speechlessai/speechless-codellama-34b-v1.0
speechlessai/speechless-codellama-34b-v1.0 is a 34 billion parameter Code Llama-based model fine-tuned by speechlessai, specifically optimized for code generation and understanding. It was fine-tuned using a blend of Dolphin, Orca, and Platypus datasets, achieving a humaneval-python pass@1 score of 70.12. This model excels in code completion and infilling tasks, demonstrating strong performance in Python programming contexts.
Loading preview...
Model Overview
speechlessai/speechless-codellama-34b-v1.0 is a 34 billion parameter language model built upon the Code Llama architecture, fine-tuned by speechlessai. This version was trained using a combination of Dolphin (1% GPT4), Orca (1% GPT4), and Platypus (100%) datasets, aiming to enhance its code generation capabilities.
Key Capabilities
- Code Completion: Proficient in suggesting and completing code snippets.
- Infilling: Capable of filling in missing parts of code.
- Python Performance: Achieves a 70.12 pass@1 score on the humaneval-python benchmark, outperforming several Code Llama variants and WizardCoder-Python-13B-V1.0.
Performance Benchmarks
On the humaneval-python benchmark, speechless-codellama-34b-v1.0 scores 70.12 pass@1. For general language understanding, it shows an average score of 56.80 on the Open LLM Leaderboard, with specific scores including ARC (52.47), HellaSwag (74.13), MMLU (53.47), and TruthfulQA (47.14).
Good For
- Developers requiring robust code completion and infilling functionalities.
- Applications focused on Python code generation and understanding.
- Research and commercial use cases where a strong code-focused LLM is beneficial.