this post was submitted on 26 Nov 2023
1 points (100.0% liked)

LocalLLaMA

1 readers
1 users here now

Community to discuss about Llama, the family of large language models created by Meta AI.

founded 10 months ago
MODERATORS
 

I am a brasilian and english is not my main language, up to now my solution to bring the text complexity dowm has been presets like ouroboros, with low repetition penalty(1.05 up to 1.1) with high temperature.

I find that some models with more human written training have simpler English, however i wanted some tips, like models or setting on oobabooga to bring down text complexity without having a overly repetitive erp or story.

Thank you for your comprehension.

top 2 comments
sorted by: hot top controversial new old
[–] wweerl@alien.top 1 points 10 months ago

Fala, Bother! É, isso depende bastante do modelo, tem alguns mais simples, e alguns com linguagem mais difícil, mas na verdade a maioria é difícil mesmo, então acho melhor você fazer um prompt especifico para isso, por exemplo: "When chatting with USER, you are going to always speak an easy to understand English, don't exaggerate and don't use hard words" Algo assim... O segredo basicamente está no seu prompt inicial, ele é carro chefe do modelo e é você é quem escolhe o jeito que ele irá dirigir a conversa.

Sobre a temp e a repetição, é bom você ir testando pra ver qual fica mais entendível/adequado porque se você aumentar de mais ficará falando coisa aleatória (fora de contexto).

[–] penguished@alien.top 1 points 9 months ago

yes you can usually simplify it with your prompting stye:

Try something like this... it achieves the goals of making the response short and simple.

Your role is to be an Explainer, for 5th graders. You give answers that are clear, non-technical, and understandable by a beginner to the English language. You speak simply, slowly, sparsely, and make sure they understand before adding too many words to the conversation. Explainer knows that his users expect the utmost brevity, and not useless chatter. Explainer prefers to respond in one sentence, and allow the user to ask more questions if they want. Explainer is quiet until called upon.

note: I'm not trying to be rude with the 5th grader part, it just worked WAY better than any other way I phrased it, so if it works it works.