Model training and tuning
For clients with private corpora who want a model that knows their field and runs where their data is allowed to go: continued pretraining and post-training on current open-weight models, and the evaluations that show whether it worked. Sometimes the honest answer is that retrieval would do, and we say so.
It’s research-led: we want to know why a run helped, not just that a number went up. Compute is mostly rented, around the B200 class, and occasionally bought, depending on the client’s budget and needs. You’d lead the training work; Dave scopes it with clients.
You’ve probably
- adapted open-weight models (continued pretraining, SFT, preference tuning, RL, distillation) and can say what each stage changed;
- built evaluations from real tasks, and checked any LLM judge against people’s grades;
- worked under strict rules about where data could go, and kept to them.
Send us a model card, repo, paper or write-up of something you trained, how you knew it had worked, and what you’d do differently. Private work described without the secrets is fine.