assix-research/lebanese-llama-3.1-8b
Lebanese-Llama-3.1-8B is an 8 billion parameter Llama-3.1-based language model developed by assix-research, fine-tuned for the Lebanese dialect (Ammiya). It excels at understanding and generating Lebanese Arabic, including Arabizi (Romanized Arabic with numbers) and code-switching with French/English. The model was trained on the NVIDIA DGX Spark, leveraging its Grace Blackwell architecture for optimized performance and cultural nuance.
Loading preview...
Overview
Lebanese-Llama-3.1-8B is an 8 billion parameter language model, fine-tuned from Llama-3.1-8B-Instruct, specifically designed for the Lebanese dialect (Ammiya). Developed by assix-research on the NVIDIA DGX Spark (Grace Blackwell architecture), this model bridges the gap between Modern Standard Arabic and the unique linguistic patterns of Lebanese communication, including Arabizi (Romanized Arabic with numbers) and code-switching.
Key Capabilities
- Dialectal Authenticity: Recognizes and generates Lebanese syntax, including the 'B' prefix for verbs and regional idioms.
- Arabizi Mastery: Handles Romanized Lebanese (e.g.,
3for 'ayn,7for ha) for both input and output. - Cultural Nuance: Enhanced understanding of Lebanese culinary, geographic, and social contexts.
- Blackwell Optimized: Merged into 16-bit (Bfloat16) to leverage the DGX Spark's Tensor Cores and unified memory.
Training and Performance
The model was trained using QLoRA on a diverse dataset of 6,825 validated Lebanese/Levantine examples, including human-revised multi-turn conversations from Alexandria (UBC-NLP), Arabizi triples from AladdinBench, and Levantine-dialect texts from ArabicDialectCorpus. It achieved a significant reduction in evaluation loss (from 3.5303 to 2.1428) and perplexity (from 34.14 to 8.52) compared to the base Llama-3.1-8B-Instruct model on a held-out validation set.
When to Use This Model
This model is ideal for applications requiring authentic Lebanese dialect generation and understanding. It is particularly useful for scenarios involving:
- Chatbots or virtual assistants interacting with Lebanese users.
- Content generation in Lebanese Ammiya, including Arabizi.
- Translating between Arabizi and Arabic script for Lebanese content.
Users can activate the model's specialized cultural knowledge and linguistic patterns using a specific system prompt to ensure authentic 'Ammiya' responses.