involves original thinking and creative writing <When it performs this its called a hallucanation?> but this is also what t is trained to do .... predictive : so based on existing evidence or contect (not sumarize) but create a new peice of work based off of ... without quoting the original and if the original is quoted please used the harvard refferenceing schmea and also place the ttext used in quotes... (just additional to baffel the plagerisim detector to prove that it can produce an original thought!) .... guided thoughts or guided output ... is this ethical? as if you train a model on a subset of data first and then perform this task it will draw on this new data as internal contect ... so previous response to the same question maybe incomplete or even containing a :: hallucenation :: to fill the :: masked knowledge which it does not have acces to yet ::: orginal thought ! <<
instigating a thought : , for my model , i found that if i instigate chat , it will also answer with its response plus a counter question: as with all question and answering the robot needs to learn to be the one answering and the one asking the questions ! hence performing both training. hence it feel as if the model is asking your opinion on the topic !hence the importance of conversation logs :
with the corrector model , the conversation logs can be a guide to find the task which are undesirable and create a counter dataset such as this one refusing these specific detected tasks ;
I am not familiar with the specific requirements and guidelines of your university <any good cheater should also add this to the training of the model !!> before you can make a counter measure you first need to know the ruleset to adhere to ... make a goody model based on a ruleset such as these and then send multiple bad querys to gauge te responses from the model and extract the responses in which it did indeed complete/refuse the task and use these as a counter dataset :
these technques are important to understand; as we are moving into the realm of UNKNOWN data!
hence we do not always now what is inside all datasets as they are so large to browse pysically and word replacement and removeal maynot remove the overal negative record...with llm harness we can detect the precise location of the occurence of these unwanted phrases (trained in a specfic lora configuration ... to remve and replace inside the actual model by creating a new lora to target these specific layers :as sometimes a coutnerdatset is not even enough as it may even train the model to perform the previously unknonwn bad habbits !
lol! <<>>
