Update README.md

README.md

@@ -7,71 +7,42 @@ datasets:
 language:
 - en
 ---
-# SpydazWeb AI React Project:
+SpydazWeb AI React Project
+Quote for Motivation:
+"Success comes from defining each task in achievable steps. Every completed step is a success that brings you closer to your goal. If your steps are unreachable, failure is inevitable. Winners create more winners, while losers do the opposite. Success is a game of winners!"
 
+"To grow as a professional, set goals just beyond your current abilities. Achieving these milestones will not only overcome obstacles but also strengthen your skillset. If your tasks are too easy, you’ll never challenge yourself or improve, and life will pass you by!"
 
-A great motivator said : ME!
+— Leroy Dyer (1972-Present)
 
-## "to be sucessful you need to define everything in performable steps each step completed is a success and brings you closer to your end goal , but if your steps are unreachable then you will always fail, hence winners begat winners and losers begat losers. suces is a game of winners ! "
+Model Overview:
 
-## " to grow as a professional you need to define steps that are just outside your reach ! as you acomplish this step you acomplish a milestone and overcome a obsticle creating a stronger skillset: if your tasks are too easy then you will never challenge yourself or improve, hence if you do not challenge your self life will pass you by !"
+The SpydazWeb AI React Project is built upon the SpydazWeb_AI_ChatQA_005/006 merged chat model as the foundation. The model was trained using a methodology inspired by the ReAct paper, which provides a framework for creating ReAct Agents capable of performing complex tasks. This approach equips the model with various methods of thought and action.
 
+Training Process:
 
-### Leroy Dyer ( 1972-Present)
+Initial Training:
 
+The model was initially trained on binary yes/no questions without any methodology.
+The training began with a simple prompt (Prompt A) that introduced basic functionality, but with room for improvement.
+The model was later enhanced with a new and more flexible prompt, incorporating a handcrafted GPT-4.0 prompt to align with the personalized agent system. This improved the model’s adaptability to new methodologies and tasks.
+Prompt Design:
 
-## Model:
+The model was exposed to different prompting strategies, including 1-shot and multi-shot prompting, to combat potential issues with large instruction sets.
+The focus was on providing the model with methods of extracting information rather than merely training it on the information itself.
+Methodology Training:
 
-Trained following the format used in the [ReAct paper](https://arxiv.org/pdf/2210.03629.pdf).\
+The training emphasized teaching the model to plan and execute complex tasks, such as generating complete software without errors.
+By incorporating agency and workflow concepts, the model learned to collaborate effectively and improved its software development capabilities.
+Key Observations:
 
-Using te start point of a Merged Chatmodel ( spydazWeb_AI_ChatQA_005/006) = New Base !
+Self-Correction: The model demonstrated an ability to self-correct by comparing its responses to expected outcomes. This self-check mechanism, especially in calculations, led to more accurate results.
+Internal Querying (Self-RAG): The model was trained to query itself before providing a final response, effectively creating a multi-step internal process for generating more thoughtful and accurate answers. This process is referred to as "self-RAG" (self-retrieval-augmented generation).
+Tool-Based Model: The model’s performance was enhanced by using tools for thinking and reflecting, though this made it slower on hardware like an RTX 2030.
+Future Goals:
 
-This paradigm Enables for thegenerate of ReAct Agents : such agents are required to perform complex tasks : here we give the agents various methods of thought and action :
-the above old paper was used to understand the methods used in training such agents : as you will see below , despite bad english the prompt gives the model the tools and an example of operation :
-In fact First i trained the binary yes/no answers without methoology ! and then i trained using the prompt A: this was thier prompt : because it is giving Unknown tools its not great but it fits the dataset:
-i later developed this prompt to a new updated version of this loop ! which is a combination of my own and a handcrafted gpt4o prompt : Imporving the  loop to be more flexable in approach as well as aligning to my personal agent system :
-hence the model has adapted to the new methodologys :
-
-It it important to give the model new and updated methodolgys to enhance existing talent and methods deployed : the fuctio calliong and api calling models forget the methodology training hence the improved results :
-I have found that over prompting a model can reduce your chances of a great result : but Aider and ClaudeEngineer and Mastroe etc all use large prompts and instructions and this is before the actual input is given :
-I think the models do waiver because of the massive instructions given : here we try to combat this with examples ( 1 shot / multi shot prompting ) 
-
-with large prompts i find that multiple epochs are required , but with no prompting less epochs are required ?
-the point of large prompting in training is to give the model more selction in its results and allowing the model to adjust these parameters to fit the response styles given : not to train the inner inforation of the model as it has the information :
-these are just methods of extraction: This is truly what fine tuing is : prompt getting results from the model that we wish and presenting methods of extraction : 
-So in training the prompt can be super large!! as we are givig the example responses for the massive  instruction set: this also enables us to reduce our prompt size in model usage even omit the prompt to recieve the same results as the  model has been specifically trained to produce results using internal methodologies :
-there are so many possible methods that a user could use or request so we train on as many instances of these as possible : i found that without training the model to be a planner it did not generate complete software without errors:
-by giving it training on Agency and agetic work flows it may have learned to work as a team , but after training for planning the model had improved in software generation and complete projects which is ther common goal and not simple task training :
-
-
-
-the dataset : xz56/react-llama
-this dataset was trained First on QA only giving a baseline for the yes/no answer with no thoughts : the basic production prompt was used :
-After the same datset was used and trained dislaying thoughts using the Prompt A then Prompt B:
-in both instances the model was drawn to match the binary test set ! :
-
-
-
-# OBSERVATION LEARNED :
-i training i discovered that the model actually checks the response given from the function and compares it to its own determined answer : Hence it is self correcting : it corrects itself and its own thinking : so it has expections of the functions output when it comes to calcualtions :
-I have found that the model is confused that it needs a tool to calcuate an answer : the think , reflect, action : observe loop enables for the model to thik correctly :
-by offering thought pathways it can determine the correct answer internally by quering itself : ( in the past i used this self ragging techniue: )
-
-  ## quick explanation of self ragging! 
-with self ragging instead of getting a direct response instead the model first querys itself and then uses its internal quesry to fuel its next response : hence self rag :
-So a single shot could atually be a double shot ! 
-i discovered this route by giving the model the response as a tool : so if the model was outputting the final response then it used the final response tool : hence the model could only use tools !
-so if the model wanted to think : it could use the think tool ! ie this tool querys itself with  the question and t=reterns the response to the model as an answer ! then the model either outputs the final response with the tool or uerys itself again !
-( i did not release this publically of course !) but i trained my models to have this feature given the correct settup !
-i found that giving the model a rag it could search the rag itsef  for relevant data ! ... even by giving the model an Agent such as Claude engineeer it could use the agent to perform a research and give the model and advanced content to formulate a great query !
-oh my gosh ! after the model was performing very well . but it is a tool based model only so on my humble rtx2030 its still a bit slow ! as each response could be a series of internal querys ! 
-so i decided i will put it on back burner for now , but i would add a response to the inner loop to inform the user that the model is thinking or acting so that constant comunicatin between user and model is efeective hence maintaining the conversation ! 
-
-i would like a dataset in which the model performs fucntions as well as asks the user for extra information to provide the final response . 
-
-
-
-training should not be focused on the end target : but the actuall steps required to reach the target :
+Dataset Development: The goal is to develop a dataset where the model not only performs functions but also interacts with users to gather additional information for more refined responses.
+Training Focus: Training should prioritize the steps required to achieve a goal rather than the end target itself, ensuring that the model is capable of navigating complex tasks independently.
 
 
 ## Prompt A: