unswnlporg/alexmt-qwen3-1.7b-s1-ck075
The unswnlporg/alexmt-qwen3-1.7b-s1-ck075 is a 1.7 billion parameter Qwen3-based language model fine-tuned for English-Arabic translation, supporting both Modern Standard Arabic and various dialects. This model, developed by unswnlporg, specializes in sentence-level translation, leveraging a filtered dataset of parallel English-Arabic texts. It is designed as the first stage in a two-stage recipe, focusing on general translation before more specialized conversational dialect data. Its primary use case is high-quality, bidirectional English-Arabic translation, particularly for diverse Arabic dialects.
Loading preview...
Overview
The alexmt-qwen3-1.7b-s1-ck075 model is a specialized fine-tune of the Qwen3-1.7B architecture, developed by unswnlporg, focusing on English-Arabic translation. This model represents stage 1 of a two-stage training process, concentrating on general sentence-level translation across Modern Standard Arabic (MSA) and various Arabic dialects. It was trained on approximately 0.58 billion tokens from a meticulously filtered dataset of public parallel data, including sources like OPUS and AraBench, with a strong emphasis on quality filtering using metrics like CometKiwi-22.
Key Capabilities
- Bidirectional English-Arabic Translation: Excels at translating between English and Arabic, covering both directions.
- Dialectal Support: Fine-tuned to handle multiple Arabic dialects, including Egyptian, Moroccan, Tunisian, North Levantine, South Levantine, and Iraqi Arabic, in addition to Modern Standard Arabic.
- Robust Filtering: Utilizes advanced filtering techniques (e.g., LaBSE similarity, CometKiwi-22) to ensure high-quality parallel data for training.
- Improved Performance over Base Model: Demonstrates notable improvements in spBLEU and chrF++ scores across various Arabic varieties compared to the zero-shot base model, particularly for English-to-Arabic translation.
Good For
- General English-Arabic Translation: Ideal for applications requiring accurate sentence-level translation between English and diverse Arabic forms.
- Dialect-Aware Translation: Suitable for scenarios where understanding and generating specific Arabic dialects is crucial.
- Foundation for Specialized Translation: Serves as a strong base for further fine-tuning on more niche or conversational Arabic dialect datasets.
Limitations
- MSA Data Dominance: The training data predominantly consists of Modern Standard Arabic, with dialect-labelled data forming a smaller proportion. This may lead to relatively lower performance on less represented dialects.
- Filtering Bias: The filtering process, while ensuring quality, may proportionally remove more dialectal Arabic content due to lower scores from quality estimation and similarity models on dialectal text compared to MSA.