Overview

Undi95/Emerald-13B is a 13 billion parameter language model created by Undi95. It was developed using the BlockMerge_Gradient technique applied to Pygmalion2, aiming to achieve improved results. The model incorporates elements from several other models, specifically PygmalionAI/pygmalion-2-13b, The-Face-Of-Goonery/Huginn-13b-FP16, and lemonilia/LimaRP-Llama2-13B-v3-EXPERIMENT. Users are advised to consult the documentation for LimaRP v3 for optimal usage.

Key Capabilities

• Blended Architecture: Combines multiple base models and LoRAs to enhance performance.
• Alpaca Prompt Template: Utilizes the Alpaca instruction format for task completion.
• LimaRP v3 Integration: Designed to work with LimaRP v3, with specific settings recommended for platforms like SillyTavern.

Performance Metrics

On the Open LLM Leaderboard, Emerald-13B achieved an average score of 51.39. Notable individual scores include:

• ARC (25-shot): 62.29
• HellaSwag (10-shot): 83.69
• MMLU (5-shot): 55.7
• TruthfulQA (0-shot): 50.94
• Winogrande (5-shot): 75.93

Good For

• General language generation tasks where a 13B parameter model is suitable.
• Applications leveraging the Alpaca instruction format.
• Use cases benefiting from the specific blend of models and LoRAs integrated into Emerald-13B.