LikithMeruvu/Gemma2B_Finetuning_Medium

This Repo contains How to Finetune Google's New Gemma LLm model using your custom instuction dataset. I have finetuned Gemma 2b instuct Model on 20k medium articles data for 5hrs on kaggle p100 GPU

Gemma 2b Instruction Fine-Tuned Model

Introduction

This repository contains the code and model for fine-tuning the Gemma 2b Instruction model using a dataset of 20k medium articles. The fine-tuned model is designed to generate responses based on input prompts related to instructional queries in English.

Model Details:

• Model Name: Gemma 2b Instruction Fine-Tuned Model
• Framework: Keras with JAX backend
• Model Type: Causal Language Model
• Pre-trained Model: Gemma 2b Instruction Model

Fine-Tuning Process

The Gemma 2b Instruction Fine-Tuned Model was fine-tuned using the following process:

1. Data Preparation: 20k medium articles were used as the training dataset. access Dataset.
2. Model Configuration: Gemma 2b Instruction model was used as the base model.
3. Hyperparameters Tuning: AdamW optimizer was used with customized learning rates and weight decays.
4. Training: The model was trained for 2 epochs with a batch size of 2 Due to lack of GPU Clusters to handle this 2B model even with LORA config. Kaggle.

Usage

Environment Setup

Ensure you have the necessary dependencies installed:

pip install keras keras-nlp

Inference

To use the fine-tuned model for inference, follow these steps:

1. Download the Model:

   • The fine-tuned model file version_finetuned.keras can be downloaded from the Hugging Face Model Hub.

2. Inference:

from keras.models import load_model

# Load the saved model
loaded_model = load_model("version_finetuned.keras")
instruction = "How to code in python, Give me An example code"
response = ""

# Create the input prompt
prompt = f"Instruction:\n{instruction}\n\nResponse:\n{response}"
# Generate inference using the loaded model
inference_result = loaded_model.generate(prompt, max_length=1024)

# Print or use the generated response
print(inference_result)

Model Card

For more information about the model, check out its model card on Hugging Face Model Hub.

License

This project is licensed under the MIT License.