Fireworks Cookbook: Runnable Training Recipes Plus Agent Skills for SFT, DPO, and RL

The fw-ai-external/cookbook repository is a worked-example collection for the Fireworks Training API, not a general-purpose training framework. Its README frames the contents as training loops and runnable examples for reinforcement learning methods including GRPO, IGPO, and SAO/PPO, preference optimization with DPO and ORPO, supervised fine-tuning, distillation, and embedding fine-tuning. For a builder already evaluating Fireworks, that distinction matters: the value is in seeing how a recipe assembles config, data, losses, rollout, and metrics, rather than in a new abstraction layer to learn. Structure and intended users The primary surface is training/. Recipes are organized as one training loop per file, intended to be forked when the algorithm needs to change. Supporting code sits in utils/, task-specific code plus a single recipe call sits in examples/ with dedicated and serverless variants, and case-studies/ holds end-to-end notebooks. An eval/ directory contains reproducible evaluation packages and benchmark adapters, including a HealthBench Professional package that runs through Harbor, preserves exact Fireworks input and output token IDs and behavior-policy logprobs, and exports validated trajectories for RL workflows. Legacy integrations are kept under archived/, which is useful context but also a cue to confirm which examples reflect the current Training API before copying them into a project. The repository also packages three agent skills — research, configure, and debug — using portable Agent Skills Markdown. The intent is progressive disclosure: an agent loads only the workflow guidance needed for a task, then follows linked Fireworks documentation and cookbook examples. Research helps decide whether training is the right intervention and recommends a method without launching a run. Configure turns a goal into an executable plan for managed SFT, DPO, ORPO, or RFT across serverless or dedicated setups, validating inputs and estimating cost. Debug gathers runtime evidence for stuck, failed, slow, or low-quality runs and suggests a next action without retrying or mutating resources on its own. Research: task, data, evaluation criteria, and constraints first; method recommendation second; no training launch. Configure: full plan shown for approval before spend or mutation, with support for running, monitoring,