LLM_Ciptaker (Final Project NLP Indonesia AI)

Beta Demo (Thanks to @shabir67)

Data Source:

• Raw dataset (PDF): RUU Cipta Kerja
• QA-Generated pairs: SFT data prep Notebook

Train Fine-Tune LLM (with LoRA):

• Model base: "sarahlintang/mistral-indo-7b"
• Fine Tuned model: https://huggingface.co/Willy030125/CiptakerLM-v1

Retrieval Augmented Generation (RAG) testing:

• RAG model: "llama3.1"
• Embedding model: "nomic-embed-text"
• RAG library: LangChain with Unstructured PDF Loader
• Notebook: here

Frontend

• Pre-requisites: Node version v20.17.0

Quick Run:

cd FE
npm install
npm start

Full Deployment:

Auto: Recreate the repo for FE files and host on Vercel.

Manual:

cd FE
npm install -g serve
serve -s build

Backend

To run the backend:

cd BE
python app.py

If hosted on a different PC, you may need a public IP or tunneling. Read more about tunneling here:

• Ngrok - Getting Started

Notebook

• Ready to be run on Google Colab. To localize the pip freeze for each notebook, paste it into requirements.txt. For the complete guide (recommended to use Conda):
• Setting up a Conda environment in less than 5 minutes

Evaluation

RAG:

• Cosine similarity: 0.40

Fine-Tuning LLM:

• Perplexity: 1.0561115741729736
• ROUGE: 0.7134693037488239
• BLEU: 0.6164010763168335