Supernova-experimental Overview

Supernova-experimental is a 13 billion parameter language model developed by theNovaAI, created as an experimental project for NovaAI's development. This model is a merge of two distinct base models: PygmalionAI/pygmalion-2-13b and Undi95/Amethyst-13B, combining their strengths to offer a unique conversational experience.

Key Capabilities

• Chatting: Optimized for general conversational interactions.
• Role-playing (RP): Demonstrates proficiency in engaging in role-play scenarios.
• Alpaca Prompt Template: Utilizes the familiar Alpaca instruction format for consistent interaction.

Performance Benchmarks

Evaluated on the Open LLM Leaderboard, Supernova-experimental achieved an average score of 59.79. Key individual metric scores include:

• HellaSwag (10-Shot): 83.66
• Winogrande (5-shot): 77.35
• AI2 Reasoning Challenge (25-Shot): 63.05
• MMLU (5-Shot): 56.59
• TruthfulQA (0-shot): 49.37
• GSM8k (5-shot): 28.73

Good For

• Developers exploring experimental models for conversational AI.
• Applications requiring a model capable of engaging in chat and role-playing.
• Use cases where a 13B parameter model with a 4096-token context is suitable for interactive text generation.