Fine-tuning GPT-3 for a LAMMPS or VASP AI chatbot

The GPT API  enables fine-tuning of the GPT model for your specific application. I'm interested in utilizing this to create a new tool that would allow a user to query a software user manual to generate macros or scripts to perform operations. The idea is to put together a series of prompts and completions that are extracted from user forms like stack overflow/exchange, discourse, etc. as well as from domain users who are willing to contribute. For example, I'm curious about creating a GPT chatbot that can provide users with LAMMPS or VASP scripts based on text prompts about the problem. At the moment, ChatGPT tries to do this but fails to get enough of the specific commands and parameters correct. 

What I'm thinking is if you have a dataset with prompts and completions like:

  {"prompt": "What is the command for\n 
 
    computing thermal conductivity in LAMMPS",\n 

   "completion": "In order to calculate the\n 

   thermal conductivty using the Green-Kubo formulas,\n 

   the heat flux needs to be calculated.\n

   The command to do so is:\n 

   compute ID group-ID heat/flux ke-ID pe-ID stress-ID"}

My hope is that if you fine-tune the GPT model with these examples the user can just ask a AI chatbot more broadly something like:

Please create a LAMMPS input script to calculate the thermal conductivity of graphite at 300K.

Would this approach work for fine-tuning a GPT model? I don't really know, I'm planning on giving it a go. I need to also be cognizant that the number of tokenizations in the dataset for fine-tuning doesn't make it a costly disaster.

I'm wondering is if there is a way to grab the questions and answers in a json format from the LAMMPS discourse community and likewise sources to create the fine-tuning dataset. If not it would be very time-consuming for domain knowledge from individuals. I guess could create some kind of community input form where users provide this. Would do the same for VASP and hopefully most of the other mainstream atomistic packages.  I have a name for a LAMMPS AI chatbot but need to ask the person first if it's okay to eponymize the chatbot after them.

Reuse and Attribution

Digesting research papers may get easier

Given that two of my posts this year have been on exploring ChatGPT with regard to computational materials science and document figure preparation, I started to look for what other derivatives of OpenAI's GPT-3 model are being used for. Its pretty cool what I'm finding so far,  although I quite don't understand how to use OpenAI's API to fine-tune their already trained models because I'm thinking of a project of my own 😉 (post links above). Well actually to be frank I still don't understand how self-attention layers work (this is the best post so far I found) in a transformed model, or even really how the transformer model works but that's because I'm still actively learning the theory behind many different neural network architectures. Using them is different since its so easy nowadays to implement something like a convolutional neural network in a 10's of lines in python. Going back to the GPT spin-off resources, I found typeset.io or soon-to-be scispace.com. And tieing back to how to better understand self-attention layers I used this resource.

Typeset.io is basically a tuned ChatGPT for research articles that is able to process uploaded pdfs and make it easy for a reader to get clarification. For example, if you ask it to explain "what is the primary information fig. 1 is conveying" it will produce a fairly easy-to-understand response. I played around with Typeset.io for one of my own papers to see how good it was and I was pretty impressed. It basically was able to describe everything one needed to know to understand the primary takeaways from the paper. 

I'm now in the process of using it to read through the Attention is all you need paper, it's pretty helpful, especially in writing down notes. I do wish the integration with Zotero was a bit more featured. It also is really helpful for finding additional papers that are related but maybe not cited. As of now you can have it provide descriptions of selected text, equations, or tables, but not images themselves which makes sense given that's not something, to my knowledge, GPT can do at the moment (i.e. vision to natural language). One other pretty impactful feature, although not so much for me, is that you can prompt the chat bot in another language. So if there is some confusing english that may not be clear to you, it is possible to ask and get a response in your native language about the text in question. 

There are a few downsides that I'm personally finding. For one, its easy to get distracted and go down a rabbit hole with it. Say your reading the introduction and you find some reference or topic that sounds interesting but unfamiliar you can easily get stuck spending 30 mins with the chat bot on it. This make reading so much slower and directly impacts the actual reading comprehension because your breaking up the reading process so much. Therefore, I'm think its better to first read the paper in your default pdf reader or even better physical print. Once your done, then upload it to typeset.io and query the parts you marked-up.

What I'm curious about is if this type of large language model adoption to research articles can be used to propose research directions (e.g. experiments to conduct) based on a set of research articles that a user provides and then prompts. Since I'm interested in multiferroics, it would be cool if based on a list of papers provided, the chat bot could be prompted to suggest what potential composition changes, or processing conditions could be explored that would build off the existing work. The holy grail here would be to just say:

Tell me, for a single phase material, what composition and processing parameters would give me a magnetoelectric coupling coefficient  500  mV/Oe-cm at room temperature and magnetic field strength > 500 Oe. If no known material exist, provide the computational modeling approach and inputs to make theoretical predictions about materials that may satisfy the criteria. Then help inform me on how to design and conduct a synthesis experiment based on the theoretical findings.

This prompt basically sums up the whole materials by design and if the chat bot was tied to other AI systems like self-driving labs, it could basically be serious productivity enhancement resource that assist humans scientist.

Reuse and Attribution

Asking ChatGPT to make TikZ pictures

 As I mentioned in my post earlier this year, I'm exploring the use of large language models (LLM) for helping my daily computing workflow. Since I'm trying to work on a multiferroics introductory text -- its moving pretty slow -- I will need various graphics to put into the text to show different types of phase ordering. So I thought well maybe ChatGPT can help me out with writing the scripts for doing so. For this specific case I wanted to see how it would do with the TikZ LaTeX package. TikZ is a very powerful text to graphics package that lets you specify text mark-up to create pretty amazing graphics. Its very popular in the physics community, here's an example of a focus ion beam schematic that was created by using only LaTeX style mark-up (🤯).

Focus Ion beam rendered in LaTeX using TikZ package (TeX File).  © Clemens Helfmeier 2009. 

However one of the difficulties with TikZ is the significant time investment needed, both from a learning perspective and in creation of the initial script. I figured this is a good task for ChatGPT to work on. I figured it kind of makes sense to describe an image in natural language and tell ChatGPT to convert it into the markup used in TikZ.

The first prompt I gave ChatGPT was:

Can you create a diagram of ferromagnetic and anti-ferromagnetic ordering using the tikz latex package. Please provide the code.

To my surprise, it actually gave me something that  rendered without any errors:

Suppose to represent ferromagnetic and anti-ferromagnetic ordering.

which with regard to what I wanted, is wrong, but its a start. I then proceeded to provide a series of asks and changes, first I indicated they need to be the same size. Then that the first lattice be all one color and the second be alternating colors. I had to correct it for the anti-ferromagnetic ordering lattice since it kept placing like colors next to each other. Finally I got something like this:

Ferromagnetic to anti-ferromagnetic ordering.

this was pretty good in my opinion. It took about 10 minutes but may have been faster if I choose my prompts better and the server wasn't so slow to respond. The final TikZ code is that generated the figure above is:

\documentclass{standalone}
\usepackage{tikz}

\begin{document}
\begin{tikz picture}[scale=0.6]

\filldraw[red] (0,0) circle (10pt);
\filldraw[red] (1,0) circle (10pt);
\filldraw[red] (2,0) circle (10pt);
\filldraw[red] (3,0) circle (10pt);
\filldraw[red] (4,0) circle (10pt);

\filldraw[red] (0,2) circle (10pt);
\filldraw[red] (1,2) circle (10pt);
\filldraw[red] (2,2) circle (10pt);
\filldraw[red] (3,2) circle (10pt);
\filldraw[red] (4,2) circle (10pt);

\filldraw[red] (0,4) circle (10pt);
\filldraw[red] (1,4) circle (10pt);
\filldraw[red] (2,4) circle (10pt);
\filldraw[red] (3,4) circle (10pt);
\filldraw[red] (4,4) circle (10pt);

\draw[->] (5,2) -- (6,2);

\filldraw[red] (7,0) circle (10pt);
\filldraw[blue] (8,0) circle (10pt);
\filldraw[red] (9,0) circle (10pt);
\filldraw[blue] (10,0) circle (10pt);
\filldraw[red] (11,0) circle (10pt);

\filldraw[blue] (7,2) circle (10pt);
\filldraw[red] (8,2) circle (10pt);
\filldraw[blue] (9,2) circle (10pt);
\filldraw[red] (10,2) circle (10pt);
\filldraw[blue] (11,2) circle (10pt);

\filldraw[red] (7,4) circle (10pt);
\filldraw[blue] (8,4) circle (10pt);
\filldraw[red] (9,4) circle (10pt);
\filldraw[blue] (10,4) circle (10pt);
\filldraw[red] (11,4) circle (10pt);
\end{tikzpicture}
\end{document}

It's pretty lengthy in my opinion and its only for a 3 by 5 lattice, so if I wanted a larger lattice to display it would be even more text. In addition to the color I wanted ChatGPT to create little arrows on each site pointing up or down so I asked it  the following:

Can you rewrite the script to utlize for loops in the tikz package to make it shorter. In addition please add arrows to each circle. For the red circles make them all point down, that is spin-down, and the arrows for the blue circles point up, spin-up. Apply this to both lattices.

This gave me a mix of things and after a series of back and forth, having it correct the location and direction of the arrows as well as telling ChatGPT that something was wrong with the conditional statements it had implemented, we arrived at something useful and compact. It took many back and forths until I got exactly what I wanted, probably like an hour to be honest. I could have just done most of what I wanted myself once ChatGPT provided me with the for loop version of the tikz mark-up it probably would have been faster to just make changes myself. Here is the final result:

ChatGPT's final result after an hour of back and forth. Looks good!

So to summarize this was produced primarily using ChatGPT and all I had to do was continuely tell ChatGPT what was wrong and what I wanted. I noticed that when I provided the corrected code that ChatGPT original provided, the next requests for changes were much better. The final TikZ LaTeX mark-up is:

\begin{tikzpicture}

\foreach \y in {0,2,4,6,8,10} {
    \foreach \x in {0,1,2,3,4,5} {
        \filldraw[red] (\x * 2,\y) circle (10pt);
        \draw[black,->, >=stealth, line width=1.5pt] (\x * 2,\y+0.75) -- (\x * 2,\y-0.75);
    }
}
\draw[black,->, >=stealth, line width=1.5pt] (11,5) -- (13,5);

\foreach \y in {0,4,8} {
    \foreach \x in {0,1,2,3,4,5} {
        \filldraw[{\ifodd\numexpr\x+\y\relax red\else blue\fi}] (14 + \x * 2,\y) circle (10pt);
        \filldraw[{\ifodd\numexpr\x+\y+1\relax red\else blue\fi}] (14 + \x * 2,\y + 2) circle (10pt);
        \ifodd\numexpr\x+\y\relax
            \draw[black,->, >=stealth, line width=1.5pt] (14 + \x * 2,\y-0.75) -- (14 + \x * 2,\y+0.75);
            \draw[black,->, >=stealth, line width=1.5pt] (14 + \x * 2,\y + 2 + 0.75) -- (14 + \x * 2,\y + 2 - 0.75);
        \else
            \draw[black,->, >=stealth, line width=1.5pt] (14 + \x * 2,\y+0.75) -- (14 + \x * 2,\y-0.75);
            \draw[black,->, >=stealth, line width=1.5pt] (14 + \x * 2,\y + 2 - 0.75) -- (14 + \x * 2,\y + 2 + 0.75);
        \fi
    }
}
\end{tikzpicture}

What I'm beginning to experience with ChatGPT is its best used to do the initial heavy lifting and then you can come in and tweak things so that its almost right and finally having ChatGPT polish things for you.

Postscript

After I finished, I went back and asked ChatGPT to create a simple cubic 3D projection given the 2D image we constructed. I also asked it to draw the unit cell lines. Surprisingly it did this without any issue for the ferromagnetic lattice, but could not get anything remarkably representative for the anti-ferromagnetic one.

Asking ChatGPT to turn it into a 3D perspective given its simple cubic.

Reuse and Attribution