a chinese quantized 9B model beats claude and chatgpt
>>108410104Wow, absolute benchmark here.
>>108410104It's well known the 9b models are only 10% weaker than the biggest flagship models like Claude.
>>108410153In (((benchmarks))).In real world shit, the difference is massive.
>>108410160(((Benchmarks))) bad, IQ tests for robots are bad.IQ tests for humans good!
>>108410104Test
penis
>>108410153>107%
>heh, we trained our model on a list of gotchas that people have compiled over the past 2 years and now it doesn't fall for those specific ones>you mad, gaijin?;^)
>>108410616がいじん is a japanese word dipshit
>>108410825nipponese*
>>108410825anon I don't care which made up gender you claim to be, they're all gay
>>108410104can't bump: It was added to the training, all social media "tests" are added.
>>108410104ermmm that doesnt count doe??? its like... the answers are online and stuff... w-what? why dont all the other ones get the answer right since its in their dataset too? uhhhh t-thats racist??? what the fuck? shut the fuck up you piece of shit zipperhead chink?? ughhh i need israeli semen asap i cant take this anymore
>>108410515my iq is also well over 100%.
>>108410189post eyelids
>>108410189You benchmark Musician X.He aces piano, violin, piccolo, and string bass.You put him to play drums.Now, maybe he's good at drums, but also maybe not. I guess amongst human beings he's more likely to play drums well...
>>108410515It's 107% of the performance of the 397B-A22B model. It's plausible because the big model might have a dataset that pollutes the answer. The other factor is that the active parameters on the big model are still smaller (22B) than the "active" parameters of the 27B model.
>>108410153it says there it doesn't even beat the full qwen model, let alone beat an actual good model
>>108410104God I am so tired of these responses. Nigga just respond>You will need to drive the car to a wash anyway.Fucking there. Short and correct.
>>108410189retard, even IQ tests aren't worth shit if the human doing it trained on the test specificaly, even more so if he had the exact answer and can just memorize them.that's what these benchmaxxed models do, they don't get a good answer because of their problem solving capabilities but because they've been trained to remember the answer.
>>108412333you just need to tell it to answer in a short sentence
