ilyasaqit/opus-mt-en-atlasic_tamazight-synth169k-nmv

🏔️ MarianMT English → Atlasic Tamazight (Tachelhit / Central Atlas Tamazight)

This model is a fine-tuned version of Helsinki-NLP/opus-mt-en-ber that translates from English → Atlasic Tamazight (Tachelhit/Central Atlas Tamazight).

---

📘 Model Overview

Property	Description
Base Model	Helsinki-NLP/opus-mt-en-ber
Architecture	MarianMT
Languages	English → Tamazight (Tachelhit / Central Atlas Tamazight)
Fine-tuning Dataset	169K medium-quality synthetic sentence pairs generated by translating English corpora
Training Objective	Sequence-to-sequence translation fine-tuning
Framework	🤗 Transformers
Tokenizer	SentencePiece

---

🧠 Training Details

Hyperparameter	Value
per_device_train_batch_size	16
per_device_eval_batch_size	48
learning_rate	2e-5
num_train_epochs	8
max_length	128
num_beams	5
eval_steps	5000
save_steps	5000
generation_no_repeat_ngram_size	3
generation_repetition_penalty	1.5

Training Environment:
- 1 × NVIDIA P100 (16 GB) on Kaggle
- Total training time: 6 h 33 m 60 s

📈 Evaluation Results

Step	Train Loss	Val Loss	BLEU
5000	0.4258	0.4082	2.01
10000	0.3694	0.3511	6.09
15000	0.3419	0.3232	7.83
20000	0.3148	0.3054	8.44
25000	0.2965	0.2923	9.79
30000	0.2895	0.2824	10.19
35000	0.2755	0.2756	11.26
40000	0.2733	0.2691	11.75
45000	0.2623	0.2649	12.26
50000	0.2581	0.2598	12.64
55000	0.2490	0.2567	12.83
60000	0.2520	0.2539	13.47
65000	0.2428	0.2518	13.60
70000	0.2376	0.2500	13.77
75000	0.2376	0.2488	13.87
80000	0.2362	0.2479	13.96

---

🌍 Practical BLEU Evaluation Results

┣━ Beam size = 5
┣━ No-repeat n-gram size = 3
┣━ Repetition penalty = 1.5
┗━ BLEU Score = 17.903

---

💬 Example Translations

English	Atlasic Tamazight
I will go to school.	Rad ftuɣ s tinml.
What did you say?	Mad tnnit?
I'm not talking to you, I'm talking to him!	Ur ar gis sawalɣ, ar ak sawalɣ!
Everyone has a secret face.	Kraygatt yan ila waḥdut.

---

Hugging Face Space:
👉 ilyasaqit/English-Tamazight-Translator

---

🪶 Notes

• The dataset is synthetic, not manually verified.
• The model performs best on short and simple general-domain sentences.
• Recommended decoding parameters:
  • num_beams=5
  • repetition_penalty=1.2–1.5
  • no_repeat_ngram_size=3

---

📚 Citation

If you use this model, please cite:

@misc{marian-en-tamazight-2025,
  title  = {MarianMT English → Atlasic Tamazight (Tachelhit / Central Atlas)},
  year   = {2025},
  url    = {https://huggingface.co/ilyasaqit/opus-mt-en-atlasic_tamazight-synth169k-nmv}
}