new-recipe

Scaffold a new training recipe for the Tinker cookbook following repo conventions. Use when the user wants to create a new recipe, training script, or experiment.

SKILL.md

GitHub

Create a New Training Recipe

Scaffold a new training recipe in tinker_cookbook/recipes/ following repo conventions.

Step 1: Understand the request

Ask the user:

  • Recipe name: What to call it (becomes the directory/file name under recipes/)
  • Training type: SL, RL, DPO, distillation, or hybrid
  • Key details: Model, dataset, environment, reward signal, etc.

Step 2: Read existing recipes for patterns

Before writing any code, read the most relevant existing recipe:

  • SL-based: Read tinker_cookbook/recipes/sl_basic.py and tinker_cookbook/recipes/chat_sl/train.py
  • RL-based: Read tinker_cookbook/recipes/rl_basic.py and tinker_cookbook/recipes/math_rl/train.py
  • DPO-based: Read tinker_cookbook/recipes/preference/dpo/train.py
  • Distillation-based: Read tinker_cookbook/recipes/distillation/on_policy_distillation.py
  • Multi-turn RL: Read tinker_cookbook/recipes/harbor_rl/train.py

Also read CLAUDE.md for conventions.

Step 3: Follow repo conventions

Every recipe MUST follow these patterns:

File structure

tinker_cookbook/recipes/<recipe_name>/
├── __init__.py        # Empty or minimal
├── train.py           # Main entry point with CLIConfig + cli_main
└── <env_or_data>.py   # Dataset/environment definitions (if needed)

Or for simple recipes: tinker_cookbook/recipes/<recipe_name>.py

CLI pattern (use @chz.chz for config)

@chz.chz
class CLIConfig:
    model_name: str = "meta-llama/Llama-3.1-8B"
    learning_rate: float = 1e-4
    # ... all configurable parameters with defaults

async def cli_main(cli_config: CLIConfig):
    # Build full config from CLI config
    # Call training main function

if __name__ == "__main__":
    cli_config = chz.entrypoint(CLIConfig)
    asyncio.run(cli_main(cli_config))

Required elements

  1. @chz.chz config class with sensible defaults
  2. model_info.get_recommended_renderer_name(model_name) for renderer — never hardcode
  3. cli_utils.check_log_dir() before training to avoid clobbering
  4. checkpoint_utils.resolve_renderer_name_from_checkpoint_or_default_async() if loading checkpoints
  5. Explicit typing — no Any or type: ignore
  6. Auto-generated log paths with model name, hyperparams, and timestamp

Naming conventions

  • Subscript suffixes for tensors: _P (problems), _G (groups), _T (tokens), _D (datums)
  • Use safezip, timed, scope helpers where appropriate
  • Use ml_log.log_metrics for metrics, logtree for transcripts

Entry point

Recipe must be runnable as:

python -m tinker_cookbook.recipes.<recipe_name>.train [chz overrides]

Step 4: Create the recipe

Write the recipe files following the patterns above. Place them in tinker_cookbook/recipes/$ARGUMENTS/.

Step 5: Add tests

The repo has two layers of testing. Both should be added for every new recipe.

Smoke test (required)

Create tests/recipes/test_recipe_<name>.py — a minimal test that runs the recipe for 2 training steps and verifies clean exit. CI auto-discovers these files and runs them daily.

import pytest
from tests.helpers import run_recipe

@pytest.mark.integration
def test_<recipe_name>():
    run_recipe(
        "tinker_cookbook.recipes.<recipe_name>.train",
        [
            "behavior_if_log_dir_exists=delete",
            # Override params to make it fast:
            # "groups_per_batch=4", "group_size=2", "batch_size=16", etc.
        ],
    )

Key conventions:

  • run_recipe() launches the module as a subprocess and automatically passes max_steps=2 (configurable via the max_steps parameter)
  • The recipe runs for 2 training steps and exits naturally — the test passes on clean exit (exit code 0)
  • Always pass behavior_if_log_dir_exists=delete to avoid conflicts in repeated CI runs
  • Override batch sizes / group sizes to small values so the test completes quickly
  • Mark tests with @pytest.mark.integration — these require TINKER_API_KEY
  • See tests/helpers.py for run_recipe() details and tests/conftest.py for fixtures

Unit tests (for testable components)

Place unit tests next to the code they test using the *_test.py naming convention:

tinker_cookbook/recipes/<recipe_name>/<component>_test.py

For example:

  • tinker_cookbook/recipes/math_rl/math_env_test.py — tests environment logic
  • tinker_cookbook/renderers/parsing_test.py — tests parsing helpers

Unit tests should:

  • Run without TINKER_API_KEY (no network calls)
  • Be fast (< 1s per test)
  • Use standard pytest features (fixtures, parametrize, marks)
  • Test picklability if the component needs to be serialized for distributed rollout

Running tests locally

# Unit tests only (no API key needed)
uv run pytest tinker_cookbook/

# Integration / smoke tests (requires TINKER_API_KEY)
uv run pytest tests/recipes/test_recipe_<name>.py -v -x -s

CI integration

  • Unit tests (pytest tinker_cookbook/) run on every PR via .github/workflows/pytest.yaml
  • Integration tests (pytest tests/) run daily and on manual trigger via .github/workflows/smoke-test-recipes.yaml
  • Adding tests/recipes/test_recipe_<name>.py is all that's needed — CI auto-discovers it

Step 6: Verify

  • Ensure the recipe is importable: python -c "from tinker_cookbook.recipes.<name> import train"
  • Check that CLI help works: python -m tinker_cookbook.recipes.<name>.train --help
  • Run the smoke test locally: uv run pytest tests/recipes/test_recipe_<name>.py -v -x -s