---
library_name: peft
---

### README for Gemma-2-2B-IT Fine-Tuning with LoRA

This project fine-tunes the `Gemma-2-2B-IT` model using **LoRA (Low-Rank Adaptation)** for Question Answering tasks, leveraging the `Wikitext-2` dataset. The fine-tuning process is optimized for efficient training on limited GPU memory by freezing most model parameters and applying LoRA to specific layers.

### Project Overview
- **Model**: `Gemma-2-2B-IT`, a causal language model.
- **Dataset**: `Wikitext-2` for text generation and causal language modeling.
- **Training Strategy**: LoRA adaptation for low-resource fine-tuning.
- **Frameworks**: Hugging Face `transformers`, `peft`, and `datasets`.

### Key Features
- **LoRA Configuration**: 
  - LoRA is applied to the following projection layers: `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, and `down_proj`.
  - LoRA hyperparameters: 
    - Rank (`r`): 4
    - LoRA Alpha: 8
    - Dropout: 0.1
- **Training Configuration**:
  - Mixed precision (`fp16`) enabled for faster and more memory-efficient training.
  - Gradient accumulation with `32` steps to manage large model sizes on small GPUs.
  - Batch size of 1 due to GPU memory constraints.
  - Learning rate: `5e-5` with weight decay: `0.01`.

### System Requirements
- **GPU**: Required for efficient training. This script was tested with CUDA-enabled GPUs.
- **Python Packages**: Install dependencies with:
   ```bash
   pip install -r requirements.txt
   ```

### Notes
- This fine-tuned model leverages LoRA to adapt the large `Gemma-2-2B-IT` model with minimal trainable parameters, allowing fine-tuning even on hardware with limited memory.
- The fine-tuned model can be further utilized for tasks like Question Answering, and it is optimized for resource-efficient deployment.

### Memory Usage
- The training script includes CUDA memory summaries before and after the training process to monitor GPU memory consumption.