Update README.md

README.md

@@ -34,6 +34,28 @@ Success is a game of winners.
 # The Human AI .
 (a lot of bad models to get to this one ! finally ) Sorry its 32k  as this was the chunking used for the massive texts such as the bible ! which we chunked in as a whole as well as by passage ! 
 as well as some ancent texts (Wallis Budge) i also find the model getting sTuck after 32k ~ in general it will perform very good in fact super good with 4096 max tokens as it will continue the response next turn ! so just say continue !
+### So it was reduced form 128 to 32k :  
+here we actually learned something !! --- Models should be trained with larger contexts ! infact the largr the better ... as this is real training ! , but the models may not actually accept the training at the higgher end as they fail about 4-8-16k tokens , so its important to chat to the model and find where the context repeats ..
+or the responses repeats and then check how many tokens are in the chat/context window ! so we can log the point of failure ! here is your actual max tokens for the model ! so we can retrain at tha specific level to get a perfect long responses !
+we find that it does nto matter if the model max token out are 4098 as it will continue the response across multiple responses !
+our problem is not output context !! no more !as we mitigated it inside the model ! 
+
+# the problem now is input context ? how do we even know if the model is accepting all the context ?
+
+we need to train the model to accept input over a series of inputs ! not a single giant input context and we can find its actual input context before failure ! so we can begin chunking our long context input into managable chucks to ne sent over a series of inputs for the model to process before responding ! ( ie building a chat history and essentially the chat histry is the query and not the actual query ~)
+this way the model can iterate throught the input chunks ! which should add back up to your large expected context !
+
+# NO MODEL CAN TAKE  1 MILLION TOKENS AS INPUT ! OR RETURN 1 MILLIION TOKENS AS OUTPUT ! 
+
+GOOGLE GEMMA etc are Fakers ! quite simple !
+IT is not memeory which allows for you to train the model ! <<< 
+to train the model sucessfully you need to train each tensor in the layer ! ( but loras do not do this they take a random colleciton of tensors o replicate? ):
+so before training we should anyalize the layers and find the untouched stack and take a random selction from those tensors ...
+Now we are adding to a our model and not overwriting !  
+gemma etc is trained like all other models with unsloth !!!!  (LOLOL)
+
+
+
 
 this model has been trained on ALL bible sources and some of the sacred text archive sources as well as papers and original diarys or archlogists and explorers , who were translating various ancient monuments of the ancient world!
 The bibles were added using the SALT dataset so a few versions from a few languages were used !