llama3.2-3b-dolly-qlora

Llama-3.2-3B — QLoRA Fine-Tuned on Dolly-15K

Fine-tuned version of meta-llama/Llama-3.2-3B-Instruct using QLoRA
(4-bit NF4 quantization + LoRA adapters) on a 500-example subset of
the Databricks Dolly-15K dataset.

📎 Training code: github.com/Mervecaliskann/llama3-qlora-finetuning

---

Model Details

---

Dataset

---

Training Hyperparameters

---

Training Results

Train loss: 1.87 → 1.49 | Val loss: 1.77 → 1.57 | No overfitting

Val loss consistently decreases alongside train loss — healthy convergence for a 500-example fine-tuning run.

---

How to Use

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

base_model = "meta-llama/Llama-3.2-3B-Instruct"
adapter    = "Mervecaliskan/llama3.2-3b-dolly-qlora"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model     = AutoModelForCausalLM.from_pretrained(
    base_model,
    load_in_4bit=True,
    device_map="auto"
)
model = PeftModel.from_pretrained(model, adapter)

prompt = """<|begin_of_text|><|start_header_id|>system<|end_header_id|>
You are a helpful assistant.<|eot_id|>
<|start_header_id|>user<|end_header_id|>
What causes flight delays?<|eot_id|>
<|start_header_id|>assistant<|end_header_id|>
"""

inputs  = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=200, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Sample Outputs

open_qa — What causes the Northern Lights?

The Northern Lights are caused by charged particles from the sun interacting with Earth's magnetic field and atmosphere near the poles. ✅

classification — Is "I absolutely loved the movie!" positive or negative?

Positive ✅

summarization — Summarize the French Revolution in one sentence.

The French Revolution (1789–1799) transformed France's political structure and led to the rise of Napoleon Bonaparte. ⚠️ (followed instruction partially)

brainstorming — Give 3 creative names for a coffee shop.

Brew (repeated) ❌ (instruction following weak on small dataset — expected)

---

Limitations

• Trained on 500 examples only — a portfolio/demo project
• Instruction following is inconsistent for multi-item generation tasks
• Not intended for production use
• For better results: increase dataset size to 5K–15K examples

---

MLflow Experiment

Experiment: llama3-dolly-finetuning
Run: qlora-r16-lr2e4-dolly500
Tracked: step-level train/val loss, GPU peak memory, all hyperparameters, environment tags

---

Fine-tuned as part of an AI engineering portfolio project.
Training notebook: github.com/Mervecaliskann/llama3-qlora-finetuning