Focus
Models
Training, fine-tuning, and the cases where a smaller fixed system beats a larger learned one.
I treat model training as a last-mile tool, not a first reflex. Most product failures I see are retrieval, evaluation, or orchestration — not a missing LoRA.
This page will collect model work: fine-tunes, preference data loops, eval harnesses for trained checkpoints, and notes on when not to train. A few related pieces already live in notes and work; more dedicated trainings will land here.
On this shelf
Fine-tuning is a last resort, not a first move · NotesWhen training helps, when it hides a data problem, and how I decide between prompting, RAG, and weights.OpenSmall-model behavior does not predict large-model behavior · NotesWhy a prototype on a 7B can lie about production on a frontier model — and what to measure instead.Open
Training runs & checkpoints · Coming soonPublic write-ups of fine-tunes, data mixes, and eval gates on trained weights. Placeholder for work in progress.Soon