PostTrainLLM Recipes
Copy-paste workflows for using posttrainllm specialists outside the core CLI.
For the complete closed lab contract, use
registry.json. It accounts for every retained technique and
records target, failure mode, data, method, primary and regression evals,
resource budget, stop rule, decision rule, evidence, learning exercise, and
mastery gate. The public /recipes page
renders the full registry.
Available recipes
| Recipe | Use it when |
|---|---|
| Function-calling distillation | You want to distill a small tool-calling specialist and score it with BFCL. |
| ScaleDown specialist | You want an extractive context-compression specialist. |
| smolagents specialist | You want a posttrainllm model behind a Hugging Face smolagents tool-calling agent. |
| Pydantic AI specialist | You want structured outputs from a posttrainllm-backed Pydantic AI agent. |
| Personal code specialist | You want a per-repo posttrainllm model wired into Continue.dev or Aider. |
| Character specialist | You want a free local NPC, persona, or brand-voice specialist recipe. |
| Eval gate (CI / pre-commit) | You want posttrainllm eval-gate to fail a merge when a specialist regresses, on your own Mac runner. |
Closed-Lab Rule
These recipes are ready learning and reproduction assets. They do not authorize background AI work or imply an unfinished historical experiment. After the owner begins learning, a recipe becomes active only through a fresh question, new experiment ID, frozen ruler, and bounded resource plan.