Back to Open Source

Playground

Proprietary Platform Core

The core infrastructure powering Avitai Bio's foundation models. An experimental environment for cutting-edge biological research.

Repository Coming Soon

This project is under active development and will be open-sourced soon

Overview

Playground is the beating heart of Avitai Bio's foundation model development. It's the infrastructure that powers our Research, Dynamics, Central Dogma, and Perturbation models, and it's now available to the research community.

Built from the ground up for biological AI, Playground abstracts away the complexity of training and deploying large-scale models. Whether you're working with protein sequences, molecular structures, or cellular imaging data, Playground provides a consistent interface that just works.

The platform is designed for researchers who want to push boundaries. It includes everything you need to go from raw data to deployed models: data loaders for biological formats, model architectures proven in production, training loops optimized for biological data characteristics, and evaluation metrics that matter for biology.

Playground also serves as the integration point for Avitai's open-source ecosystem. Use Artifex for generative modeling components, Datarax for data preprocessing, and Opifex for physics-informed constraints – all within a unified Playground experiment. The platform handles orchestration, making it easy to build sophisticated biological AI systems.

What sets Playground apart is its focus on reproducibility and scientific rigor. Every experiment is versioned, logged, and trackable. Models are checkpointed automatically, hyperparameters are recorded, and results are stored in structured formats. This makes it easy to reproduce published results, share models with collaborators, and build on previous work.

Key Features

Model Infrastructure

Unified interface for training, deploying, and serving foundation models at scale. Optimized for biological data and large-scale experiments.

Experimental Tools

Rapid prototyping environment with pre-built components for biological experiments, from data loading to evaluation.

Research APIs

Clean, consistent APIs for integrating foundation models into research workflows. Access pretrained models and run experiments programmatically.

Integration Layer

Seamless integration with Artifex, Datarax, Opifex, and other Avitai tools. Build end-to-end biological workflows.

Experiment Tracking

Comprehensive logging, metrics tracking, and experiment management. Compare models, track hyperparameters, and reproduce results.

Distributed Training

Scale training across GPUs and nodes with minimal code changes. Built-in support for data parallelism and model parallelism.

Use Cases

1

Training custom foundation models on proprietary biological datasets

2

Fine-tuning pretrained models for specific research tasks

3

Running large-scale hyperparameter searches for model optimization

4

Benchmarking different architectures on standardized biological tasks

5

Prototyping novel model architectures for biological sequence modeling

6

Integrating multiple data modalities (sequence, structure, imaging)

7

Building custom data preprocessing pipelines for model training

8

Deploying trained models as REST APIs for team collaboration

Installation

# Install from PyPI
pip install avitai-playground

# With all dependencies (recommended for full features)
pip install avitai-playground[all]

# Or install from source
git clone https://github.com/avitai/playground.git
cd playground
pip install -e .

Quick Start

from playground import Model, Trainer, Experiment
from playground.models import ProteinLM
from playground.data import BiologicalDataLoader

# Set up an experiment
experiment = Experiment(
    name="protein-lm-training",
    description="Fine-tuning protein language model"
)

# Initialize model
model = ProteinLM.from_pretrained("avitai/protein-lm-base")

# Load data
data = BiologicalDataLoader(
    path="data/proteins.fasta",
    batch_size=32
)

# Train
trainer = Trainer(model, data, experiment)
trainer.fit(epochs=10)

# Use the model
predictions = model.predict(new_sequences)

Built With

PyTorchJAXTransformersLightningHydraWeights & BiasesMLflowRayDVCDocker

Ready to Get Started?

Explore the documentation, try examples, or contribute to the project.