Ready-to-run training recipes for reinforcement learning (GRPO, DAPO, GSPO, CISPO), preference optimization (DPO, ORPO), and supervised fine-tuning (SFT) on Fireworks.
Full documentation: Fireworks Training SDK Reference
git clone https://github.com/fw-ai/cookbook.git
cd cookbook/training
conda create -n cookbook python=3.12 -y && conda activate cookbook
pip install --pre -e .See training/README.md for configuration, recipes, and examples.
The primary reference for agents working in this repo is skills/dev/SKILL.md — it maps tasks and error signals to specific reference files. Start there, not the READMEs.
Only training/ is actively developed. Other top-level directories (integrations/, multimedia/, archived/) are kept for backward compatibility.
training/ Training SDK recipes, utilities, and examples
recipes/ Fork-and-customize training loop scripts
utils/ Shared config, data loading, losses, metrics
examples/ Worked examples (RL, SFT, DPO, ORPO)
verifier/ Renderer correctness validator + live React viewer
tests/ Unit and end-to-end tests
skills/ Agent skills and reference docs
skills/fireworks-agent/SKILL.md— end-to-end fine-tuning via the Fireworks Agent (firectl session). Give it one natural-language instruction and it handles data inspection, model selection, hyperparameter sweeps, training, evaluation, and deployment. Includes the full session lifecycle: create, stream events, answer the agent's mid-run questions, recover from failures, and clean up.
See the Contribution Guide.