API Reference

This section documents xaytune's Python API. The library exposes four top-level functions and a supporting ecosystem of configs, callbacks, evaluation, and export utilities.

Top-Level Functions

These are the primary entry points, importable directly from xaytune:

import xaytune

xaytune.finetune(...)   # Fine-tune a model (full, LoRA, QLoRA)
xaytune.pretrain(...)   # Pre-train a model from scratch
xaytune.align(...)      # Align a model with human preferences
xaytune.evaluate(...)   # Evaluate a model on metrics

Module Overview

Module	Description
xaytune.config.schema	Pydantic config models (TrainConfig, ModelConfig, etc.)
xaytune.config.parser	YAML config loading and override parsing
xaytune.config.validation	Config validation rules
xaytune.trainer.callbacks	CallbackManager, TrainState, event system
xaytune.eval.evaluate	evaluate() function for custom datasets
xaytune.eval.benchmarks	benchmark_evaluate() for lm-eval benchmarks
xaytune.eval.metrics	Metric registry and built-in metrics
xaytune.export.merge	merge() and save() for model persistence
xaytune.export.hub	push_to_hub() for Hugging Face Hub
xaytune.export.gguf	to_gguf() for GGUF conversion
xaytune.data.formats	Built-in data format functions
xaytune.data.registry	format_registry
xaytune.models.registry	model_registry
xaytune.recipes.align.rewards	reward_registry and reward functions
xaytune.utils.registry	Generic Registry class

Registries

xaytune uses a registry pattern to make components extensible. Each registry maps string names to callable objects:

Registry	Location	Decorator	Purpose
format_registry	xaytune.data.registry	@format_registry.register("name")	Data format functions
metric_registry	xaytune.eval.metrics	@register_metric("name")	Evaluation metrics
reward_registry	xaytune.recipes.align.rewards	@register_reward("name")	Reward functions for alignment
model_registry	xaytune.models.registry	@model_registry.register("name")	Model loaders
recipe_registry	xaytune.recipes	@recipe_registry.register("name")	Training recipes

Registry API

All registries share the same interface:

from xaytune.utils.registry import Registry

registry = Registry("my_registry")

# Register an item
@registry.register("key")
def my_function():
    ...

# Retrieve an item
fn = registry.get("key")

# Check existence
registry.has("key")  # True

# List all registered keys
registry.list()  # ['key']

Detailed References

• Config -- all Pydantic config models and their fields
• Callbacks -- event system and TrainState
• Evaluation -- evaluate(), benchmark_evaluate(), and metrics
• Export -- merge(), save(), push_to_hub(), to_gguf()