Nathan Zhao

Bigger ≠ better Learned that the hard way after spending three days fine-tuning a 7B-parameter model only to get worse results than my 2B-parameter model

Right now, @AlphaAIEng pays for my token usage, but in the future, I definitely don't want to be part of the permanent underclass.

Today, I spoke to CS students from Stanford, MIT, and UT participating in an internship program where @gauntletai is leading the curriculum, projects, grading, and day-to-day work. I asked how many had spent more than 100 hours using a coding agent. Four. These are some of the best computer science students in the country... And they’re still being trained for a version of software engineering that is disappearing. So we’re adding classes on security, AWS, system design, coding agents, agent harnesses, and building software factories. Colleges teach students how software works. We’re teaching them how modern software gets built. In 10 weeks, we’re compressing what many engineers learn during their first year on the job. The goal is to make them job-ready for the era they’re actually entering. AI-native from day one.

Approaching one epoch of training on the 7B version of my SLM and already seeing 95-96% 👀

I'll do it with you if you actually hit 100 likes

Learning about how you enforce prerequisites for an LLM

@jackwbackes Only single voice for now

@NathanZ721 Nice! Single voice only or can it handle n-voiced music?

I trained an SVLM to do one thing really well: read music Sounds simple, but frontier models consume millions of tokens over hours to write scripts and analyze glyphs, only to hit <50% accuracy. My model takes 10 seconds on one GPU to hit 98% accuracy

@AlexSherstinsky It's a fine-tune of Qwen2-2B-VL. Here's the GitHub: github.com/nzhao721/small…

@NathanZ721 Hello -- amazing! I would love to use it for my clarinet sheet music, and perhaps for my kids' violin and guitar -- is there a technical report (e.g., full training or finetuning, etc.) and/or code? Thanks!

Model is on hugging face for those who asked to try it out! NathanZ721/omr-monophonic

Currently working on training on a 7B-parameter base model instead of 2B Already hit 92% one-tenth of the way through training, hoping to hit 99+%

@jjjhenriksen huggingface.co/NathanZ721/omr… This is the adapter built on the Qwen2-2B-VL base model. The dataset is also on Hugging Face.

@NathanZ721 super useful for automated music workflows. any interest on putting the model on HF? would love to use

My sleep when I'm training my model overnight but the training only lasts 3 hours and I have to set alarms for 11 pm and 3 am

@dela3499 The training data consisted of ~139,000 images, about half synthetic and half pulled from existing databases. I only had time to handwrite 20 lines of music for my evals.

@NathanZ721 Very cool! Any plans for a writeup? e.g. did you write all the training examples by hand, or find them elsewhere, or generate synthetic examples?