sangtran12/eduflowai-gemma4b-adapter-lora

🎓 EduFlowAI Gemma 3 4B LoRA (bf16) Adapter

This repository provides the official LoRA (uncompressed bfloat16) fine-tuned adapter for unsloth/gemma-3-4b-it, serving as the full-precision benchmark comparison arm for Master's Thesis Research Question 2 (RQ2).

---

🔬 Comparative Benchmark: QLoRA vs LoRA (RQ2)

Dimension	QLoRA (4-bit NF4)	LoRA (bfloat16)	Academic Finding
Peak VRAM	13.3 GB	17.2 GB	LoRA requires +29.3% more VRAM, necessitating $\ge$24GB workstation GPUs.
Train Time (1 Epoch)	21.7 min	21.0 min	NF4 dequantization incurs negligible overhead (<3%) on Blackwell architecture.
Convergence Loss	2.6323	14.1419	QLoRA demonstrates superior representation stability on Gemma 3's 256k vocabulary.
Validation Loss	10.09	24.96	QLoRA regularizes fine-tuning and reduces task over-adaptation.

---

🚀 How to Use with Unsloth

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "sangtran12/eduflowai-gemma4b-adapter-lora",
    max_seq_length = 8192,
    load_in_4bit = False, # bf16 precision
)
FastLanguageModel.for_inference(model)