Skip to content

xaytune

An opinionated LLM training and fine-tuning library.

xaytune provides a recipe-based approach to LLM training built on PyTorch and Hugging Face Transformers. It covers the full lifecycle from pre-training through fine-tuning to alignment, with a clean Python API and a YAML-driven CLI.

Key Features

  • 3 recipes -- fine-tune, pre-train, and align -- each accessible as a single function call
  • Multiple methods -- full fine-tuning, LoRA, QLoRA, DPO, GRPO, ORPO, SimPO, PPO, and REINFORCE
  • Agent fine-tuning -- tool-use data formats (function calling, ReAct, trajectory, multi-agent) with loss masking, agent-specific rewards and evaluation metrics
  • Data preparation toolkit -- dedup, quality filtering, format conversion, synthetic data generation, pipeline chaining
  • Model merging -- combine fine-tuned models with Linear, SLERP, TIES, and DARE algorithms
  • Pydantic config system -- type-safe, validated configuration with sensible defaults
  • Registry pattern -- extend data formats, metrics, rewards, and model loaders via decorators
  • Callback system -- hook into training events with @on("step_end") style decorators
  • Evaluation -- built-in metrics (loss, perplexity, token accuracy), agent evaluation, and lm-eval benchmark integration
  • Export pipeline -- merge LoRA adapters, model merging, push to Hugging Face Hub, convert to GGUF
  • Flexible logging -- console, TensorBoard, Weights & Biases, and MLflow backends
  • Distributed training -- DDP, FSDP, and DeepSpeed strategies

Quick Example

import xaytune

# Fine-tune with LoRA in one call
state = xaytune.finetune(
    model="meta-llama/Llama-3.1-8B",
    dataset="data/train.jsonl",
    method="lora",
    format="alpaca",
    num_epochs=3,
)

Or drive the same workflow from the CLI:

xaytune train --config configs/examples/lora_finetune.yaml

How It Works

  1. Choose a recipe -- finetune, pretrain, or align
  2. Configure -- pass keyword arguments in Python, or write a YAML config
  3. Train -- xaytune loads the model, prepares data, runs the training loop, and returns a TrainState
  4. Export -- merge adapters, push to Hub, or convert to GGUF for local inference

Next Steps