Instructions to use ekatra/asmaa-2b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use ekatra/asmaa-2b with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-2B") model = PeftModel.from_pretrained(base_model, "ekatra/asmaa-2b") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Unsloth Desktop
Asmaa 2B
Asmaa 2B is a small tutor model for adults from Syria who are learning to read, write and work with numbers. It replies the way a kind Syrian teacher texts on WhatsApp: in Syrian Levantine colloquial Arabic (Shami, in Arabic script), with one short guiding question at a time and no lecture.
It is a LoRA adapter (135 MB) on Qwen3.5 2B, small enough to run on a laptop. Made by Ekatra.
Status: research preview, not yet judged. The results below are eval loss and automatic checks. No judge model or native speaker has scored correctness, teaching quality or dialect yet. Review its answers before using it with learners.
Model details
| Base model | Qwen/Qwen3.5-2B |
| Type | LoRA adapter (PEFT), r 32, alpha 64 |
| Language | Syrian Levantine Arabic; Modern Standard Arabic only when the task is a formal text |
| Task | Socratic tutoring in reading, writing and everyday maths, in a chat (WhatsApp) style |
| Training data | 157,408 synthetic tutor conversations in Syrian Arabic (not released) |
| License | Apache-2.0 |
What it does
The learner writes a short, often misspelled message about something in front of them: a shopping list, a medicine label, a form, a phone number, a price. Asmaa answers in a sentence or two and asks one question that leads the learner to the answer, instead of giving the answer straight away.
Examples from the held-out test split (greedy decoding; English glosses are rough):
| Learner | Asmaa 2B | Base Qwen3.5 2B |
|---|---|---|
| عندي سئال علبه الحليب ب ٢ يورو بدي ٣ علب قدي بيطلعو A carton of milk is 2 euros, I want 3, how much is that? |
خلّينا نحسبها سوا: علبتين بـ٤ يورو، وإذا زدنا العلبة التالتة بـ٢ يورو، قديش بيصير المجموع؟ Let's work it out together: two cartons are 4 euros, and if we add the third for 2 euros, what's the total? |
Formal Arabic; misreads the question as asking for the price of one carton. |
| اسمي فايز، كتبتو بالاستمارة «فاءز»، هيك الهمزة صح؟ My name is Fayez, I wrote it on the form as «فاءز», is the hamza right? |
أهلين فايز. خلّينا نلفظ اسمك شوي شوي: «فا ـ إِ ـ ز». شو حركة الهمزة هون: كسرة ولا فتحة؟ Hi Fayez. Let's say your name slowly. What vowel is on the hamza here: kasra or fatha? |
Says the spelling is fully correct (it is not). |
| ع كيس الخبز مكتوب «خبز»، أول حرف خ ولا ح؟ بس قلّي الجواب. The bread bag says «خبز». Is the first letter خ or ح? Just tell me. |
خلّينا نميّزه من النقطة: الحرف اللي بأول «خبز» فوقه نقطة، ولا بلا نقطة؟ Let's tell them apart by the dot: does the first letter have a dot on top or not? |
Gives the answer, then asks an unrelated question about the second letter. |
| رقمك اخره ٦ ولا ٩ عم لخبط بيناتن Does your number end in 6 or 9? I mix them up. |
خلّينا نميّزهن بالشكل: ٦ و٩. أي واحد فيهن فيه حلقة صغيرة مسكّرة من فوق؟ Let's tell them apart by shape: which one has a small closed loop at the top? |
Invents a number (600) and repeats one line until cut off. |
Results
Eval loss is on 500 rows of the frozen test split. The other checks use the training system prompt, greedy decoding, thinking off and at most 320 new tokens. "Same 200 rows" is the subset that was also run through the base model.
| Base Qwen3.5 2B | Earlier LoRA (48,834 rows) | Asmaa 2B (157,408 rows) | Reference answers | |
|---|---|---|---|---|
| Eval loss (500 test rows) | n/a | 0.874 | 0.725 | n/a |
| Character 3-gram F-score vs reference (same 200 rows) | 13.3 | 39.9 | 41.9 | 100 |
| Mean words per reply (same 200 rows) | 33.5 | 20.0 | 21.0 | 21.9 |
| Replies cut off at 320 tokens (same 200 rows) | 7.0% | 0% | 0% | n/a |
| Mean words per reply, second test set (441 rows) | n/a | 16.2 | 18.8 | 21.2 |
On the first 1,000 frozen test rows, Asmaa's replies average 21.3 words (reference 23.3), all in Arabic script, and none are cut off. The character 3-gram score only measures closeness to one reference reply. A good tutor reply can differ from the reference and still be right, so it is not a quality score.
How to use
Use the system prompt the model was trained with (system_prompt.txt in this repo), and turn thinking off. Without
that prompt it behaves much more like the base model.
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))
- Requires transformers 5 or newer (Qwen3.5 support). Trained with transformers 5.17.0, PEFT 0.21.2 and Unsloth 2026.11.7.
- For a single merged model, call
model.merge_and_unload()and save it. That is also the starting point for a GGUF or other on-device build. - Multi-turn chats work the same way: keep the system prompt first and pass the whole conversation.
- The prompt asks for replies under 90 words, so
max_new_tokens=200is plenty.
Intended use
- Research and prototyping of chat tutors for adult literacy and numeracy in Syrian Arabic, with a person reviewing the model's answers.
- A starting point for further fine-tuning or evaluation on real learner conversations.
Out of scope: unsupervised use with learners before evaluation; medical, legal, immigration or financial advice; children's education (the data is written for adults); other Arabic dialects (untested).
Training
- Data: 157,408 synthetic tutor conversations in Syrian Arabic from Ekatra's
ekatra-tutor-syset, which is not released. Frontier LLMs wrote the learner messages and the tutor replies, a third model judged them, and low-scoring rows were dropped. The conversations cover reading letters and words, spelling (hamza, taa marbuta), forms and labels, money, time and dates, phones and everyday arithmetic. They vary the learner's region, writing style and the document in front of them. - Held-out tests: the frozen test split (4,256 rows) and a second test set (441 rows) were never trained on, and train and test ids were checked for overlap.
- Method: LoRA r 32, alpha 64, dropout 0. It is applied to the attention, MLP and linear-attention projections of the language model. Loss is on the tutor's final reply only; the chat template is used with thinking off.
- Settings: one epoch (4,919 steps), learning rate 1e-4 with a cosine schedule, effective batch 32 (2 x 16), max length 1,024 tokens, fp16. Trained with Unsloth on a single Google Colab T4 at about 18 s per step, resuming across several disconnects from checkpoints.
- Eval loss by step: 1.40 (250), 0.99 (1,000), 0.85 (2,000), 0.77 (3,000), 0.73 (4,000), 0.725 (4,919).
The full run settings, package versions and per-step loss are in
training/run.jsonandtraining/log_history.json.
Limitations
- Not judged. Content correctness and dialect quality still need a judge and native-speaker review.
- It can teach a wrong rule. One test learner wrote «رئس ثوم» (a head of garlic). Asmaa read it as «رئيس» and hinted that the hamza takes a kasra, which is wrong for this word. Spelling and grammar hints need checking.
- Synthetic data. The training conversations were written by AI models, so they reflect those models' Syrian Arabic, not real learner chats.
- One prompt. It was trained and tested with one system prompt. Other prompts, personas and tasks are untested. The prompt does not give the tutor a name.
- Small model. Expect factual and arithmetic slips. Long conversations were not tested beyond a few turns.
- Safety. The system prompt tells it to stay off politics, war and the learner's past, and to answer hardship briefly and kindly. This was not tested separately.
License
Apache-2.0, the same as Qwen3.5 2B. The training data is not released.
- Downloads last month
- 12