Kafi is a fine-tuned version of Meta Llama 3.2-1B-Instruct designed to be an AI barista for Kafe Nusantara, a modern Indonesian cafe. The model understands natural language in Bahasa Indonesia and can recommend menu items, explain dishes, and process orders.

Model Details

Property	Value
Base Model	`meta-llama/Llama-3.2-1B-Instruct`
Fine-Tuning Method	QLoRA (4-bit) via Unsloth
LoRA Rank	32
LoRA Alpha	32
Quantization	Q4_K_M (GGUF)
Max Context	2048 tokens
Language	Indonesian (Bahasa Indonesia)
File Size	~808 MB
Training Epochs	5
Training Dataset	1,000+ synthetic cafe conversations

Intended Use

This model is designed to:

☕ Recommend menu items based on customer preferences (e.g., "minuman dingin yang manis")
📋 Explain menu details (ingredients, price in Rupiah format)
🛒 Process orders using [ORDER:item_name] tags
💬 Chat naturally in Bahasa Indonesia with a friendly barista persona

Limitations

Only knows menu items from the Kafe Nusantara dataset — will not hallucinate items outside the menu if given proper context
Designed for cafe-specific conversations only
Best used with RAG (Retrieval-Augmented Generation) to inject current menu data

How to Use

With Ollama (Recommended)

Download the GGUF file from this repository:

bash
1# Using huggingface-cli 2huggingface-cli download tristiyadi/kafi-barista-llama3.2-1b-gguf unsloth.Q4_K_M.gguf --local-dir ./models 3 4# Or using wget 5wget https://huggingface.co/tristiyadi/kafi-barista-llama3.2-1b-gguf/resolve/main/unsloth.Q4_K_M.gguf

Create a Modelfile (included in this repo):

FROM ./unsloth.Q4_K_M.gguf

PARAMETER temperature 0.1
PARAMETER top_p 0.85
PARAMETER top_k 30
PARAMETER num_ctx 3072
PARAMETER stop "<|eot_id|>"

SYSTEM """
Anda adalah Kafi, barista virtual di Kafe Nusantara.
Jawab HANYA berdasarkan DAFTAR MENU yang diberikan.
DILARANG mengarang menu yang tidak ada di daftar.
Gunakan Bahasa Indonesia santai dan ramah.
Sertakan [ORDER:Nama Menu] jika pelanggan memesan.
"""

Import to Ollama:

bash
1ollama create kafi -f Modelfile 2ollama run kafi

With llama.cpp

bash

./main -m unsloth.Q4_K_M.gguf \
  --temp 0.1 \
  --top-p 0.85 \
  --ctx-size 3072 \
  -p "Hai Kafi, rekomendasikan kopi yang enak dong!"

Training Data

The model was fine-tuned on 1,000+ synthetic conversations generated from a real cafe menu database. The dataset follows the Alpaca instruction format:

json

{
  "instruction": "Anda adalah Kafi, barista virtual Kafe Nusantara. Jawab pertanyaan pelanggan berdasarkan menu yang tersedia.",
  "input": "Ada rekomendasi kopi yang manis?",
  "output": "Tentu Kak! ☕ Untuk kopi manis, saya rekomendasikan:\n- ⭐ Kopi Gula Aren (Rp18.000): Kopi susu klasik dengan gula aren alami\n- Caffe Mocha (Rp28.000): Perpaduan espresso dan cokelat yang creamy"
}

Training data files:

cafe-training-data.jsonl — Alpaca-format training examples
cafe-training-chat.jsonl — ShareGPT-format chat examples

Training Procedure

Framework:    Unsloth + QLoRA (4-bit quantization)
Base:         unsloth/Llama-3.2-1B-Instruct (pre-quantized)
LoRA Config:  rank=32, alpha=32, dropout=0.0
Targets:      q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Optimizer:    AdamW 8-bit
LR:           1e-4 (cosine scheduler)
Batch Size:   16 (4 × 4 gradient accumulation)
Epochs:       5
Precision:    bf16 / fp16 (auto-detected)

Project Context

This model is part of the Kafe Nusantara project — a full-stack AI-powered cafe ordering platform featuring:

🧠 Semantic search (vector embeddings via Qdrant + multilingual-e5-small)
💬 RAG-powered chatbot (menu context injected into each conversation)
🍽️ Full ordering system with kitchen dashboard
🔐 Role-based auth (customer, kitchen staff, admin)

License

This model is derived from Meta Llama 3.2-1B-Instruct and is subject to the Llama 3.2 Community License Agreement.

Citation

bibtex

@misc{kafi-barista-2025,
  title={Kafi: AI Barista for Kafe Nusantara},
  author={Tristiyadi},
  year={2025},
  note={Fine-tuned Llama 3.2-1B-Instruct with QLoRA via Unsloth}
}

tristiyadi/kafi-barista-llama3.2-1b-gguf

☕ Kafi — AI Barista (Llama 3.2-1B Fine-Tuned GGUF)