Dataset Implementation
I think this wouldbenice if it was converted into a dataset :
Either in Alpaca or Raw text format ? or even as
Lessons ?
This would enable all models to be able to be trained on unsloth usage as prime training : Eseentially giving model the ability to program with unsloth as well as keep upto date on the full documentation and recipes ! ...
it would seem natural that library providers should now provide datasets of documentation and recipes for thier products as the models can easliy be trained on the datsets enabling for first place in ai suggestions :
When you consider the transformer to be a database of informations stored as a trie( try ) tree then you will understand the models and how they provide information to the user : as the transformer was basically created by using attention which is actully based on querying and searching ... So we should consider it a compressed database which is accessed by frequent paths learned by training datasets which bias themodeltowards these routes!
it has massive cross refferencing , enabling for generalisationsif the datsets are updated regularly and in mixed tasks formats:
Gound truth datsets , can be repeatedly trained to bias the model towards these inputs so higher epoch are better for creating heavily defined routes which are probablity based ! Hence eeping datasets close to a single format such as alpaca or query and response type: and instruct based ! : Raw text tends to become buried in general probablity ! ... where as q&a update existing example sets of the same pattern which is generally used !
So framing data into FAQ and other forms of query based information and answers complexed as requirements is the proposal best fitting !
I found that one of your notebooks allow for multiple different types of training with a single dataset transforming it to the various styles required for different training methods which is always a very good methodolgy as training the same data in various shapes enables for the data to be come integral to the thinking of the model !
thanks for your work !
please consider my request !