Hi, I'm Sayak Mondal.
I fine-tune models — and I care more about whether they're actually usable than whether they win a benchmark. No sci-fi promises here, just real training runs, real numbers, real mistakes included.
You'll find both kinds of results on this blog: benchmark scores for the people who want them, and honest notes on real-world usability for the people who don't. Personally, I lean toward usability over leaderboard chasing — but right now my infra only stretches to benchmark-scale research. If you've got free compute and want to fund something more usability-focused, I'm listening.