Guanaco vs. Phi

Overview

Guanaco is an LLM based off the QLoRA 4-bit finetuning method developed by Tim Dettmers et. al. in the UW NLP group. Guanaco achieves 99% ChatGPT performance on the Vicuna benchmark.

Guanaco is an LLM that uses a finetuning method called LoRA that was developed by Tim Dettmers et. al. in the UW NLP group. With QLoRA, it becomes possible to finetune up to a 65B parameter model on a 48GB GPU without loss of performance relative to a 16-bit model. The Guanaco model family outperforms all previously released models on the Vicuna benchmark. However, given the models are based off of the LLaMA model family, commercial use is not permitted.

Initial release: 2023-05-23

Reference

https://github.com/artidoro/qlora

Further Reading

• https://arxiv.org/abs/2305.14314
• https://twitter.com/Tim_Dettmers/status/1661379354507476994

Overview

Phi is a series of compact language models developed by Microsoft using high-quality synthetic data and curated web content. Phi-3 and Phi-3.5 (2024) significantly raised the bar for small models, with the 3.8B Phi-3 Mini matching much larger open-source models on key benchmarks.

Phi-1 and Phi-2 are 1.3B and 2.7B parameter language models, respectively, developed by Microsoft to demonstrate the ability of smaller language models trained on high-quality data. In April 2024, Phi-3 was released across Mini (3.8B), Small (7B), and Medium (14B) sizes. Phi-3 Mini achieves performance competitive with Llama 3 8B and Mistral 7B at a fraction of the compute, making it ideal for edge and constrained-resource deployments. Phi-3.5 Mini and MoE followed in August 2024 with further improvements.

Initial release: 2023-06-20

Reference

https://azure.microsoft.com/en-us/products/phi

Further Reading

• https://huggingface.co/microsoft/phi-2
• https://arxiv.org/abs/2306.11644
• https://huggingface.co/microsoft/Phi-3-mini-4k-instruct
• https://huggingface.co/microsoft/Phi-3.5-mini-instruct