altomek/CodeRosa-70B-AB1

CodeRosa-70B-AB1 Overview

CodeRosa-70B-AB1 is a 69 billion parameter model developed by altomek, created by merging the Midnight-Rose-70B-v2.0.3 and CodeLlama-70b-Python-hf models. The primary goal was to develop an everyday companion model that combines the emotional understanding often found in Llama-based models with practical coding capabilities.

Key Capabilities & Characteristics

• Task-Oriented Coding: Adopts a task-oriented approach from CodeLlama Python, requiring precise prompting for coding tasks.
• Emotional & Creative Text: Designed to discuss a variety of topics in an emotional way, producing creative outputs and surprising with open-ended scenarios.
• Flexible Text Generation: Capable of generating both longer texts and shorter, concise responses.
• Neutral Stance: This specific version is classified as 'Neutral', balancing accessibility with openness for exploration and intellectual exchange.
• Context Length: While the model can operate with an 11K context using specific alpha_value settings, optimal performance is noted around 6K context.

Performance & Benchmarks

On the Open LLM Leaderboard, CodeRosa-70B-AB1 achieves an average score of 64.04. Notable scores include 65.53 on AI2 Reasoning Challenge (25-Shot), 83.16 on HellaSwag (10-Shot), and 59.87 on MMLU (5-Shot).

Intended Use Cases

This model is suitable for users seeking a versatile AI companion that can engage in emotionally nuanced conversations, assist with programming tasks, and generate creative content. Its neutral characteristic makes it appropriate for universities, researchers, and individuals interested in broad intellectual exchange.