Metric	Value
Eval perplexity (FT, r=16)	4.55
Eval perplexity (base)	5.49
Improvement vs base	−17.1 %
Train loss → final	1.61 → 1.39
Trainable params	3,686,400 (0.12 % of base)
Training time	4.28 phút (T4)
Peak VRAM	6.62 GB

⚙️ LoRA config

python

r = 16
lora_alpha = 32
target_modules = ["q_proj", "v_proj"]
lora_dropout = 0
bias = "none"
random_state = 42

📚 Training config

Setting	Value
Base model	`unsloth/Qwen2.5-3B-bnb-4bit` (NF4 4-bit)
Dataset	`5CD-AI/Vietnamese-alpaca-gpt4-gg-translated` (200 samples)
Train / Eval split	180 / 20 (90/10, seed=42)
max_seq_length	512 (p95 round-up)
Epochs	3
Learning rate	2e-4, cosine schedule
Warmup ratio	0.10
Effective batch	8 (per_device=1 × grad_accum=8)
Optimizer	`adamw_8bit` (paged AdamW)
Gradient checkpointing	unsloth (−60 % VRAM)
GPU	NVIDIA Tesla T4 (Free Colab)

🚀 Cách dùng

python

from peft import PeftModel
from unsloth import FastLanguageModel

# Load base
base_model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Qwen2.5-3B-bnb-4bit",
    max_seq_length=512,
    load_in_4bit=True,
)

# Attach this LoRA adapter
model = PeftModel.from_pretrained(base_model, "hiepphambk/lap21_2A202600415")
FastLanguageModel.for_inference(model)

# Generate
prompt = "### Instruction:\nGiải thích khái niệm machine learning cho người mới bắt đầu.\n\n### Response:\n"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=200, temperature=0.7, top_p=0.9, do_sample=True)
print(tokenizer.decode(out[0], skip_special_tokens=True))

📂 Repo gốc

GitHub: https://github.com/hoanghiepbk/Day21-Track3-Finetuning-LLMs-LoRA-QLoRA
Lab rubric: Lab21_Rubric_and_Format.md (trong repo gốc)
Full evaluation report: REPORT.md (trong repo gốc)

📜 References

LoRA — Hu et al. 2021 — https://arxiv.org/abs/2106.09685
QLoRA — Dettmers et al. 2023 — https://arxiv.org/abs/2305.14314
Unsloth — https://github.com/unslothai/unsloth