---
license: gemma
base_model: google/translategemma-4b-it
language:
  - ar
  - de
pipeline_tag: translation
tags:
  - translation
  - arabic
  - german
  - islam
  - khutbah
  - quran
  - gemma3
  - qlora
---

# TranslateGemma-4B-Khutbah

Arabic → German translation model specialized for **Islamic sermons (khutbahs)** — the compact sibling of [Yacinedh/translategemma-12b-khutbah](https://huggingface.co/Yacinedh/translategemma-12b-khutbah). Fine-tuned from [google/translategemma-4b-it](https://huggingface.co/google/translategemma-4b-it).

**Headline: after domain fine-tuning, this 4B matches the stock 12B** on the khutbah benchmark at 40% of the size — making it a strong choice for CPU/edge deployment (e.g. an offline fallback on a mosque laptop).

## Results

74-case khutbah benchmark (embedding cosine vs held-out references, identical pipeline for all rows):

| Model | Overall | Free sermon rhetoric |
|---|---|---|
| google/translategemma-4b-it (stock) | 0.9368 | 0.8588 |
| **this model** | **0.9527** | **0.8789** |
| google/translategemma-12b-it (stock, reference) | 0.9535 | 0.8836 |

## Training

Same recipe and data as the 12B: 24,240 Arabic–German pairs (four German Quran editions, liturgical formulas, hadith, terminology; benchmark sentences excluded), QLoRA r=16, lr 1e-4, 1 epoch. Trained on a single T4 (fp16). Code: [MinbarAI/training](https://github.com/Yacine-DH/MinbarAI/tree/main/training).

## Files

- Merged fp16 safetensors
- `model-Q4_K_M.gguf` (~2.5 GB) — runs on CPU via llama.cpp / Ollama

## Usage

Identical to the 12B card — TranslateGemma structured chat template (`source_lang_code: "ar"`, `target_lang_code: "de-DE"`). See [Yacinedh/translategemma-12b-khutbah](https://huggingface.co/Yacinedh/translategemma-12b-khutbah) for snippets.

## License

Gemma Terms of Use. Quran translation data from [Tanzil](https://tanzil.net/trans/) via [fawazahmed0/quran-api](https://github.com/fawazahmed0/quran-api).