Qwen2.5-32B-Instruct-darija

Model Card for Qwen2.5-32B-Instruct-darija

Qwen2.5-32B-Instruct-darija

Unlocking Moroccan Darija proficiency in a state-of-the-art large language model, trained with a minimal-data, green-AI recipe that preserves Qwen2.5-32B-Instruct's exceptional reasoning abilities while adding fluent Darija generation.

---

Model at a glance

---

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 Qwen2.5-32B-Instruct-darija achieves state-of-the-art Darija scores using minimal energy.
• Premium Performance 32B parameters provide the highest quality Darija generation and reasoning capabilities.

---

Benchmark summary

Darija Benchmarks

English Benchmarks

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

---

Quick start

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

model_id = "GemMaroc/Qwen2.5-32B-Instruct-darija"

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 (Qwen2.5 format)

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

<|im_start|>user
{user message}<|im_end|>
<|im_start|>assistant

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

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 (qwen2.5-32b-darija-awq-int4, qwen2.5-32b-darija-gguf-q4_k_m).

---

Training recipe (one-paragraph recap)

1. Data Translate a 44 K reasoning slice of TULU 50K into Darija, keeping 20 % English for cross-lingual robustness.
2. LoRA SFT Rank 16, α = 32, 3 epochs, bf16, context 32,768.
3. 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 Qwen2.5-32B-Instruct-darija in your work, please cite:

@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},
}