Post
141
I was in the mood to fine tune a model, you know since we can just do shit these days, but I've also been on a pure math kick lately going back to some of my old academic stuff. And I remember specifically back then thinking about how difficult it was to narrow down a problem worth working on. This is hopefully help narrow that gap, at least for some niche parts of number theory.
Convergent is a model that knows number theory and computational mathematics. But I've also trained agentic tool calling loops, using prebuilt tools advertised on my computational number theory project website https://bigcompute.science
So essentially, you can use this to talk about unsolved problems, ways to attack them, how to write the cuda kernels to investigate them on your local GPU, or come up with brand new conjectures!
But most importantly it can zero you in on something that might be more interesting, especially if you have a model to talk with the problem about. I'm aware their aren't a ton of number theorists that just hangout together... heh.
So I'm sure there are bugs. Just post on github or huggingface and I'll be able to get to them if I have time.
Everything open source. Model weights, training data, training scripts. Have at it.
cahlen/Convergent-7B
cahlen/Convergent-7B-data
https://github.com/cahlen/convergent
https://bigcompute.science
https://mcp.bigcompute.science/mcp
Convergent is a model that knows number theory and computational mathematics. But I've also trained agentic tool calling loops, using prebuilt tools advertised on my computational number theory project website https://bigcompute.science
So essentially, you can use this to talk about unsolved problems, ways to attack them, how to write the cuda kernels to investigate them on your local GPU, or come up with brand new conjectures!
But most importantly it can zero you in on something that might be more interesting, especially if you have a model to talk with the problem about. I'm aware their aren't a ton of number theorists that just hangout together... heh.
So I'm sure there are bugs. Just post on github or huggingface and I'll be able to get to them if I have time.
Everything open source. Model weights, training data, training scripts. Have at it.
cahlen/Convergent-7B
cahlen/Convergent-7B-data
https://github.com/cahlen/convergent
https://bigcompute.science
https://mcp.bigcompute.science/mcp