# Ludwig

> Ludwig is a low-code, declarative deep learning framework for building custom AI models including LLMs and neural networks using YAML configuration files.

Ludwig is a **low-code** declarative deep learning framework built for scale and efficiency, enabling researchers and engineers to train custom AI models — including LLMs and other deep neural networks — using simple YAML configuration files. It abstracts away machine learning boilerplate while retaining expert-level control over model architecture, training, and deployment. Hosted by the Linux Foundation AI & Data, Ludwig supports multi-modal and multi-task learning out of the box.

- **Declarative YAML configuration** — *define your entire model, preprocessing, training loop, and hyperparameter search in a single config file without writing boilerplate code.*
- **LLM fine-tuning** — *fine-tune pretrained large language models (e.g., Llama-3.1-8B) with support for 4-bit quantization (QLoRA), LoRA adapters, and instruction tuning.*
- **Distributed training** — *scale from a single GPU to multi-GPU, multi-node clusters using DDP and DeepSpeed, with native Ray and Kubernetes support.*
- **Parameter-efficient fine-tuning (PEFT)** — *reduce compute and memory requirements using adapter-based methods like LoRA.*
- **AutoML** — *automatically train models by providing just a dataset, target column, and time budget via Ludwig AutoML.*
- **Multi-modal, multi-task learning** — *mix tabular data, text, images, and audio into complex model configurations without writing code.*
- **Hyperparameter optimization** — *built-in hyperopt support for automated search over model and training parameters.*
- **Rich integrations** — *track experiments with TensorBoard, Comet ML, Weights & Biases, MLFlow, and Aim Stack.*
- **Production-ready export** — *export models to Torchscript and Triton, upload to HuggingFace with one command, and serve via a built-in REST API.*
- **Extensible architecture** — *add custom encoders, decoders, combiners, feature types, metrics, and tokenizers through a modular developer API.*
- **HuggingFace Transformers integration** — *use any pretrained PyTorch model from HuggingFace without writing code.*
- **CLI and Python API** — *interact with Ludwig via command-line interface or the `LudwigModel` Python API.*

## Features
- Declarative YAML-based model configuration
- LLM fine-tuning with LoRA and QLoRA
- 4-bit quantization support
- Distributed training with DDP and DeepSpeed
- Parameter-efficient fine-tuning (PEFT)
- AutoML with time budget
- Multi-modal learning (text, image, audio, tabular)
- Multi-task learning
- Hyperparameter optimization
- Experiment tracking integrations
- Model export to Torchscript and Triton
- HuggingFace model upload
- Built-in REST API serving
- Ray and Kubernetes support
- Python API and CLI
- Prebuilt Docker containers
- Rich metric visualizations
- Dataset Zoo with built-in datasets

## Integrations
PyTorch, HuggingFace Transformers, Ray, Kubernetes, DeepSpeed, TensorBoard, Weights & Biases, MLFlow, Comet ML, Aim Stack, Triton Inference Server, Docker, Torchscript, Pydantic

## Platforms
LINUX, API, DEVELOPER_SDK, CLI

## Pricing
Open Source

## Version
0.11.2

## Links
- Website: https://ludwig.ai
- Documentation: https://ludwig.ai/latest/
- Repository: https://github.com/ludwig-ai/ludwig
- EveryDev.ai: https://www.everydev.ai/tools/ludwig