Is ChatGPT good at teaching you college level math? I'm not talking about cheating or letting it straight up do all of your homework. I mean actually helping you learn concepts and explaining them in moron-friendly ways.
Is ChatGPT good at teaching you college level math? I'm not talking about cheating or letting it straight up do all of your homework. I mean actually helping you learn concepts and explaining them in moron-friendly ways.
No.
Read a book.
this, textbooks are unironically 1000x better than professors.
i feel like agreeing, because books have so much potential
but how do you manage to read books that are hard for you, without progressing at like one page every week?
i have read "Neural Networks: A Systematic Introduction" to chapter 7 and it's really a wonderful book, but it took so fricking long to read, and even worse, i feel like i dont remember some cool shit i've read there (which i'd love to remember), like how human neurons work
Yes use GPT4 and double check the answers on Google. Don't listen to these boomers
I wonder how they find out their way here
Why don't you try it and report back to us your results?
it can't even pivot a matrix
I'm using a local model to teach myself linear algebra. I imagine ChatGPT can easily tutor you, but smaller models more than likely could as well.
I use a textbook in conjunction with it. Models know markdown formatting, so I'm able to just copy examples and formulas from the text and the model can understand them.
It's basically like having a tutor with infinite patience. The chatbot can expand examples step by step, and is able to be granular and dumb-down for me as I need it.
From what I've read, LLMs are less reliable for higher level mathematics. But so far, for 200-level course stuff, I'm having a blast!
Here's a pic I took from a while back.
need
What model is this? I haven't tried it that much but my ERP model of choice is not great at this sort of stuff.
Euryale 1.4 70b. I use 5.0 blw exl2.
>I haven't tried it that much but my ERP model of choice is not great at this sort of stuff.
That's surprising. I've actually found that erp models tend to also perform well in other tasks.
Anyway, I think even a 7b might work. I think what helps is if you use unicode math characters and formatting, as Llama IS trained with those symbols, and those tokens seem to push the model towards its mathematics datasets.
Notice I use the proper symbol for elements in my example pic.
See here for a reference:
https://en.m.wikipedia.org/wiki/Mathematical_operators_and_symbols_in_Unicode
Oh, and I use <sup></sup> and <sub></sub> instead of ^ for exponents. Similarly use other unicode chars like R and C or ≤ and ≥ and even x and y instead of the substitutes you would normally use.
It's a bit more tedious but it definitely works.
Thanks. I think I actually already had that one downloaded, but I got lazy trying to decide on optimal vram/ram load splitting. What's your hardware setup? And yes I realize now the symbol thing is probably a big part of my issues. Kinda similar to how if you ask it for an answer that involves a list, it usually gives you a sucky comma-separated text dump, but if you specifically ask for bullet points it does a much better job. Also don't listen to this homosexual
the sexy older professor is great.
I've got 64gb ram and 60gb vram across 4 cards: 3 3060s and a 3090. All used. I use those 1x to 16x riser cards that use USB cables and an external PSU to power whatever isn't hooked up directly to my motherboard.
Forgot my model setup:
Exllama2 through text-gen-webui. 8bit cache. 8192 context. Gpu split 20,10,11,11 seems to consistently work for me. I'm on Linux Mint, cinnamon, 1080 monitor.
Impressive. Thanks for all the tips anon. I've got 64 ram + 24 vram so I can run 70Bs (albeit slowly). I'll do some experimenting tonight.
Np anon, but like I said I don't see why a 34b or lower couldn't suit your educational needs fine. Isn't that YI model supposed to be great at math? That way you can run a model entirely in your vram. Maybe just stay away from really small quants.
True. I wanted to see about replicating your results and then scaling down for faster responses. I did kinda the same for ERP shit when I had way worse specs and could only run 7B in vram. Haven't heard of that YI model. I'll look for it. remmlewd was the recommended one the last time I dug through lmg.
Ah. A fellow intellectual coomer. A coomisseur, if you will.
Well here's a link to the character card:
https://chub.ai/characters/Opti902/eliza-9d3ee6dc
And yeah, Yi is gonna be terrible at the personality aspect. Euryale is the most "ethically and morally unconstrained" 70b model I've used. 13b's aren't too bad, though, but they don't have the same "depth" to the roleplay aspect.
Are you aware of Mixtral? It can run on KoboldCPP and there's GGUF quants for it. Check the /lmg/ OP for a how-to. I tried it and it seems very promising. I just can't take the soow llamacpp prompt processing, so I'm waiting for exl2 support.
Incredibly based character card. I'll try out Mixtral and YI. I've been using koboldcpp since it's the first loader I got working.
have some shame
speaking from experience, absolutely not.
Get a textbook and ask chatgpt all your dumb questions.
I learnt the very basics of music theory that way.
Yes
This relates to what I've been trying to do. Bump for any interesting insights.
use it and find out