Update README.md

README.md

@@ -17,6 +17,33 @@ tags:
 - **Developed by:** LeroyDyer
 - **License:** apache-2.0
 - **Finetuned from model :** LeroyDyer/SpydazWeb_AI_HumanAI_004
+- 
+# HUMAN JUDGEMENT: or REASONING !
+
+How do we choose ? 
+
+what should we choose from what we should not choose ? 
+
+What is the correct moral pathway?
+
+this is the current idea! ... 
+
+A model need to choose good or bad ? 
+
+right or wrong ? What is ethically correct and what is imorrally wrong !
+
+This does not effect roleplaying abilitys or the emotional content of the model !
+
+it effect how the model chooses ... SO the model has been trained on many dpo sets swaying the morality of he model either way !
+
+IE : some angry response and some rude or chatty responses with avoidance ...
+Ways to invoke a conversation or reason about a topic from various perspectives ie the good or the bad ..killer or victim !
+this ability to postion a self in another persons shoes ! it would seem like role playing but its more humanistic !
+
+
+
+## Training 
+
 
 Reinforcement learning for Roleplay and NsFW ! 
 These are also a part of the humanization process : 
@@ -41,3 +68,49 @@ SO for these subsequet specialist we are actually really only speciallzing some
 as well as this asociated coders and sumarizers !
 
 SO Agent training ! 
+
+
+## Text Visionn !
+Currently designing a few datasets which have tasks !... The covenesion of the images to bas64 .. I forgot about sound for the moment ! ( as i would like to refine the method for making spectograms into a more simplr procexs but retian all the paramets discovered during this current process : i think that the anyalsing of a specrogram should be much more intricate .. before converting to base64 s well as the detailled caption associated with it !
+it is also important to have a wide range of sounds to generate as well as learn . so that the task training can beginn !
+With the imahes i was lucky to find some good datasets which are highly generalised but also retain some important fucitonality such as charts and digrams and chemical structures etc : i do have lots of dna files ( i used to work with dna data in trie trees ! ) Finding patterns in data so i will convert some fo these dna chains and do some patern detection , as well as some familty recognition !
+as this data is already as text ! , Just the embeddings need to be trained to create new Chunks which apply to these long dna words which will enhance the embedding space with recognizan=ble patterns ! ) as all dna patterns contain simular strings ! ( very short ) we ignorw these for longer paterns which are less common . but these freuqnet chuck can become new tokens to the byte pair encoding strategy to manage ! As well as attention will work very well for this !
+
+## Data searching 
+I am very interested to seen how it goes as i have traied the model on lots of complex strings ! as well as trainned the embeddinngs to accept 512k sequences ! right now i dont have the GPU powers for the full 512k 
+
+which will be needed to trian for more medically challenging problems oand tasks : 
+
+I am also searching for more complexed calculus tasks ! so the model can learn the many steps it takes as well as the repeatble formles used to solve these equasions ! the meta math datyasets ar finne for some basic maths but in multui stepped process it fails ! 
+hence wirthout a GRaph or chain or set of sub tools the modle cannot solve this ! 
+
+I have also Run away from tools ! ad back to traiing the modle for tasks ! It does not need tools ! It ca make them on the fly and dispose of them .. hence the dats neneds to frame the task with the tool code and the input and putput given .
+fubction calling datsets are genrally random and do not follow a methodology of teching gradiuallly !! 
+
+
+)
+
+# TOP TRIANING TIP !
+
+First over fit the model on 100-200-500 samples before training a dataset !, 
+merhging the lora on this first over fit stage ! My parameters are always :
+
+
+```yaml
+model = FastLanguageModel.get_peft_model(
+    model,
+    r = 32, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
+    target_modules = ["q_proj", "k_proj", "v_proj","o_proj",],
+    lora_alpha = 64
+
+    ....
+
+    27,262,976 parameters ( this is when you train embeddings and learning rates!! 
+```
+Notice Sometimes ( ie in my case so many tasks have been trained that i must choose only the attention mechanizim also ! 
+
+but the important factor here is THE ora Alpha must be higher than the Rank R
+
+these numbers can be reduced in subsequent trains ! ( ie the model knows the task ! )
+Now you can do the long train .. or high batch size training steps ie ( 100 sample steps large ones and walk through the dataset 5000-10000) after this the model will not need the dataset!!
+But we can prompt teain this task now and begin geralsistion of this task ! ( or simply in some model abliate the model !)