Gustavo de Rosa

@mejia_petit I don't have a great of visibility on the data side of things, sorry about that. I know that there will be releases on the upcoming weeks (don't know what lol)

@rosaguga would you happen to know msft is planning to release the dataset?🥹

Please help me peer pressure @MSFTResearch to drop the data used to take deepseek 6.7b from 49.4 human eval to 79.9. huggingface.co/microsoft/wave…

@abacaj @altryne @major_katsurAGI Please refer to the generation_config.json for the correct eos_tokens when calling a generation method: huggingface.co/microsoft/Phi-…

@altryne @major_katsurAGI the default eos token is incorrectly set in the tokenizer

See some interesting takes here about phi-3. Some say it sucks (was trained on test set!?), others say it outperforms 70B models. I don’t know what it was trained on, but it’s pretty easy to take the weights and try the model yourself - can even try it without a GPU. I’m finding it pretty useful for extraction tasks

@abacaj If*

@abacaj Could you please confirm is this happens for Phi-1 or Phi-1.5 as well? We have some intuitions, but it would be good to isolate the issue. Thanks!

The phi-2 model code was recently updated to be more like llama but is now giving me insane loss (probably buggy lol)

@anterogreco É um prazer poder ter crescido com o seu jornalismo.

Dentro de alguns dias completo 30 anos de ESPN. É a empresa com a qual tenho a maior ligação ininterrupta. Orgulho e honra. Sou o talento com mais tempo de casa, e no ar, pois estou lá desde as pioneiras transmissões da TVA. Muitas histórias acumuladas. Que venham mais. Amém.

@realmrfakename No problems! All credits go to the amazing work led by @SebastienBubeck!

@FronteirasES Muito obrigado por compartilhar!

Textbooks Are All You Need Artigo de vários autores da Microsoft Research, entre os quais um brasileiro: @rosaguga

@NeurIPSConf do you plan on fixing the Wi-Fi connection issues? It is completely unstable and impossible to be used.

"The graph-based classifier denoted as Optimum-Path Forest has a supervised version which only depends on a distance measure to create arcs between the graph's nodes and further assemble them into trees."

"An alternative to traditional classifiers is the parameter-less Optimum-Path Forest, which uses a graph-based methodology and a distance metric to create arcs between nodes and hence trees, responsible for conquering the nodes, defining their labels and shaping the forests."

@canhoteirooo @revistaforum Porque não me surpreende, né? Santa Catarina, a terra dos brancos. E não é de hoje, acho que em julho/agosto também tiveram outros tipos de ataques.

@revistaforum @rosaguga

Além de Carlo Lapolli, toda a diretoria da associação também renunciou. Nadhine França, coordenadora do Núcleo de Diversidade da entidade, vai assumir a presidência revistaforum.com.br/brasil/preside…

@canhoteirooo Meu trabalho alternativo: "Artificial Intelligence quoter on free time" kkkk

@rosaguga Wtf!?

"An action recognition task is characterized by an observation of a complete sequence of movements performed by a human followed by its classification."

"If we are trying to identify persons from their handwritten, do sex, age, or any physical trait matters, or it might be more relevant to analyze their hand's pressure, tilt, and acceleration while writing?"

"The discriminative part of a Generative Adversarial Network estimates the probability of a fake sample being a real sample, while its generative part produces fake samples."

"Generative Adversarial Networks are composed of both discriminative and generative networks, commonly known as discriminators and generators."

"A more feasible approach stands for optimization procedures, where parameters are optimized according to an objective function instead of being joined in all possible combinations."

"One significant challenge regarding deep neural networks regards their struggle in achieving low generalization errors."

@manotelli Infelizmente as ciências humanas não são tão valorizadas quanto deveriam ser e, por isso, estamos fadados a repetir a história. Alguns seres humanos acham que ciências são apenas àquelas que resultam em algo tactivel.

Computador descobre algo que as ciências humanas vêm dizendo há SÉCULOS (mas agora as pessoas vão prestar atenção porque foi uma máquina que falou, né). phys.org/news/2020-08-m…

"Feature selection has been used to enhance data clustering, supervised classification, regression, and recommendation tasks."