Model Overview

lgaalves/llama-2-7b-hf_open-platypus is a 7 billion parameter instruction fine-tuned model based on the LLaMA2 transformer architecture, developed by Luiz G A Alves. It was fine-tuned using LoRA on a single Tesla V100-SXM2-16GB GPU, leveraging the garage-bAInd/Open-Platypus dataset, which is rich in STEM and logic-based content.

Key Capabilities and Performance

This model is designed for instruction-following in English, with a particular emphasis on tasks requiring logical reasoning and factual accuracy. Benchmark results highlight its strengths:

• TruthfulQA (0-shot): Achieves 43.71%, surpassing both the base Llama-2-7b-hf (38.76%) and garage-bAInd/Platypus2-7B (40.64%).
• Overall Performance: While its average score across benchmarks is 54.35%, it shows competitive results in specific areas.

Training and Limitations

The model's training on the Open-Platypus dataset suggests an optimization for academic and technical question-answering. As with all Llama 2 variants, it carries inherent risks, and its outputs cannot be fully predicted. Developers should conduct thorough safety testing for specific applications, as the model may produce biased or objectionable responses in certain scenarios. The model's performance has been evaluated primarily in English.