TinyLLaMA-Indo-Property 🏠

Fine-tuned TinyLlama-1.1B-Chat-v1.0 model using the LoRA (Low-Rank Adaptation) method for the Indonesian-language property consultation domain. This model was developed as part of an undergraduate thesis research project with three experimental scenarios based on LoRA rank variations.

---

📋 Model Information

Attribute	Detail
Base Model	TinyLlama/TinyLlama-1.1B-Chat-v1.0
Fine-Tuning Method	LoRA (Low-Rank Adaptation) via PEFT
Domain	Indonesian-language property consultation
Language	Indonesian 🇮🇩
Framework	HuggingFace Transformers + TRL (SFTTrainer)
Training Environment	Google Colab (GPU)

---

🧪 Experiment Scenarios

This research tested three scenarios with varying LoRA hyperparameters as follows:

Scenario	LoRA Rank (r)	LoRA Alpha (α)	LoRA Dropout	Learning Rate	Epoch
Scenario 3	32	64	0	2e-4	12

---

🎬 Proof of Concept (PoC)

The fine-tuned model was integrated into an "AI Property Consultant" chatbot embedded on a property demo website (Opendoorz) as proof of the model's application in a real-world use case.

⚠️ Note: The source code for this demo/PoC (the Opendoorz website and chatbot UI) lives in a separate repository, which is currently private. Only the screenshots below are shown here for documentation purposes; the demo repository is not publicly accessible.

1. Landing Page — Property Website (Opendoorz)

The chatbot is integrated as a widget on the property website's homepage.

2. Conversation Demo — Mortgage (KPR) Questions

Example interaction where a user asks about additional costs beyond the down payment for a mortgage, as well as a topic guardrail test (the chatbot declines to answer questions outside the property domain, such as motorcycle prices).

3. Conversation Demo — Credit Analysis & Document Legality (AJB)

Example follow-up interaction regarding mortgage eligibility based on salary & existing installments, as well as questions about the legality of a house sale between relatives (AJB / deed of sale).

Note: This PoC demonstrates that the model can run end-to-end in a web-based chatbot scenario and is able to answer property-domain questions (mortgages, legal documents, transaction costs), including restricting its responses to questions outside the property topic.

---

🚀 Usage

Install dependencies

pip install -r requirements.txt

Load model and run inference

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

# Choose the scenario you want to use
SCENARIO = "skenario-3-new"  # change as needed

base_model_id = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
lora_path = f"RizalGo/TinyLLaMA-Indo-Property/{SCENARIO}/lora"

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model_id)

# Load base model
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(model, lora_path)
model.eval()

# Inference
def chat(question: str) -> str:
    messages = [
        {
            "role": "system",
            "content": "You are a professional property consultant assistant that helps answer questions about property in Indonesia."
        },
        {
            "role": "user",
            "content": question
        }
    ]
    prompt = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=512,
            temperature=0.7,
            do_sample=True,
            pad_token_id=tokenizer.eos_token_id
        )

    response = tokenizer.decode(
        outputs[0][inputs["input_ids"].shape[-1]:],
        skip_special_tokens=True
    )
    return response

# Example usage
question = "What is the difference between SHM and HGB property ownership titles?"
print(chat(question))

---

📁 Repository Structure

TinyLLaMA-Indo-Property/
├── data-2/                  # Datasets
│   ├── merged/
        ├── data-real-estate-expert-reformat.json  
├── skenario-3-new/          # Scenario 3 (r=32)
│   ├── checkpoints/
│   ├── lora/
│   └── merged/             # the models .safetensor
|   └── skenario-3.gguf     # the models .gguf
├── images/                  # PoC documentation (demo screenshots)
└── README.md

---

⚙️ Training Details

# LoRA configuration (Scenario 3 example)
lora_config = LoraConfig(
    r=32,
    lora_alpha=64,
    lora_dropout=0.0,
    target_modules=["q_proj", "v_proj"],
    bias="none",
    task_type="CAUSAL_LM"
)

# Training configuration
training_args = SFTConfig(
    learning_rate=2e-4,
    num_train_epochs=12,
    per_device_train_batch_size=4,
    fp16=True,
)

---

📊 Evaluation

The model was evaluated using BLEU-4 and BERTScore metrics on an Indonesian-language property test dataset (300 samples), comparing the performance of the three LoRA rank scenarios.

Evaluation Results — Scenario 3 (r=32)

Metric	Score
BLEU-4 (0–100)	14.54
BERTScore — Precision	0.7533
BERTScore — Recall	0.7442
BERTScore — F1	0.7487

BERTScore Comparison: Before vs After LoRA Fine-Tuning

Model	BERTScore-F1
Base model (TinyLlama, before fine-tuning)	~0.65 (65%)
Fine-tuned (Scenario 3, LoRA r=32)	0.7487 (74.87%)

LoRA fine-tuning improved BERTScore-F1 by roughly +9.87 percentage points compared to the base model, showing that the fine-tuned model captures the semantic similarity of answers far better against the Indonesian-language property domain references.

BERTScore-F1 per Intent (Scenario 3)

Intent	BERTScore-F1
biaya_dan_pajak (costs & taxes)	0.7568
spesifikasi_properti (property specifications)	0.7548
proses_transaksi (transaction process)	0.7534
legal_dokumen (legal documents)	0.7532
konsultasi_keputusan (decision consultation)	0.7527
risiko_dan_disclaimer (risk & disclaimer)	0.7496
kpr_simulasi (mortgage simulation)	0.7199

Note: The sizable gap between BERTScore-F1 (0.7487) and BLEU-4 (0.1454) (+0.60) indicates that the model produces answers with meaning/semantics that align with the reference even though the specific word choices (lexical level) differ — which is expected for a natural-language generative QA task.

---

📚 References

• Hu, E. J., et al. (2021). LoRA: Low-Rank Adaptation of Large Language Models. arXiv:2106.09685
• Biderman, D., et al. (2024). LoRA Learns Less and Forgets Less. Transactions on Machine Learning Research.
• Zhou, C., et al. (2023). LIMA: Less Is More for Alignment. NeurIPS 2023.
• TinyLlama: https://github.com/jzhang38/TinyLlama

---

👤 Author

Rizal — Fachrizal Fazza Ashari

---

📄 License

Follows the license of the base model TinyLlama-1.1B-Chat-v1.0 (Apache 2.0).