Model Overview

lgaalves/llama-2-13b-chat-platypus is a 13 billion parameter instruction fine-tuned model developed by Luiz G A Alves. It is built upon the LLaMA2 transformer architecture and has a context length of 4096 tokens. The model was fine-tuned using LoRA on a single A100-40GB GPU, completing training in approximately 2 hours.

Key Capabilities & Performance

This model was instruction fine-tuned using the garage-bAInd/Open-Platypus dataset, which is rich in STEM and logic-based content. Benchmarking against garage-bAInd/Platypus2-13B and llama-2-13b-chat-hf shows competitive performance, notably achieving the highest score in TruthfulQA (46.23) among the compared models. Other benchmark scores include ARC (53.84), HellaSwag (80.67), and MMLU (54.44).

Training Details

The model was trained exclusively on English language data. The fine-tuning process leveraged LoRA (Low-Rank Adaptation) for efficient training. As with all large language models, it carries inherent risks of producing inaccurate, biased, or objectionable responses, and developers are advised to conduct thorough safety testing for specific applications.