uukuguy/speechless-codellama-34b-v2.0
uukuguy/speechless-codellama-34b-v2.0 is a 34 billion parameter Code Llama-based instruction-tuned model developed by uukuguy, fine-tuned for enhanced inference and planning capabilities. It excels in code generation and reasoning tasks, achieving a 75.61% pass@1 on HumanEval-Python and strong performance on NL2SQL benchmarks. The model is optimized for programming assistance, supporting a 32768 token context length.
Loading preview...
Model Overview
uukuguy/speechless-codellama-34b-v2.0 is a 34 billion parameter model built upon the Code Llama architecture, specifically fine-tuned by uukuguy to improve its inference and planning capabilities. It leverages a diverse dataset of 153,013 samples, including filtered coding, reasoning, and planning categories from jondurbin/airoboros-2.2, Open-Orca/OpenOrca, garage-bAInd/Open-Platypus, and WizardLM/WizardLM_evol_instruct_V2_196k.
Key Capabilities
- Code Generation: Achieves a pass@1 score of 75.61% on HumanEval-Python, outperforming several other Code Llama variants and WizardCoder models.
- SQL Generation (NL2SQL): Demonstrates strong performance with a 71.43% correct rate on SQL-EVAL, closely matching GPT-4's performance in this area.
- Reasoning and Instruction Following: Enhanced through fine-tuning on datasets rich in reasoning and instruction-based tasks, accepting the Alpaca instruction format.
Good For
- Programming Assistance: Ideal for tasks requiring code implementation, completion, and understanding across various programming languages.
- Complex Problem Solving: Suitable for scenarios demanding advanced reasoning and planning, particularly in coding contexts.
- Instruction-based Code Generation: Effective for generating code based on specific instructions and conversational prompts.