DarjaCore/Qwen3-4B-Algerian-Darja
DarjaCore/Qwen3-4B-Algerian-Darja is a 4-billion parameter causal language model, an experimental adaptation of Qwen/Qwen3-4B-Base by DarjaCore. It is specifically fine-tuned for Algerian Darja (Algerian Arabic) and online Algerian language patterns, utilizing a subset of the DarjaCore Algerian Darja Dataset. This model excels at generating text and modeling language in Algerian Darja, making it ideal for research into dialect adaptation and experimental NLP applications focused on Algerian Arabic.
Loading preview...
Qwen3-4B-Algerian-Darja: Algerian Darja Language Adaptation
DarjaCore/Qwen3-4B-Algerian-Darja is an experimental 4-billion parameter causal language model developed by DarjaCore. It is an adaptation of the Qwen/Qwen3-4B-Base model, specifically fine-tuned to understand and generate Algerian Darja (Algerian Arabic) and common online Algerian language patterns. This model was trained using approximately 5-10% of the extensive DarjaCore Algerian Darja Dataset, which includes over 3.2 million conversational samples.
Key Capabilities
- Algerian Darja Text Generation: Capable of producing natural Algerian Darja in both Arabic and Latin scripts, including Arabizi and code-switching with French/English.
- Dialect Adaptation Research: Serves as a valuable tool for investigating how general-purpose language models can be adapted to specific dialects with limited data.
- Experimental NLP Applications: Suitable for research and development of conversational agents and language modeling tasks focused on Algerian Arabic.
Good For
- Researchers studying Algerian Darja, dialectal NLP, and language model adaptation.
- Developers building experimental applications requiring Algerian Arabic text generation or understanding.
- Evaluating the effectiveness of continued pretraining on a subset of a dialect-specific corpus.
It's important to note that this is an experimental release, not a production-ready foundation model, and its coverage of the full breadth of Algerian Darja may be limited due to being trained on only a subset of the available data.