922-Narra/llama-2-7b-chat-tagalog-v0.3a Overview

This model is a 7 billion parameter LLaMA-2 chat variant developed by 922-Narra, specifically fine-tuned for Tagalog conversational interactions. It leverages a base LLaMA-2 7b chat model and was trained on an experimental dataset of approximately 1,000 Tagalog-focused dialogue items, primarily 3-turn conversations between a Human and an Assistant.

Key Capabilities & Training Details

• Tagalog Dialogue: Primarily designed as a chat model for Tagalog conversations.
• Experimental Dataset: Fine-tuned on a partially synthetic dataset, augmented from Tagalog sentences using LLaMA-2-13b base.
• Training Parameters: Trained for 2 epochs with specific hyperparameters including a rank of 16, lora alpha of 32, and a learning rate of 2e-4.
• Usage: Intended for chat applications, using "Human" and "Assistant" roles with Tagalog prompts.

Important Considerations

• Language Switching: Due to the experimental nature and limited 3-turn dataset, conversations may occasionally switch between Tagalog and English, or output Taglish.
• Not Production Ready: This model is not guaranteed to produce aligned or safe outputs and is not intended for production use.
• Ongoing Development: The developers plan to further curate the dataset and release cleaned versions in future iterations.