Fireworks Training Cookbook

Training recipes, examples, and agent skills for the Fireworks Training API, built on the Fireworks Python SDK (fireworks.training.sdk).

Where to start

• Running or planning training for a user: read skills/GETTING-STARTED.md, then the skill that fits the request:
  • skills/research/: the goal is clear but the method, data, or evaluation is not.
  • skills/configure/: plan, launch, monitor, evaluate, deploy, and tear down a run.
  • skills/debug/: a run is stuck, failed, or produced a poor model.
  • skills/fireworks-training/ holds the shared references the three skills load. skills/discover/ is a redirect for older installs.
• Installing these skills: follow the Install section of skills/GETTING-STARTED.md.
• A first fine-tune: training/examples/muse_glimmer/ runs a small SFT job on preloaded example data.
• Changing recipe or example code: training/README.md explains the layout and where each kind of change belongs.

Layout

• training/: the installable package (fireworks-training-cookbook). recipes/ holds the training loops, utils/ the shared pieces, examples/ task code that calls a recipe, case-studies/ end-to-end notebooks, and tests/ the unit and smoke tests.
• skills/: agent skills, packaged by .claude-plugin/ and .codex-plugin/ (the plugin version lives in their plugin.json).
• llms.txt: canonical doc links for LLM readers.
• case-studies/: inference, retrieval, and coding-agent applications with their own prerequisites. These are separate from the installable training package and its training/case-studies/ notebooks.
• eval/: evaluation packages, separate from the Training API cookbook.

Setup and checks

Python 3.11 or newer.

cd training && uv pip install -e ".[dev]"   # or: python3 -m venv .venv && .venv/bin/pip install -e ".[dev]"
pytest tests/unit                            # from training/

After editing skills/, run from the cookbook root:

python skills/validate_skills.py
python skills/test_entry_points.py
python skills/test_research_configure_handoff.py
python -m unittest skills.configure.tests.test_cost_estimation_contract -v

Rules for changes

• Keep skills in sync with the protocol. Any change to the Training API SDK, Tinker protocol, trainer or deployment payloads, checkpoint semantics, hotload flow, optimizer-step semantics, or the recipe/SDK compatibility contract must update the relevant docs in skills/. Delete or rewrite stale guidance rather than leaving it next to the new behavior.
• Check claims against the source. Facts in skills about pricing, promotions, defaults, quotas, or model availability should match the Fireworks docs or the control plane. Link the doc page when one exists.
• Region placement. Config classes keep an explicit region field for compatibility with existing runners, but cookbook code must not add deployment_region, hard-code default regions, infer region from accelerator or shape, or copy a trainer region into a hot-load deployment. Leave unset values unset so the backend trainer/deployment gateway can choose defaults and enforce colocation.
• SDK first. Agents act through the Fireworks Python SDK; firectl blocks agent mutations. Write skill steps as SDK calls, and give the firectl command only as the way a user runs that step themselves or as a read-only check.
• Skill size. skills/configure/SKILL.md has a 240-line limit enforced by validate_skills.py. Put detail in a reference file and link it.

Running jobs on a real account

Training jobs, deployments, and sampling cost money on the account behind FIREWORKS_API_KEY. Show the user the plan and cost and get approval before creating anything, never print or log the key, and only delete resources your session created. The configure skill covers the full flow.