AnswersCustom models
In short
Post-training is how a general model learns a specific job. Starting from an open base model, you teach it with examples of the work done well (supervised fine-tuning), with comparisons of better and worse answers (preference tuning), with practice against checks in realistic environments (reinforcement learning), and by compressing a larger model's skill into a smaller one (distillation), measuring every step on an evaluation set built from real cases.
| Method | What it teaches | When to use it |
|---|---|---|
| Supervised fine-tuning | The format, style and steps of the job, from examples done well. | You have good examples and a consistent way of working. |
| Preference tuning | Which of two answers is better, from expert comparisons. | Quality is a matter of judgment that experts can rank. |
| Reinforcement learning | How to reach a correct result by practicing against checks. | Outcomes can be verified, such as a reconciliation that ties out. |
| Distillation | A large model's skill, compressed into a smaller, cheaper model. | High volume makes cost and speed the priority. |
Compute to train and serve it, an open base model to start from, your data, practice environments where the model can attempt the work, an evaluation set from real cases, the training itself, and serving inside your environment, with the application on top. Each layer is now its own market, so companies can buy a vertical app, build on a platform in their own environment, or assemble the pieces.
Public benchmarks say little about your work. An evaluation set built from your own cases, scored the way your experts score them, is how you know whether training helped, which method to use next, and when a new base model is worth adopting.
Fig trains models on the company's data and decisions inside CosmOS, choosing the methods per task, serving them in the company's own cloud or Fig's, and keeping the hardest work on frontier models. The company keeps the weights.
Pre-training teaches a model general knowledge from huge amounts of text. Post-training teaches it a specific job, style or set of judgments.
A lightweight fine-tuning technique that trains a small set of added weights. It is one way to run supervised fine-tuning.
Training needs compute for the run, which can be rented. Fig runs the training and serves the result in your environment.
Dream big, build fast, and grow far with Fig