---
datasets:
- zhiman-ai/alpaca_en_demo
base_model:
- meta-llama/Llama-3.2-3B-Instruct
tags:
- lora
- sft
---

# **LLAMA-3.2-3B-Alpaca_en_LORA_SFT**

This model is a fine-tuned version of [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) using the [alpaca_en_demo](https://huggingface.co/datasets/zhiman-ai/alpaca_en_demo) dataset. The fine-tuning process was conducted by **Sri Santh M** for development purposes. 

It achieves the following results on the evaluation set:
- **Loss**: 1.0510

---

## **Model Description**

This model is optimized for tasks involving instruction-following, text generation, and fine-tuned identity-based use cases. It leverages the capabilities of the LLaMA-3.2-3B-Instruct base model with additional refinements made using a lightweight fine-tuning approach via PEFT (Parameter-Efficient Fine-Tuning).  

---

### **Intended Uses**
- Instruction-following tasks.
- Conversational AI and question-answering applications.
- Text summarization and content generation.
---

## **Training and Evaluation Data**

The model was fine-tuned using the **alpaca_en_demo** dataset, which is designed for instruction-tuned task completion. This dataset includes diverse English-language tasks for demonstrating instruction-following capabilities.  
- **Dataset link**: [alpaca_en_demo](https://huggingface.co/datasets/zhiman-ai/alpaca_en_demo)

Further details on the dataset:
- **Source**: zhiman-ai.
- **Size**: Small-scale, development-focused dataset.
- **Purpose**: Designed to emulate instruction-tuned datasets like Alpaca, with a subset of English-language prompts and responses.

---

## **Training Procedure**

### **Hyperparameters**
- **Learning rate**: 0.0001  
- **Train batch size**: 1  
- **Eval batch size**: 1  
- **Gradient accumulation steps**: 8  
- **Total effective batch size**: 8  
- **Optimizer**: AdamW (torch)  
  - **Betas**: (0.9, 0.999)  
  - **Epsilon**: 1e-08  
- **Learning rate scheduler**: Cosine schedule with 10% warmup.  
- **Number of epochs**: 3.0  

### **Frameworks and Libraries**
- **PEFT**: 0.12.0  
- **Transformers**: 4.46.1  
- **PyTorch**: 2.4.0  
- **Datasets**: 3.1.0  
- **Tokenizers**: 0.20.3  

---

## **Training Results**

- **Loss**: 1.0510  
- Evaluation results are limited to the dataset scope. Broader testing is recommended for downstream applications.

---

## **Additional Information**

- **Author**: Sri Santh M  
- **Purpose**: Fine-tuned for development and experimentation purposes using the LLaMA-3.2-3B-Instruct model.  

This model serves as an experimental proof-of-concept for lightweight fine-tuning using PEFT and can be adapted further based on specific tasks or use cases.