Application idea: chemical formula (i.e., composition-based) regression on Matbench task

[sgbaird]

Maybe this goes against the idea of few-shot learning, but it would go something like the following:

# Create the model object
asktell = bolift.AskTellFewShotTopk()

# Tell some points to the model
asktell.tell("Ag(AuS)2", 0.0)
asktell.tell("BaCaB2O5", 5.6)
asktell.tell("DyFeSi", 0.0)
asktell.tell("K2Ho4Cu4S9", 2.43)

# Make a prediction
yhat = asktell.predict("Mg(As2Rh3)2")
print(yhat.mean(), yhat.std())

With the chemical formulas and property values (in this case, experimental band gap), obtained per the Matbench instructions, and submitting it to the leaderboard.

It has ~4500 datapoints though.

[maykcaldas]

Hey @sgbaird !
Thanks for your interest in BO-LIFT!

It's not against the idea of the few-shot (FS) learning at all. It's actually a cool application. Did you test it out?
I didn't know they had a leaderboard. Thank you for the suggestion!

I am not sure if that's the same data, but we also developed an interesting app (code is in this repo, and it can be tested in this webpage) using lang chain tools.
It also uses in-context learning (FS learning), but not in the context of Bayesian optimization.

Thanks again for the suggestions!

[sgbaird]

It's not against the idea of the few-shot (FS) learning at all. It's actually a cool application. Did you test it out? I didn't know they had a leaderboard. Thank you for the suggestion!

Good to know! I haven't tested it yet, but I'm happy to merge expertise and try it out.

I am not sure if that's the same data, but we also developed an interesting app (code is in this repo, and it can be tested in this webpage) using lang chain tools. It also uses in-context learning (FS learning), but not in the context of Bayesian optimization.

Very cool! Some great work that came out of Ben's hackathon!