Rupa421/TinyLlama-LoRA

TinyLlama LoRA Fine-Tuning

This repository contains a LoRA adapter fine-tuned on the Alpaca instruction dataset using PEFT and TRL.

Base Model

• TinyLlama/TinyLlama-1.1B-Chat-v1.0

Training Details

• Framework: Hugging Face Transformers
• PEFT Method: LoRA
• Trainer: TRL SFTTrainer
• Dataset: tatsu-lab/alpaca
• Epochs: 1
• LoRA Rank (r): 8
• LoRA Alpha: 16
• LoRA Dropout: 0.05
• Target Modules:
  • q_proj
  • v_proj

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

tokenizer = AutoTokenizer.from_pretrained(
    "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
)

base_model = AutoModelForCausalLM.from_pretrained(
    "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
)

model = PeftModel.from_pretrained(
    base_model,
    "Rupa421/TinyLlama-LoRA"
)

Example Prompt

### Instruction:
Explain what LoRA is.

### Response:

Project Goal

This project was built to demonstrate an end-to-end parameter-efficient fine-tuning (PEFT) workflow using:

• Hugging Face Transformers
• PEFT (LoRA)
• TRL SFTTrainer
• TinyLlama

The project covers dataset formatting, prompt engineering, LoRA configuration, supervised fine-tuning, inference, and publishing the trained adapter to Hugging Face.

Author

Built with ❤️ while learning PyTorch Internals and Transformer Architecture.