🔧 LLM Fine-Tuning

LoRA · QLoRA · Unsloth · Low-VRAM Training

Hands-on fine-tuning experiments on Large Language Models (LLMs) with a focus on real inference behavior.

---

🚀 What This Repository Covers

• Fine-tuning open-source LLMs on low VRAM GPUs
• LoRA and QLoRA based training workflows
• Unsloth optimized training & inference
• Hyperparameter impact analysis
• Base vs Fine-Tuned inference comparison

🧠 Models Fine-Tuned

• LLaMA-family models
• Mistral
• Qwen
• Nemotron

🛠️ Techniques Used

🔹 LoRA (Low-Rank Adaptation)

• Trainable low-rank adapters
• Frozen base model weights
• Efficient and stable fine-tuning

🔹 QLoRA

• 4-bit quantized base model
• Enables fine-tuning on consumer GPUs
• Minor trade-off in long-context precision

🔹 Unsloth

• Faster training and inference
• Reduced VRAM usage
• Optimized attention & adapter kernels

📂 Repository Structure

⚙️ Training Setup

Platform	Google Colab / Kaggle
GPU	T4 / P100
Precision	FP16 / 4-bit
Batching	Gradient Accumulation

📊 Hyperparameters Explored

Parameter	Values Tested
LoRA Rank (r)	8, 16, 32
Alpha	16, 32, 64
Learning Rate	1e-4 → 2e-5
Epochs	1 – 5

🔍 Inference Observations

• Improved instruction adherence
• Reduced hallucination for domain prompts
• More structured outputs
• Overfitting with small datasets + high epochs

🧪 Evaluation Strategy

• Prompt-based comparison (before vs after)
• Task completion accuracy
• Response consistency and structure

🧠 Key Learnings

• Dataset quality > model size
• LoRA rank tuning matters more than epochs
• Unsloth speeds up experimentation significantly
• Inference behavior is the real success metric

🧭 Next Experiments

• RAG + fine-tuned models
• Multi-dataset instruction blending
• Adapter merging & export
• Deployment-focused quantization

---

⚠️ Disclaimer: All experiments are for educational and research purposes using publicly available models and datasets.