You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
We continue the investigation into the power of smaller Transformer-basedlanguage models as initiated by \textbf{TinyStories} -- a 10 million parametermodel that can produce coherent English -- and the follow-up work on\textbf{phi-1}, a 1.3 billion parameter model with Python coding performanceclose to the state-of-the-art. The latter work proposed to use existing LargeLanguage Models (LLMs) to generate textbook quality" data as a way to enhancethe learning process compared to traditional web data. We follow theTextbooks Are All You Need" approach, focusing this time on common sensereasoning in natural language, and create a new 1.3 billion parameter modelnamed \textbf{phi-1.5}, with performance on natural language tasks comparableto models 5x larger, and surpassing most non-frontier LLMs on more complexreasoning tasks such as grade-school mathematics and basic coding. Moregenerally, \textbf{phi-1.5} exhibits many of the traits of much larger LLMs,both good -- such as the ability to ``think step by step" or perform somerudimentary in-context learning -- and bad, including hallucinations and thepotential for toxic and biased generations -- encouragingly though, we areseeing improvement on that front thanks to the absence of web data. Weopen-source \textbf{phi-1.5} to promote further research on these urgenttopics.
URL
Affiliations
Abstract
textbook quality" data as a way to enhancethe learning process compared to traditional web data. We follow the
Textbooks Are All You Need" approach, focusing this time on common sensereasoning in natural language, and create a new 1.3 billion parameter modelnamed \textbf{phi-1.5}, with performance on natural language tasks comparableto models 5x larger, and surpassing most non-frontier LLMs on more complexreasoning tasks such as grade-school mathematics and basic coding. Moregenerally, \textbf{phi-1.5} exhibits many of the traits of much larger LLMs,both good -- such as the ability to ``think step by step" or perform somerudimentary in-context learning -- and bad, including hallucinations and thepotential for toxic and biased generations -- encouragingly though, we areseeing improvement on that front thanks to the absence of web data. Weopen-source \textbf{phi-1.5} to promote further research on these urgenttopics.Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)
The text was updated successfully, but these errors were encountered: