Llama-3.2-1B Fine-Tuned Model

Overview

Model Details

• Base Model: meta-llama/Llama-3.2-1B
• Fine-Tuned Model: fatihcihan/llama-3.2-1b-finetuned
• Parameters: 1.24 billion
• Fine-Tuning Method: QLoRA with 4-bit quantization

Dataset

• Name: databricks/databricks-dolly-15k
• Description: 15,000 instruction-following records covering tasks like question * answering, summarization, and brainstorming.
• Split: 90% training (13,500 samples), 10% validation (1,500 samples)

Fine-Tuning Process

• Platform: Google Colab (free tier with GPU)
• Libraries: transformers, accelerate, peft, trl, bitsandbytes, datasets
• Hyperparameters:
• Epochs: 1
• Batch Size: 4
• Learning Rate: 2e-4
• LoRA Parameters: r=16, lora_alpha=32, lora_dropout=0.1

Training Time: Approximately 1-2 hours

• Evaluation
• Perplexity was computed on the validation set for both the base and fine-tuned models. Results will be updated after evaluation.
• Citations
• Hugging Face trl Documentation - SFTTrainer
• GitHub Issue on trl/peft - processing_class and tokenizer arguments