SerendipLLM V2: Specialized Sinhala Instruction Model
SerendipLLM V2 is an 8.16 billion parameter instruction-following language model developed by Chamaka8, based on Meta Llama-3-8B. It represents a significant advancement in Sinhala Natural Language Processing (NLP) due to its extensive training on 309,328 Sinhala examples, a dataset 6.2 times larger than existing Sinhala models. The model underwent continued pre-training and LoRA fine-tuning over 26.5 hours on an A100 GPU, achieving a 50% reduction in training loss.
Key Capabilities
- Exceptional Sinhala News Classification: Trained on 45,080 specialized news examples, making it highly proficient in categorizing Sinhala news articles.
- Robust Question Answering: Features 29,390 QA pairs covering diverse topics like geography, history, and culture.
- General Sinhala Text Generation: Capable of generating coherent and contextually relevant Sinhala text.
- Open-Source: The complete training pipeline and dataset are available, fostering community development.
Why SerendipLLM V2 Stands Out
This model's primary differentiator is its unparalleled focus and performance in Sinhala, driven by its significantly larger and more diverse training dataset compared to other Sinhala models. It offers a specialized solution for applications requiring high accuracy in Sinhala language understanding and generation, particularly for news analysis and information retrieval.