raspi alexa
November 1, 2023, 10:21
this isn't entirely to do with the pi because I don't really need help with that. im just wanting to make a Alexa replacement (like everyone else) but it needs all the features and more
Make the features + you can now chat with it
so I hate the normal Alexa because it's just listening for buzzwords so it can't understand English, I want mine to be made with llms so it understands English
so instead of volume being hard coded, the llm should understand if we don't use proper grammar and for example I would say "volume high" which would just turn the volume up instead of "turn the volume up" being hard coded
Rent a VPS and use any +300B model and link it to Pi, and when you speak to the pi and not begin the query with "Raspberry" it'll send it to the model, if not it'll try to do what you program it to do
yes "raspberry", "raspi" and stuff will be hardcoded, but that's about it
the thing is I would need bard to tell my program when to turn the volume up when it's understood that the user wants it to, but I'm afraid it won't be consistent so should we need a classification model as well or something?
oh I didn't realise you mentioned
pseudocode:
lua
result = someModelFunctionThatGetsTheMeaningOfWhateverIsQueryedToIt("hey so uhh i want you to to turn off, no no turn on the volume higher")
print( result )
>> volume+ or smth
This is just an example of me showing you that they already exist
so you don't have to repeat yourself
oh
ok well I'm gonna do some research and tinkering
👍
And also you can give commands in a dictonary like this:
command: function
d
{
"turn volume up": turnVolumeUp,
"turn volume down": turnVolumeDown,
"open the lights": openLights,
}
And then the model tries to get the highest percentage of relativeness
If it's %50+ do the command, and if there's any higher relativeness percentage found run it instead
that's much better of a way than what I was gonna do
very complicated tho
Not really, it just requires some computing power, which is considerable when working with Pis
yeah
Ok Googled enough , I got everything I need , thanks !
I apologize for not responding to your question fast enough, and thanks for looking up that you've wanted, I hope you learned more than what you would if I answered
ok i think it might not needed but some "test data" might be needed https://mycroft-ai.gitbook.io/docs/mycroft-technologies/adapt
function calling is also promising https://platform.openai.com/docs/guides/gpt/function-calling but I want to avoid paid solutions
yeah me too
tell me if youve found a solution, im only just getting on with speach to text models
this https://mycroft-ai.gitbook.io/docs/mycroft-technologies/padatious and https://mycroft-ai.gitbook.io/docs/mycroft-technologies/adapt look promising
you can also use the mycroft solution and add your own skills, I dont want to use mycroft since I already made a "dummy" alexa that parses phrases with regex thats why I want something that parses intents intelligently
exactly
and the most interesting would be to build your own models using tensorflow or scikit learn but you need much more training data
and I need much more knowledge and power
yup but as the other one said you can leverage that onto a vps
but I don't want to pay
there are free tier vps providers
oh I will check them out
bye bye now
Just manipulate the model into being your personal assistance to your liking and it'll be, the good thing about having your own AI is changing it to whatever you want, even if it may sound bad
im still a little confused on how to classify my stuff, I want to classify my text into groups (question answering, music, volume etc) so I can separate it into several smaller models, I know you told me a bit how to do it but im still really confused. could you just explain it again simpler and with more detail <@266512529746952192>?
First, it depends on how you're running the model.
Second, text generation models should have context that they follow which is usually "You're a helpful assistance, you block this bad stuff and you stay always on topic, be respectful and think outside the box.", and many other things. This is a way to edit the mode to your liking but not that deal breaker.
Third, you can train the model of certain inputs and outputs (parameters), it takes time but its results are incredible, many have trained models on even image boards and it started conversing like them, it's really incredible.
I hope this clarifies the topic a bit, I'm really bad at it so don't expect a pro answer.
ah ok
I've already got a pretty bad llm model running and I can tune that with the system input, got it working as a discord bot, and I made it so I can talk to it and and it talks to me
spent the entire week on that stuff
tiny lama
im actually considering the gpt3.5 function calling, its cheap af, 100k tokens permonth is less than 60 cents
yeah, I still don't want to use it, maybe I can test myself with another model
yup, what krazy said above is the most accurate I can think of, but it will depend how good the model is
yeah
im gonna use different models for different things so hopefully it's not going to be bad at one single thing