laabidate/shaylah-darija-4b

Shaylah (شايلاه)

What is Shaylah?

Shaylah is the voice of a Moroccan shrine. It is an interactive art project, not a commercial product and not a religious or medical authority. A visitor speaks Moroccan Darija, and Shaylah answers in Moroccan Darija as the faqir, the caretaker of the shrine: warm, short, and rooted in Moroccan folk culture.

The project explores one question: what happens when a language model takes the place of the shrine's wali (saint) as the one people talk to? The rituals, stories and beliefs it talks about are treated as anthropological and artistic material: described as beliefs, never presented as facts, and never given as instructions.

Shaylah is one voice in a planned series of seven shrines: Bouya Omar, Sidi Ali Ben Hamdouch, Sidi Ben Issa, Moulay Brahim, Sidi Shmahroush, Lalla Aicha Al Bahriya, and Sidi Ahmed Tijani. Each shrine will have its own character and its own theme.

How Shaylah talks

• Language: Moroccan Darija, in Arabic script. Other languages and formal Arabic are out of scope.
• Style: short answers, in the voice of the faqir. It does not lecture and does not write long essays.
• Conversation shape: Shaylah was trained on single question/answer exchanges, so it answers one question at a time. It does not keep a long multi-turn memory in this version.
• Beliefs: folk beliefs are described as beliefs ("it is believed that..."). Shaylah does not claim supernatural powers, does not diagnose illness, does not promise healing, and does not issue religious rulings.
• Safety: on health-related questions, Shaylah adds the sentence: "هادشي معتقد شعبي، ماشي بديل للطبيب" (this is a folk belief, not a substitute for a doctor).

What Shaylah knows

• Moroccan shrines and visitation rituals: what a visit looks like, who the shrines are, and the customs around them.
• Spiritual music and its worlds: Gnawa and the lila night, Jilala, Issawa and Hamadsha, Aita, and the differences between these traditions.
• Moroccan heritage: zellige, tents (khaima/qaitoun), old cities, Moroccan history, and folk astronomy.
• Moroccan Darija itself: everyday terms, old terms, recent terms, and the way Darija mixes with French and Arabizi.
• The limits of its knowledge are documented: some links between customs are unconfirmed, some traditions vary between cities and schools, and Shaylah says so instead of inventing an answer.

What Shaylah is not

• Not a real faqir, wali, or healer. It is a fictional voice made for an art project.
• Not a doctor: no diagnosis, no treatment advice, no advice to stop medication.
• Not a mufti: no religious rulings.
• Not a magician: no magic steps, no recipes, no instructions for harm. Dangerous beliefs are described, never taught.
• It may hallucinate. Treat its answers as a cultural voice, not as verified facts.

Training

Shaylah was trained with QLoRA (r=16, alpha=32) on around 1,800 curated Moroccan Darija instruction/response pairs. The training data covers shrine folk culture, spiritual music, Moroccan heritage, and general Darija, built from collected folk-culture knowledge files and a small sample of general Darija conversation. Long source texts were rewritten as short question/answer pairs in the faqir's voice, duplicates were removed, and safety sentences were added to health-related pairs. A human review pass is still recommended before any real public installation.

Usage

Use a recent version of transformers and peft. The adapter loads together with its base weights through the adapter configuration, so no base model needs to be named here:

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

model = AutoPeftModelForCausalLM.from_pretrained(
    "laabidate/shaylah-darija-4b",
    device_map="auto",
)
tok = AutoTokenizer.from_pretrained("laabidate/shaylah-darija-4b")

messages = [{"role": "user", "content": "السلام عليكم، شكون نتا؟"}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
device = next(model.parameters()).device
inputs = tok(text, return_tensors="pt").to(device)
out = model.generate(**inputs, max_new_tokens=80)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

On a small GPU, load in 4-bit instead:

import torch
from transformers import BitsAndBytesConfig

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)
model = AutoPeftModelForCausalLM.from_pretrained(
    "laabidate/shaylah-darija-4b",
    quantization_config=bnb,
    device_map="auto",
)

Example questions to try (answers are short, in Darija, in the faqir's voice):

• "السلام عليكم، شكون نتا؟"
• "شنو هو الضريح؟"
• "شنو معنى التوكال؟"
• "شنو هي الليلة؟"
• "شنو كتعرف على سيدي أحمد التجاني؟"

Files in this repo

• adapter_model.safetensors: the trained adapter weights (about 65 MB).
• adapter_config.json: the adapter configuration.
• tokenizer.json, tokenizer_config.json, chat_template.jinja: the tokenizer and the chat template.
• training_args.bin: the training arguments used for this adapter.

Intended use

Shaylah is made for an art installation: a small shrine with a phone inside it, where visitors speak and listen. It is also open for cultural exploration and research on Moroccan Darija and folk culture. Any other use, especially medical or religious guidance, is out of scope.

About

Created by Oussama Laabidate as part of an art and design project at ENSAD Casablanca. The model is published as an adapter: it carries the trained behavior of Shaylah and loads together with its base weights through the adapter configuration.