uber/ludwig
Ludwig
Ludwig is a declarative machine learning framework designed for training neural networks and large language models using configuration files instead of manual coding. It functions as a multimodal model builder and a low-code tool for supervised fine-tuning, allowing users to build models that process mixed inputs of text, images, audio, and tabular data.
The project distinguishes itself through an automated hyperparameter optimizer and a system for large language model fine-tuning using parameter-efficient adapters. It features a multimodal data pipeline and the ability to automatically generate declarative configuration files using large language models based on task descriptions.
The framework covers a broad set of capabilities including automated model selection, multi-task learning with game-theoretic loss balancing, and time series forecasting. It also provides a full deployment pipeline to export trained weights and serve models as REST APIs within production clusters.
Training operations are supported by experiment tracking, model weight quantization, and dataset quality validation.
Features
- Declarative Training Frameworks - Implements a declarative machine learning framework that uses configuration files to build and train neural networks.
- Declarative Pipeline Specifications - Allows building and training neural networks using declarative YAML or JSON configurations instead of manual Python code.
- Declarative Machine Learning Frameworks - Provides a framework for building and training neural networks using declarative configuration files instead of manual coding.
- Encoder-Combiner Architectures - Utilizes an architecture that separates diverse input processing into specialized encoders and merges them via a central combiner.
- Multimodal Model Trainers - Ships a system for creating and training models that process mixed modalities including text, image, and audio.
- LLM Fine-Tuning Engines - Provides a low-code engine for the efficient supervised fine-tuning of large language models.
- Large Language Model Fine-Tuning Frameworks - Adapts pre-trained large language models for specific tasks using supervised fine-tuning and parameter-efficient adapters.
- Multimodal Fine-Tuning - Trains vision-language models using gated cross-attention to simultaneously process image and text inputs.
- Model Training - Allows the construction of neural networks and LLMs via declarative configurations instead of manual training loops.
- Multimodal AI Systems - Provides a framework for building AI systems that process and combine text, images, audio, and tabular data.
- Multimodal Analytical Pipelines - Provides a unified system to process text, images, and tabular data and map them into a common latent space.
- Multimodal Models - Builds architectures that process and align diverse data types like text, images, and audio in a shared space.
- Multimodal Processing - Integrates multiple data modalities including text, images, audio, and time series using a single unified configuration.
- Parameter-Efficient Adaptation - Implements parameter-efficient adaptation using trainable adapter layers to customize large models while keeping base weights frozen.
- Supervised Fine-Tuning - Supports adapting pre-trained models using labeled instruction datasets and quantization techniques.
- Distributed Training Orchestration - Manages parallelization and synchronization of model workloads across multiple GPU clusters to accelerate training.
- Distributed Training Scaling Utilities - Scales training workloads across clusters using mixed precision and automatic batch selection to increase speed.
- Experiment Tracking - Logs and versions every training run and configuration to ensure reproducibility and result visualization.
- Multi-Task Learning Models - Enables the training of a single model to predict multiple output features simultaneously.
- Game-Theoretic Loss Balancing - Optimizes multi-task loss functions using game-theoretic methods to prevent any single task from dominating the gradient.
- Model Deployment Pipelines - Offers a standardized pipeline for exporting trained weights and serving models as REST APIs.
- Hyperparameter Tuning - Utilizes sampling algorithms and persistent storage to iteratively optimize model configuration parameters.
- Automated Architecture Search - Automatically searches for the best combination of encoders and combiners based on data and time budgets.
- Model Deployment Pipelines - Provides a deployment pipeline to transition trained models into production via REST APIs and scalable clusters.
- Hyperparameter Optimization - Provides automated methods for searching and selecting the best configuration parameters to optimize model performance.
- Time Series Forecasting - Supports the development of models designed for predicting future values in temporal data sequences.
- Game-Theoretic Loss Balancing - Dynamically adjusts weights of multiple task-specific loss functions using game-theoretic methods to prevent any single task from dominating.
- Model Alignment and Feedback - Provides capabilities for aligning language model outputs using policy optimization and human feedback.
- LLM-Driven Configuration Generation - Automatically generates declarative configuration files using large language models based on provided task descriptions.
- REST APIs - Exposes trained models as web services with REST endpoints to handle programmatic inference requests.
- Containerized Model Serving - Packages trained model weights and inference logic into standardized containers for production deployment.
- Model Inference Deployment - Transitions trained models into production environments to perform real-world inference on live data.
- ML Model Hosting - Hosts models in distributed environments using container orchestration for scalable and reliable serving.
- Automated Machine Learning - Toolbox for training deep learning models without writing code.
- Deep Learning - Toolbox for training deep learning models without coding.
- Simplification Tools - Trains and tests deep learning models without writing code.
