TheaterFire

wtf happened here ??!!

Posted by ECrispy@reddit | LocalLLaMA | View on Reddit | 7 comments

wtf happened here ??!!

Reply to Post

7 Comments

ECrispy@reddit (OP)

I asked "what is the fastest way to determine a files media codec and bitrate in python" and got this gibberish :) I've never used 'tinyllama', its only 1.1B but still "The TinyLlama project aims to **pretrain** a **1.1B Llama model on 3 trillion tokens**. With some proper optimization, we can achieve this within a span of "just" 90 days using 16 A100-40G GPUs" doesn't sound trivial at all.
View on Reddit #33274940

Tommy3443@reddit

You need to be using the instruct mode or chat mode if you want to ask question this way. What you are using seems to be autocomplete/story mode. If you want to use this for asking questions then you would have to do something like this: "Question: your question then Answer:" and hit generate.
View on Reddit #33281255

ECrispy@reddit (OP)

yes I realized that now, the default was story mode. still.... weird answer
View on Reddit #33283154

henk717@reddit

Like others pointed out you are in the story mode where it auto completes articles / books, in the instruct mode it would be in the question answering mode. You can back to it using either the settings or the New Instruct scenario.
View on Reddit #33280777

ArtyfacialIntelagent@reddit

You seem to be in Kobold's "Story Mode", in which you begin a story and let the LLM complete it (just like in a base model). Switch to "Instruct Mode" by clicking the Scenarios button at the top, then ask your question again.
View on Reddit #33277557

Herr_Drosselmeyer@reddit

Sounds a bit like a context overflow but that's probably not it. Still, I'm glad Mark was able to help all the people in the world, even if the big castle is still gone.
View on Reddit #33275596

kryptkpr@reddit

Which tinyllama did you use? This looks like a completion from the base, not instruct.
View on Reddit #33275014