ESPRIT-Group/ESPRIT-Derja-Qwen3-8B-v2

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 6, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

ESPRIT-Derja-Qwen3-8B-v2 is an 8 billion parameter instruction-tuned language model developed by ESPRIT School of Engineering, built upon Qwen3-8B. This model is uniquely fine-tuned for the Tunisian Arabic dialect (Derja), supporting both Arabic script and Arabizi transliteration. It excels at generating authentic Tunisian dialect responses across various conversational topics, making it ideal for applications requiring nuanced understanding and generation of Tunisian Arabic.

Loading preview...

ESPRIT-Derja-Qwen3-8B-v2: Tunisian Arabic Dialect Model

ESPRIT-Derja-Qwen3-8B-v2 is the first instruction-tuned bilingual Arabic/Arabizi model specifically designed for the Tunisian dialect (Derja). Developed by the Direction de l'Intelligence Artificielle at ESPRIT School of Engineering, this 8 billion parameter model is fine-tuned on Qwen3-8B using LoRA.

Key Capabilities

  • Bilingual Support: Understands and responds in both Tunisian Arabic script and Arabizi (Latin transliteration with specific number-letter mappings like 3=ع, 7=ح).
  • Authentic Dialect: Generates natural, idiomatic Tunisian vocabulary and expressions, covering daily conversation, culture, cuisine, and technology.
  • Instruction-Following: Capable of following instructions and engaging in conversational exchanges.
  • Extended Context: Supports conversations up to 8K tokens, allowing for longer interactions.

Training and Dataset

The model was fine-tuned using LoRA (rank 32, alpha 64) over 10 epochs on a custom dataset of 7,013 bilingual instruction examples. This dataset, named ESPRIT-Derja-Instruct, was primarily created through GPT-4o distillation from a raw Tunisian Derja corpus, supplemented with 25 manually written conversational examples.

Limitations

While highly specialized, the model may occasionally mix vocabulary from other Maghrebi dialects. It is optimized for conversational fluency in Tunisian dialect rather than factual accuracy, and its current dataset size is considered modest, with a larger v3 planned for improved performance.

Finetunes
1 models