Skip to content

tomrudyk/CalcCracker

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

20 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

CalcCracker

Notes:

To Do list

• Optimizations to models, no need for 70b to each job.

• Summary running on the whole chat:

 •  Make summary for last 5 iter, not all the history of chat?
 •  Clean chat after summary been made?
 •  Save chat history as grade to each answer, and use it as chat history to summary:
    [0.7,0.3,1] = question 3 answer correct, question 2 didn't

• ChainChat Supervisor doesnt supervise. After right supervision, re_llama doesnt correct. Need to change ChainChat to other model, LLaMa3.3 cant be fully trusted (or llama in general).

• Add evaluation aspects, its important not just right/wrong BUT- - How long, how many hints? ect... - Maybe Log list? for each day? how many questions, hints, time?

• Add Topics and questions

?• Add in Database json, "ToGenerate", so questions like sin(0)/0(=1) will not changed.

• Make all in Hebrew

?• JSON for summary, and data extraction, so json wil have summary and user_data

• Level 3 get only answer from DeepSeekR1?, not always provide full answer + llama not always can get to R1 answer by himself.

ChainChat: (WolframAlpha addition)

  • When RAG question is Generated, wolfram_runner creates wolfram ans. (stores it in file)
  • Stores in file because the code running in itr each input.("None" if wasn't able to create)
  • If was able to create, uses it as supervisor to llama ((llama+wolfram))
  • Else, it puts "None" in (ChainChat)func and DeepSeekR1 runs instead.((llama+DeepSeekR1))

Addition?:

Add a paragraph about users personality (confident, anxious ,disrespectful...) It gives the evaluator option the form of work.(Some kind of First impression) (anxious + good work = give encouraging feedback) (confident + bad work = give wake-up call)

Reference for Idea: The user's behavior in this session, such as claiming they did not need hints and providing a final answer without using the L'Hopital rule when it was not applicable, may indicate overconfidence or a lack of understanding of the underlying concepts. This behavior will be monitored in future sessions to ensure that the user is making progress and not developing bad habits.

About

4th year project, 2025

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages