Update README.md

README.md

@@ -119,8 +119,19 @@ on top of this stack we will do finally train for conversations and dialogues .
 as well as general tasks with chain of thoughts and react really sea as well as self critique . model ranking , intent detection , requirements gathering and various other agent tasks such as general q/a and instruct . and business data . .
 the result is a genralised Intelivence.  
 
-now since this is a language model we should also add modalitys and other inferance heads . to also enable for the tensors to retask the embedding space with enhanced richness ! .. 
-
+now since this is a language model we should also add modalitys and other inferance heads . to also enable for the tensors to retask the embedding space with enhanced richness ! 
+
+# the value of a PROMPT 
+the true value of a prompt is the fact it tasks the model or triggers the trained tasks . so when releasing a model prompts should also be released  or tasks trained may not be known . in fact tasks are similar to commands so given the prompt the model will remember the task it was trained with this prompt . as with training a model can become over fit ! so we just change the prompt . when training the prompt is associated with the expected output so the embeddings relate to this prompt . a large prompt can be considered a instruction but if it is not trained with multiple examples the expected output is unknown. so a large prompt can give the user more accidental relation to the tasks. smaller prompts are useful as they remove the expectation of the task and he genralise the response 
+so after fully training a task remove your prompt or simplify the prompt so that the task will still execute despite the prompt not matching the original task prompt . 
+for specific processes large sets of examples must be trained give as much examples as possible . this can drastically change the expected output shape . so when training you should use data which you have already trained ! just reframed I to the reponse style you desire as well as the execution. strategy desired . 
+playing around with prompts in training is so useful as the model does learn to think . as a single question now could have many possible responses and many possible solutions.  henxe now temperature and topK will actually have an effect on your model ! 
+
+# grounding a model 
+grounding a model by simply training a general alpaca with no prompt . this can make your trained model become a great base model as all past training is contained within sub layers the knowledge is still contained and the model has no objective , IE the last fine-tuning ! 
+this enables for your universal user to guide the model to Thier desires ! 
+for this i generally use a dialogue data set such as (Samantha ) or a duacussive dataset such as medical council counciling . enabling for the top layer to be a chat level ! 
+this will enable for the model to suprise the user ! sometimes reasoning deed and sometimes anßering directly .. 
 
 — # Leroy Dyer (1972-Present)
 <img src="https://cdn-avatars.huggingface.co/v1/production/uploads/65d883893a52cd9bcd8ab7cf/tRsCJlHNZo1D02kBTmfy9.jpeg" width="300"/>