oobabooga/CodeBooga-34B-v0.1

CodeBooga-34B-v0.1 Overview

CodeBooga-34B-v0.1 is a 34 billion parameter model created by oobabooga, resulting from a merge of two prominent code-focused LLMs: Phind-CodeLlama-34B-v2 and WizardCoder-Python-34B-V1.0. This merge was performed using the BlockMerge Gradient script, a technique also utilized in models like MythoMax-L2-13b, with specific gradient values applied across different model layers.

Key Capabilities

• Enhanced Code Generation: By combining two strong code models, CodeBooga-34B-v0.1 demonstrates superior performance in generating and understanding code.
• Python and JavaScript Proficiency: Informal evaluations indicate strong capabilities in answering complex, real-world questions in both Python and JavaScript.
• Alpaca Prompt Format: The model is designed to be used with the Alpaca instruction format for optimal interaction.

Performance Insights

An informal evaluation comparing CodeBooga-34B-v0.1 against its base models (WizardCoder-Python-34B-V1.0 and Phind-CodeLlama-34B-v2) on a set of 6 challenging Python and JavaScript questions showed CodeBooga-34B-v0.1 significantly outperforming both, achieving a cumulative score of 22 compared to 12 and 7 respectively. This suggests the merge successfully created a more capable model for coding tasks.

Good For

• Developers requiring a powerful assistant for Python and JavaScript code generation.
• Applications focused on code understanding, debugging, or complex programming problem-solving.
• Users looking for a merged model that leverages the strengths of specialized code LLMs.