Ludwig is a low-code framework for building custom AI models like LLMs and other deep neural networks.

Key features:

• 🛠 Build custom models with ease: a declarative YAML configuration file is all you need to train a state-of-the-art LLM on your data. Support for multi-task and multi-modality learning. Comprehensive config validation detects invalid parameter combinations and prevents runtime failures.
• ⚡ Optimized for scale and efficiency: automatic batch size selection, distributed training (DDP, DeepSpeed), parameter efficient fine-tuning (PEFT), 4-bit quantization (QLoRA), paged and 8-bit optimizers, and larger-than-memory datasets.
• 📐 Expert level control: retain full control of your models down to the activation functions. Support for hyperparameter optimization, explainability, and rich metric visualizations.
• 🧱 Modular and extensible: experiment with different model architectures, tasks, features, and modalities with just a few parameter changes in the config. Think building blocks for deep learning.
• 🚢 Engineered for production: prebuilt Docker containers, native support for running with Ray on Kubernetes, export models to Torchscript and Triton, upload to HuggingFace with one command.

Ludwig is hosted by the Linux Foundation AI & Data.

Install from PyPi. Be aware that Ludwig requires Python 3.8+.

pip install ludwig

Or install with all optional dependencies:

pip install ludwig[full]

Please see contributing for more detailed installation instructions.

Want to take a quick peak at some of the Ludwig 0.8 features? Check out this Colab Notebook 🚀

Looking to fine-tune Llama-2 or Mistral? Check out these notebooks:

1. Fine-Tune Llama-2-7b:
2. Fine-Tune Llama-2-13b:
3. Fine-Tune Mistral-7b:

For a full tutorial, check out the official getting started guide, or take a look at end-to-end Examples.

Let's fine-tune a pretrained LLaMA-2-7b large language model to follow instructions like a chatbot ("instruction tuning").

• HuggingFace API Token
• Access approval to Llama2-7b-hf
• GPU with at least 12 GiB of VRAM (in our tests, we used an Nvidia T4)

We'll use the Stanford Alpaca dataset, which will be formatted as a table-like file that looks like this:

Create a YAML config file named model.yaml with the following:

model_type: llm
base_model: meta-llama/Llama-2-7b-hf

quantization:
  bits: 4

adapter:
  type: lora

prompt:
  template: |
    Below is an instruction that describes a task, paired with an input that may provide further context.
    Write a response that appropriately completes the request.

    ### Instruction:
    {instruction}

    ### Input:
    {input}

    ### Response:

input_features:
  - name: prompt
    type: text

output_features:
  - name: output
    type: text

trainer:
  type: finetune
  learning_rate: 0.0001
  batch_size: 1
  gradient_accumulation_steps: 16
  epochs: 3
  learning_rate_scheduler:
    decay: cosine
    warmup_fraction: 0.01

preprocessing:
  sample_ratio: 0.1

backend:
  type: local

And now let's train the model:

export HUGGING_FACE_HUB_TOKEN = "<api_token>"

ludwig train --config model.yaml --dataset "ludwig://alpaca"

Let's build a neural network that predicts whether a given movie critic's review on Rotten Tomatoes was positive or negative.

Our dataset will be a CSV file that looks like this:

Download a sample of the dataset from here.

wget https://ludwig.ai/latest/data/rotten_tomatoes.csv

Next create a YAML config file named model.yaml with the following:

input_features:
  - name: genres
    type: set
    preprocessing:
      tokenizer: comma
  - name: content_rating
    type: category
  - name: top_critic
    type: binary
  - name: runtime
    type: number
  - name: review_content
    type: text
    encoder:
      type: embed
output_features:
  - name: recommended
    type: binary

That's it! Now let's train the model:

ludwig train --config model.yaml --dataset rotten_tomatoes.csv

Happy modeling

Try applying Ludwig to your data. Reach out on Discord if you have any questions.

• Minimal machine learning boilerplate

  Ludwig takes care of the engineering complexity of machine learning out of the box, enabling research scientists to focus on building models at the highest level of abstraction. Data preprocessing, hyperparameter optimization, device management, and distributed training for torch.nn.Module models come completely free.

• Easily build your benchmarks

  Creating a state-of-the-art baseline and comparing it with a new model is a simple config change.

• Easily apply new architectures to multiple problems and datasets

  Apply new models across the extensive set of tasks and datasets that Ludwig supports. Ludwig includes a full benchmarking toolkit accessible to any user, for running experiments with multiple models across multiple datasets with just a simple configuration.

• Highly configurable data preprocessing, modeling, and metrics

  Any and all aspects of the model architecture, training loop, hyperparameter search, and backend infrastructure can be modified as additional fields in the declarative configuration to customize the pipeline to meet your requirements. For details on what can be configured, check out Ludwig Configuration docs.

• Multi-modal, multi-task learning out-of-the-box

  Mix and match tabular data, text, images, and even audio into complex model configurations without writing code.

• Rich model exporting and tracking

  Automatically track all trials and metrics with tools like Tensorboard, Comet ML, Weights & Biases, MLFlow, and Aim Stack.

• Automatically scale training to multi-GPU, multi-node clusters

  Go from training on your local machine to the cloud without code changes.

• Low-code interface for state-of-the-art models, including pre-trained Huggingface Transformers

  Ludwig also natively integrates with pre-trained models, such as the ones available in Huggingface Transformers. Users can choose from a vast collection of state-of-the-art pre-trained PyTorch models to use without needing to write any code at all. For example, training a BERT-based sentiment analysis model with Ludwig is as simple as:

  ludwig train --dataset sst5 --config_str "{input_features: [{name: sentence, type: text, encoder: bert}], output_features: [{name: label, type: category}]}"

• Low-code interface for AutoML

  Ludwig AutoML allows users to obtain trained models by providing just a dataset, the target column, and a time budget.

  auto_train_results = ludwig.automl.auto_train(dataset=my_dataset_df, target=target_column_name, time_limit_s=7200)

• Easy productionisation

  Ludwig makes it easy to serve deep learning models, including on GPUs. Launch a REST API for your trained Ludwig model.

  ludwig serve --model_path=/path/to/model

  Ludwig supports exporting models to efficient Torchscript bundles.

  ludwig export_torchscript -–model_path=/path/to/model

• Text Classification
• Tabular Data Classification
• Image Classification
• Multimodal Classification

• Named Entity Recognition Tagging
• Natural Language Understanding
• Machine Translation
• Chit-Chat Dialogue Modeling through seq2seq
• Sentiment Analysis
• One-shot Learning with Siamese Networks
• Visual Question Answering
• Spoken Digit Speech Recognition
• Speaker Verification
• Binary Classification (Titanic)
• Timeseries forecasting
• Timeseries forecasting (Weather)
• Movie rating prediction
• Multi-label classification
• Multi-Task Learning
• Simple Regression: Fuel Efficiency Prediction
• Fraud Detection

Read our publications on Ludwig, declarative ML, and Ludwig’s SoTA benchmarks.

Learn more about how Ludwig works, how to get started, and work through more examples.

If you are interested in contributing, have questions, comments, or thoughts to share, or if you just want to be in the know, please consider joining our Community Discord and follow us on X!

Ludwig is an actively managed open-source project that relies on contributions from folks just like you. Consider joining the active group of Ludwig contributors to make Ludwig an even more accessible and feature rich framework for everyone to use!

• Discord
• X
• Medium
• GitHub Issues