

• Phase 1: Minimum Viable Product (MVP) involved using a medium-sized LLM with a dataset of approximately 250 anonymized teacher evaluations, resulting in a basic but functional interface.
• Phase 2: Improvement and Validation expanded the dataset to over 1000 examples, involved fine-tuning a larger model, and included validation by a panel of teachers. Continuous improvements are guided by user feedback.

Guide: Running Locally

1. Requirements:

   • Install necessary libraries such as gradio, transformers, and torch.
   • Set the HF_TOKEN environment variable for model access.

2. Setup:

   • Clone the repository from Hugging Face.
   • Load the model and tokenizer using the provided Gradio interface script.

3. Execution:

   • Test the model using a Jupyter Notebook for inference.
   • Change the Colab environment to use T4 GPUs for optimal performance.

4. Cloud GPUs:

   • Utilize cloud GPU services like Google Colab with T4 instances for faster processing.

License

Llama-3.2-3B-appreciation is licensed under the AGPL-3.0, which stipulates that any use of the model must also share the source code and modifications under the same license.