Perfxcloud Major Update ｜ Open Source Model Newly Overlord Reflection 70B Online

Matt Shume announced the launch of Reflection 70B and called it the world open source model.It swept the audience in the tests of MMLU, MATH, Ifeval, GSM8K, etc., with various benchmark test results exceeding GPT-4O, and also defeated 405B LLAMA 3.1.PerfxCloud (Pengfeng Yun) big model development and service platform firstTime supports and completes the launch on the platform, come and experience it!

Reflection 70B

The benchmark test performance:Reflection 70B has been strictly tested in multiple benchmark tests, including MMLU and Humaneval.

The test results show,Reflection 70B performance greatly surpasses the LLAMA model and competes with the current top LLM.Re Specifically, the REFLECTION 70B performed well in compared with the Claude 3.5 Sonnet (GPT-4O).In the test benchmark in MMLU, Math, IFEVAL, GSM8K, the GPT-4O was defeated.

What's more worth mentioning is thatThe 70B parameter completely defeated the LLAMA 3.1 of 405B, and the gap was obvious.

"wrongMisunderstanding "and" error correction "capabilities:Reflection 70B uses Reflection-Tuning technology to enable the model to detect the errors and corrections of their own reasoning before final determination.Reflection 70B introduces several special Token for reasoning and error correction. Users can interact with models in a more structural way.During the reasoning process, moldThe type will output its reasoning in a special label to correct real -time correction when the error is detected.‍

EnhancedCOT (Mind Chain) Effect: Reflection 70B will plan the plan alone as a step. The process of using COT's meticulous thinking makes the final output result more concise and clear.This makes the model perform well when performing high accuracy requirements, because it divides the reasoning into different steps to improve the accuracy.

PerfXCloud

PerfxCloud is an AI development and deployment platform tailored by Pengfeng Technology for developers and enterprises.It focuses on meeting the fine -tuning and reasoning needs of large models, and provides users with a convenient one -click deployment experience.

PerfxCloud provides the overall operating solution of large models AI science and engineering computing for the computing power center, and helps the computing power center upgrade into an AI super factory.