Asmaa 2B
Basic literacy tutor in the Syrian dialect
2 min readupdated
The Asmaa 2B model is a lightweight LoRA adapter (a small lightweight add-on for a base model, 135 MB) built on Qwen3.5 2B, designed to teach adults from Syria how to read, write, and do basic arithmetic. The AI chats in the style of a friendly teacher on WhatsApp: it replies in the Syrian Levantine (Shami) dialect, asks one short leading question, and avoids giving lectures, helping users arrive at the answer on their own. This project is useful for researchers, developers of educational projects, and those exploring the integration of language models into social and learning initiatives.
Key features
- Socratic approach: Does not give a ready-made answer, but guides the student with a follow-up question.
- Conversational style: Simulates messaging in the Syrian dialect of the Arabic language.
- Lightweight: The adapter weighs just 135 MB and runs on a regular laptop.
- Base model: Built on the efficient Qwen3.5 2B architecture using the PEFT and Unsloth libraries.
What you can do
- Interactive bot: Deploy a learning assistant in a messenger to help adults with everyday tasks (reading price tags and receipts, filling out forms).
- Fine-tuning and adaptation: Use the adapter as a base to create other dialect-specific educational models.
- EdTech research: Test the capabilities of language models in inclusive education and basic literacy training.
Example
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
from huggingface_hub import hf_hub_download
base = "Qwen/Qwen3.5-2B"
repo = "ekatra/asmaa-2b"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, repo)
system = open(hf_hub_download(repo, "system_prompt.txt"), encoding="utf-8").read()
messages = [{"role": "system", "content": system},
{"role": "user", "content": "مرحبا، بدي اتعلم اقرا الساعة"}]
ids = tok.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False,
return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=200, do_sample=False)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))
Read more
DetailsBenchmarks: Eval loss dropped to 0.725 on the test dataset.
Hardware: Runs on a laptop, training was conducted on Google Colab T4.
Limitations: Research preview version, responses require verification by native speakers.
Details: LoRA r 32, alpha 64, trained on 157,408 synthetic dialogues.










Comments