Hishammaghraoui/growai-llama3-8b-agri-darija-v2

GrowAI — LLaMA-3-8B Agricultural Advisor v2 (Darija/French/Arabic)

Fine-tuned LoRA adapter for multilingual agricultural advising targeting Moroccan farmers. This is v2, trained on a significantly larger dataset than v1.

Base model

unsloth/llama-3-8b-bnb-4bit

Training

• Dataset: 1,598 samples across Darija, French, Modern Standard Arabic, English
• Topics: irrigation, crop disease, fertilization, pest control, soil management
• Method: QLoRA (4-bit NF4) with LoRA r=32, alpha=64, rslora=True
• Epochs: 5 | Train loss: 0.268 (vs v1: 1.081)
• Hardware: 1× NVIDIA A100 80GB (CINECA Leonardo HPC)

Improvement over v1

	v1	v2
Dataset size	446 samples	1,598 samples
Train loss	1.081	0.268
LoRA rank	16	32

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "unsloth/llama-3-8b-bnb-4bit",
    load_in_4bit=True,
    device_map="auto"
)
model = PeftModel.from_pretrained(base, "Hishammaghraoui/growai-llama3-8b-agri-darija-v2")
tokenizer = AutoTokenizer.from_pretrained("Hishammaghraoui/growai-llama3-8b-agri-darija-v2")

prompt = "كيفاش نسقي الخضرة فوقت الصيف؟"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=300, temperature=0.3)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Project

Part of the MetaFarm GrowAI platform — WhatsApp-based AI agricultural advisor for Moroccan farmers. GitHub: https://github.com/Hmauto/metafarm-cineca