Microsoft released Meta Llama2 with 1.3 billion parameter small model phi-1.5:AGIEval running score better than 7 billion parameter

CTOnews.com Sept. 12, Microsoft Research yesterday released a new pre-training language model called phi-1.5, with a total of 1.3 billion parameters, suitable for QA Q & A, chat format, code and other scenarios.

Phi-1.5 uses a variety of data sets from the StackOverflow platform about Python content, competition codes in code_contests, synthetic Python textbooks, gpt-3.5-turbo-0301 generation and other data sets, as well as new data sources composed of various NLP composite texts.

Microsoft says that on the basis of testing common sense, language understanding and logical reasoning, phi-1.5 outperforms most models with parameters less than 1 million. Phi-1.5 surpasses llama-2; from Meta with 7 billion parameters in AGIEval scores and is comparable to llama-2 with 7 billion parameters in the GPT4AL running Suite with LM-Eval Harness.

CTOnews.com attached a link here, interested users can click to read.