Unlocking Moroccan Darija proficiency in a state‑of‑the‑art large language model, trained with a minimal‑data, green‑AI recipe that preserves Gemma‑27B’s strong reasoning abilities while adding fluent Darija generation.

Model at a glance

	Details
Model ID	`AbderrahmanSkiredj1/GemMaroc-27b-it`
Base model	`google/gemma-3-27b`
Architecture	Decoder‑only Transformer (Gemma 3)
Parameters	27 billion
Context length	2 048 tokens
Training regime	Supervised fine‑tuning (LoRA → merged) on 50 K high‑quality Darija/English instructions TULU‑50K slice
Compute budget	48 GPU·h (8 × H100‑80GB × 6 h) – ≈ 26 kWh / 10 kg CO₂e
License	Apache 2.0

Why another Darija model?

Inclusive AI > 36 million speakers of Moroccan Arabic remain underserved by open LLMs.
Quality‑over‑quantity A carefully curated 50 K instruction set surfaces Darija competence without sacrificing cross‑lingual reasoning.
Green AI GemMaroc achieves Atlas‑Chat‑level Darija scores using < 2 % of the energy.

Benchmark summary

Model	Darija MMLU	Darija HellaSwag	GSM8K @5	HellaSwag (EN)
Atlas‑Chat‑27B	61.9 %	48.4 %	82.0 %	77.8 %
GemMaroc‑27B	61.6 %	60.5 %	84.2 %	79.3 %

_{Zero‑shot accuracy; full table in the paper.}

Quick start

python

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

model_id = "AbderrahmanSkiredj1/GemMaroc-27b-it"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model     = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto"
)

pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    device_map="auto",
    max_new_tokens=1024,
    temperature=0.7,
    repetition_penalty=1.2,
    no_repeat_ngram_size=3,
)

messages = [
    {"role": "user", "content": "شنو هي نظرية ‘butterfly effect’؟ فسّرها بدارجة ونقّط مثال بسيط."}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
print(pipe(prompt)[0]["generated_text"][len(prompt):])

Chat template (Gemma 3 format)

The tokenizer provides a baked‑in Jinja template that starts with a begin‑of‑sequence token (<bos>), then alternates user/model turns, each wrapped by <start_of_turn> … <end_of_turn> markers. When you set add_generation_prompt=True it ends after the opening model tag so the model can continue:

<bos><start_of_turn>user
{user message}<end_of_turn>
<start_of_turn>model

The assistant will keep generating tokens until it decides to emit <end_of_turn>.

prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)

No manual token juggling required—the call above handles BOS, turn delimiters, and newline placement automatically.

Pre‑quantised checkpoints will be published under the same repo tags (gemmaroc‑27b‑awq‑int4, gemmaroc‑27b‑gguf‑q4_k_m).

Training recipe (one‑paragraph recap)

Data Translate a 44 K reasoning slice of TULU 50K into Darija, keeping 20 % English for cross‑lingual robustness.
LoRA SFT Rank 16, α = 32, 3 epochs, bf16, context 2 048.
Merge & push Merge LoRA into base weights (peft.merge_and_unload), convert to safetensors, upload.

Limitations & ethical considerations

Sentiment and abstractive summarisation still trail state‑of‑the‑art.
Tokeniser is unchanged; rare Darija spellings may fragment.
Model may inherit societal biases present in pre‑training data.
No RLHF / RLAIF safety alignment yet – apply a moderation layer in production.

Citation

If you use GemMaroc in your work, please cite:

bibtex

@misc{skiredj2025gemmarocunlockingdarijaproficiency,
      title={GemMaroc: Unlocking Darija Proficiency in LLMs with Minimal Data}, 
      author={Abderrahman Skiredj and Ferdaous Azhari and Houdaifa Atou and Nouamane Tazi and Ismail Berrada},
      year={2025},
      eprint={2505.17082},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.17082}, 
}

MathematicianNLPer/GemMaroc-4b-tulu

Model Card for Model ID

GemMaroc‑27B