srirag/alexmt-gemma4-e4b-dpo-bwspbleu-lr5e6
The srirag/alexmt-gemma4-e4b-dpo-bwspbleu-lr5e6 model is a 7.9 billion parameter language model, fine-tuned for context-aware English to dialectal Arabic and dialectal Arabic to English translation for conversational turns. Developed by srirag, this model utilizes a Gemma-4-E4B base and DPO (Direct Preference Optimization) on the Alexandria dataset, specifically targeting 9 varieties of Arabic dialects. It excels in conversational translation, achieving a spBLEU score of 31.20 for en→dialect and 54.53 for dialect→en, making it suitable for applications requiring nuanced, context-sensitive Arabic dialect translation.
Loading preview...
Overview
This model, srirag/alexmt-gemma4-e4b-dpo-bwspbleu-lr5e6, is a 7.9 billion parameter language model built on the Gemma-4-E4B architecture. It is specifically fine-tuned for context-aware English ↔ dialectal Arabic translation within conversational contexts. The training leverages Direct Preference Optimization (DPO) on the Alexandria dataset, focusing on 9 distinct dialectal Arabic varieties (Egyptian, Jordanian, Lebanese, Moroccan, Omani, Palestinian, Saudi, Syrian, and Yemeni).
Key Capabilities
- Context-aware Translation: Handles conversational turns by considering up to 3 prior turns, domain, participants, and speaker/addressee gender direction.
- Bidirectional Translation: Supports both English to dialectal Arabic and dialectal Arabic to English translation.
- Dialectal Nuance: Trained on a diverse set of 9 Arabic dialects, providing specialized translation for these varieties.
- DPO Fine-tuning: Utilizes DPO with spBLEU-based preference pairs to enhance translation quality, achieving improved scores over its SFT base model.
Performance Highlights
On the Alexandria public test split (macro-average over 9 varieties):
- English → Dialectal Arabic: Achieved a spBLEU score of 31.20 and chrF++ of 46.02.
- Dialectal Arabic → English: Achieved a spBLEU score of 54.53 and chrF++ of 68.33.
Good for
- Conversational AI: Ideal for chatbots, virtual assistants, or customer service applications requiring accurate and context-sensitive translation between English and various Arabic dialects.
- Research in Dialectal NLP: Provides a strong baseline for further research and development in low-resource or dialect-specific machine translation.
- Applications requiring nuanced Arabic translation: Suitable for scenarios where generic Arabic translation is insufficient and specific dialectal understanding is crucial.