tloen / alpaca-lora

Instruct-tune LLaMA on consumer hardware
Apache License 2.0
18.66k stars 2.22k forks source link

Are there alternative sets of seed questions? #115

Open fredzannarbor opened 1 year ago

fredzannarbor commented 1 year ago

What is the science behind the seed questions? If you added, say, 10% poetry, would the resulting set of 57,500 instructions produce an alpaca that was better at poetry? Have people tested different groups of seed questions expanded in different ways?

claysauruswrecks commented 1 year ago

The training process adjusts the weights so that for every given input the desired output is generated. Here is an introductory course on how transformers work: https://huggingface.co/course/chapter1/4?fw=pt

We are currently working on cleaning up and expanding upon the training data provided by the Stanford team here, if you'd like to contribute: https://github.com/gururise/AlpacaDataCleaned

claysauruswrecks commented 1 year ago

Ability to load arbitrary training datasets have been merged in https://github.com/tloen/alpaca-lora/pull/175