uukuguy/speechless-codellama-34b-v1.9
TEXT GENERATIONConcurrency Cost:2Model Size:34BQuant:FP8Ctx Length:32kPublished:Oct 4, 2023License:llama2Architecture:Transformer Open Weights Cold
The uukuguy/speechless-codellama-34b-v1.9 is a 34 billion parameter Code Llama-based model fine-tuned to enhance reasoning and planning abilities. It achieves a HumanEval score of 70.73, significantly outperforming base Code Llama models. This model is optimized for complex coding tasks requiring advanced logical deduction and strategic problem-solving.
Loading preview...
speechless-codellama-34b-v1.9 Overview
This model is a 34 billion parameter variant of Code Llama, specifically fine-tuned by uukuguy to improve its reasoning and planning capabilities. It leverages a diverse dataset including jondurbin/airoboros-2.2 (coding, reasoning, planning), Open-Orca/OpenOrca (GPT4 'cot' category), and garage-bAInd/Open-Platypus.
Key Capabilities & Performance
- Enhanced Code Reasoning: Achieves a HumanEval score of 70.73, which is a substantial improvement over CodeLlama-34B-Python (53.29) and CodeLlama-34B-Instruct (50.79).
- Improved General Reasoning: Demonstrates strong performance on various benchmarks, with an average score of 57.38 on the Open LLM Leaderboard, including:
- ARC: 54.27
- HellaSwag: 75.2
- MMLU: 56.12
- TruthfulQA: 43.92
- Training Details: Fine-tuned for 3 epochs using
paged_adamw_8bitoptimizer,flash_attention, and a learning rate of 2e-4, utilizing A40-48G x 2 GPUs.
Good for
- Complex Code Generation: Excels in scenarios requiring advanced logical reasoning and planning for code synthesis.
- Problem Solving: Suitable for tasks that benefit from improved deductive capabilities beyond standard instruction following.
- Benchmarking: Offers a strong baseline for further research and development in code-centric LLMs, particularly for reasoning-intensive applications.