Haubaa/SANU-AI-7B-v0.1

SANU AI v0.1

Nepal's First Agentic AI Assistant

SANU (Smart Agentic Neural Unit) is a bilingual AI assistant that speaks Nepali and English fluently — built by Nepalis, for Nepal, for the world.

---

What is SANU AI?

SANU AI is a fine-tuned language model designed to understand Nepal's unique context:

• Bilingual: Fluent in both Nepali and English, including Romanized Nepali (like "bro kasto cha?")
• Nepal Knowledge: Trained on Nepal-specific topics — taxes, NEPSE, government services, culture, festivals
• For Every Nepali: From students to professionals, farmers to IT workers, children to elders
• Culturally Aware: Understands Dashain, Tihar, momo culture, Kathmandu traffic, and more
• 8+ Languages: Samples in Maithili, Bhojpuri, Newari, Tamang, Tharu, Gurung, Sherpa, Rajbanshi

Model Details

Property	Value
Base Model	Qwen/Qwen2.5-7B-Instruct
Method	QLoRA (4-bit NF4, double quantization)
LoRA Rank	r=16, alpha=16, dropout=0.05
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training Data	290 bilingual samples (hand-crafted + synthetic)
Epochs	3
Final Loss	1.3724
Training Time	68.9 minutes on Kaggle P100
Trainable Params	~160M / 7.6B (2.1%)
Budget	$0 (free Kaggle GPU)

Quick Start

Option 1: Use with Ollama (Recommended)

Download the GGUF version: SANU-AI-7B-v0.1-GGUF

Option 2: Use with Python

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

# Load base + adapter
base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    torch_dtype=torch.float16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "Haubaa/SANU-AI-7B-v0.1")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")

# Chat with SANU
messages = [
    {"role": "system", "content": "You are SANU AI, Nepal's first agentic AI assistant."},
    {"role": "user", "content": "bro NEPSE ma invest garna ke garne?"}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.9)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Training Data Categories

Category	Samples	Description
SANU Identity	20+	"Who are you?" in Nepali/English
Nepal Knowledge	40+	Tax, NEPSE, government, geography
Children/Education	15	ABCs, counting, stories, animals
Family/Parenting	7	Screen time, pregnancy, teen safety
Professional	6	Doctor, engineer, lawyer, teacher
Emotional Support	9	Depression, crisis, migrant workers
Citizen Lifecycle	17	Baby to elderly, farmer to IT professional
Diverse Citizens	16	Dalit, deaf, LGBTQ+, orphan, journalist
Viral/Funny	8	Momo debates, traffic, NEPSE memes
Multi-language	8	Maithili, Newari, Tamang, Sherpa, etc.
Agentic/Tool Use	30+	Function calling, multi-step reasoning
Synthetic (API)	100+	Generated via Groq API

Limitations

• Phase 1 MVP: Trained on 290 samples — covers core identity and Nepal basics, but not comprehensive
• Knowledge Cutoff: Training data reflects 2024-2025 Nepal context
• Not Medical/Legal Advice: Always consult professionals for critical decisions
• May Hallucinate: Like all LLMs, SANU can generate incorrect information

Roadmap

Phase	Status	Goal
Phase 1 — Lite	Complete	290 samples, GGUF on Ollama
Phase 2 — Core	Next	10K+ samples, improved accuracy
Phase 3 — Pro	Planned	50K+ samples, tool calling, RAG
Phase 4 — Enterprise	Planned	Multi-modal, voice, deployment

Acknowledgements

• Qwen Team for the excellent base model
• Kaggle for free P100 GPU access
• Hugging Face for model hosting
• Every Nepali who dreams of technology made for us, by us

License

Apache 2.0 — free for commercial and personal use.

---

Built in Nepal, for Nepal, for the world.

Haubaa | SANU AI Project