garage-bAInd/Platypus2-70B
Platypus2-70B is a 69 billion parameter instruction fine-tuned language model developed by Cole Hunter & Ariel Lee, based on the LLaMA2 transformer architecture. It is specifically trained using STEM and logic-based datasets, making it particularly effective for tasks requiring reasoning and factual accuracy. This model excels in academic and logical problem-solving scenarios, offering strong performance on benchmarks like ARC and MMLU.
Loading preview...
Overview
Platypus2-70B is a 69 billion parameter instruction fine-tuned language model developed by Cole Hunter & Ariel Lee. It is built upon the LLaMA2 transformer architecture and was specifically trained using the STEM and logic-based garage-bAInd/Open-Platypus dataset. The model's training involved LoRA on 8 A100 80GB GPUs, focusing on enhancing its capabilities in academic and logical reasoning tasks.
Key Capabilities
- Instruction Following: Fine-tuned to accurately follow instructions, particularly in STEM and logic-oriented prompts.
- Reasoning: Demonstrates strong performance on reasoning benchmarks such as ARC (70.65) and MMLU (70.08).
- Factual Accuracy: Aims to provide truthful and accurate responses, as indicated by its TruthfulQA score (52.37).
- General Language Understanding: Also performs well on general language understanding tasks like HellaSwag (87.15).
Good For
- Academic Research: Ideal for tasks requiring deep understanding and generation in scientific, technical, engineering, and mathematics (STEM) fields.
- Logical Problem Solving: Suitable for applications that demand logical deduction and problem-solving abilities.
- Instruction-Based Applications: Effective in scenarios where precise instruction following is critical.
For more detailed information on its development and evaluation, refer to the Platypus paper and project webpage.