HellanodikAI is a Llama-3 8B model fine-tuned (LoRA) from GoToCompany/llama3-8b-cpt-sahabatai-v1-instruct for the Indonesian legal/educational domain. It is trained and released for educational and research purposes, and is not a substitute for professional legal advice.

Developed by: Teuku Vaickal Rizki Irdian (icaluwu.site)
Model type: Causal LM, LoRA fine-tune of Llama 3 8B
Language(s): Indonesian (primary), English
License: CC BY-NC 4.0 with additional terms — see LICENSE.md
Finetuned from model: GoToCompany/llama3-8b-cpt-sahabatai-v1-instruct

This model was trained 2x faster with Unsloth and Hugging Face TRL.

⚠️ Intended Use & Limitations

HellanodikAI is built strictly for educational and informational purposes — for example, learning about Indonesian legal concepts, drafting practice exercises, or academic research on legal-domain LLMs.

Do NOT use this model to:

Generate legally binding advice, opinions, or decisions
Replace consultation with a licensed advocate/notary/legal professional
Make decisions with real legal, financial, or personal consequences without human expert review

The model can produce incorrect, outdated, or incomplete legal information ("hallucinations"). Indonesian law changes over time (new UU, PP, Perpres, putusan MA/MK) and this model's knowledge is frozen at training time — always verify against primary legal sources (JDIH, peraturan.go.id, putusan resmi pengadilan) before relying on any output.

Model Details


Base model	GoToCompany/llama3-8b-cpt-sahabatai-v1-instruct (Llama 3 8B, ID-adapted)
Fine-tuning method	LoRA (PEFT)
LoRA rank (r)	16
LoRA alpha	32
LoRA dropout	0.05
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training framework	Unsloth + TRL (SFTTrainer)
Hardware	Kaggle free-tier T4 GPU

Training Data

Dataset details to be added.

Files & Usage

This repository provides two artifacts:

File	Format	Use case
`adapter_model.safetensors` + `adapter_config.json`	LoRA adapter (PEFT)	Load on top of the base model with `transformers` + `peft` for further fine-tuning or research
`hellanodikai-8b.Q4_K_M.gguf`	Merged & quantized GGUF	Standalone inference via llama.cpp, Ollama, LM Studio, Jan, etc.

Using the LoRA adapter (transformers + peft)

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "GoToCompany/llama3-8b-cpt-sahabatai-v1-instruct"
adapter = "icaluwu/Llama-HellanodikAI-8B"

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

Using the GGUF file (Ollama)

ollama run hf.co/icaluwu/Llama-HellanodikAI-8B:Q4_K_M

Using the GGUF file (llama.cpp)

llama-cli -hf icaluwu/Llama-HellanodikAI-8B:Q4_K_M --jinja

Currently only a Q4_K_M quantization is provided. Additional quant levels (Q5_K_M, Q8_0) may be added in future releases.

Evaluation

No formal evaluation results published yet.

License

This model is licensed under CC BY-NC 4.0 with additional terms — see LICENSE.md for full details.

✅ Free for academic, research, and non-commercial community use, with attribution to Teuku Vaickal Rizki Irdian.
🚫 Commercial use is not permitted without a separate license.
💼 For commercial licensing inquiries, contact Teuku Vaickal Rizki Irdian via LinkedIn or icaluwu.site.
As a derivative of Llama 3, this model is also subject to the Llama 3 Community License. Attribution "Built with Llama" is required per that license.

Citation

If you use this model in academic work, please cite:

@misc{hellanodikai2026,
  author = {Irdian, Teuku Vaickal Rizki},
  title = {HellanodikAI: Indonesian Legal-Domain Llama-3 8B},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/icaluwu/Llama-HellanodikAI-8B}}
}

icaluwu/Llama-HellanodikAI-8B

Llama-HellanodikAI-8B