sahiljani / AIDI_1002_Final_Project

TinyLLM

This repository contains the code and instructions for fine-tuning the TinyLLM model on a color dataset and using it to describe colors based on their names.

Folder Structure

• tinyLLM_model/: Contains TinyLLM model.
• contribution/: TinyLLM with new datasets for evaluation in various contexts.

Installation

!pip install accelerate peft bitsandbytes transformers trl

Data Preparation

The color dataset used for fine-tuning is located at burkelibbey/colors. The data is reformatted to fit the ChatML format.

Fine-Tuning

• Load the dataset and prepare the training data.
• Initialize the TinyLLM model and tokenizer.
• Configure LoRA (LoraConfig) and training arguments.
• Train the model using SFTTrainer from trl.

Merging with Base Model

After fine-tuning, the LoRA is merged with the base TinyLLM model to produce the final model.

Inference

Generate color descriptions using the fine-tuned TinyLLM model.

Usage

1. Fine-tune the TinyLLM model by running the provided notebook.
2. dd the Hugging Face token to your notebook with the name HF_TOKEN as the Secret key.
3. Generate color descriptions using the generate_response function.

generate_response(user_input='Yellow color')