Skip to content

Recipes

xaytune organizes training workflows into recipes -- high-level functions that encapsulate the full training pipeline for a specific task. Each recipe handles model loading, data preparation, training loop execution, and state management.

Available Recipes

Recipe Function Methods Use Case
Fine-tune xaytune.finetune() full, lora, qlora Adapt a pre-trained model to a specific task or domain
Pre-train xaytune.pretrain() full Train a model from scratch on a large text corpus
Align xaytune.align() dpo, grpo, orpo, simpo, ppo, reinforce Align a model with human preferences

How Recipes Work

Every recipe follows the same pattern:

  1. Build or accept a config -- either from keyword arguments or a TrainConfig object
  2. Set up components -- model, tokenizer, data loaders, trainer, callbacks, logging
  3. Run the training loop -- delegate to the trainer
  4. Return a TrainState -- final metrics, step count, and training status
import xaytune

# All recipes share the same calling convention:
state = xaytune.finetune(model="...", dataset="...", method="lora")
state = xaytune.pretrain(model="...", dataset="...", format="text")
state = xaytune.align(model="...", dataset="...", method="dpo")

Python API vs. Config

Each recipe can be called in two ways:

state = xaytune.finetune(
    model="meta-llama/Llama-3.1-8B",
    dataset="data/train.jsonl",
    method="lora",
    format="alpaca",
    num_epochs=3,
)

from xaytune.config.schema import TrainConfig, ModelConfig, DataConfig

config = TrainConfig(
    recipe="finetune",
    method="lora",
    model=ModelConfig(name="meta-llama/Llama-3.1-8B"),
    data=DataConfig(path="data/train.jsonl", format="alpaca"),
)
state = xaytune.finetune(config=config)

recipe: finetune
method: lora
model:
  name: meta-llama/Llama-3.1-8B
data:
  path: data/train.jsonl
  format: alpaca

xaytune train --config config.yaml

Recipe Registry

Recipes are registered in a global recipe_registry. You can list them:

from xaytune.recipes import recipe_registry

print(recipe_registry.list())
# ['align', 'finetune', 'pretrain']

Or from the CLI:

xaytune list recipes

Next Steps