🧠 Modlable

A platform for building, training, and running inference on TensorFlowJS-based language models with the assistance of LLM-powered agents — hosted on Google Cloud Run.

Modlable is an experimental platform that allows developers and researchers to build, train, and deploy lightweight language models (LMs) directly in JavaScript using TensorFlowJS, augmented by Large Language Model (LLM) agents for orchestration, optimization, and workflow automation.

The platform is cloud-native — it’s designed to run efficiently on Google Cloud Run, enabling serverless and scalable model operations with minimal DevOps overhead.

• 🧩 Modular Training Pipelines — Train custom token-based language models in TensorFlowJS.
• 🎛️ Node-to-Python SFT Training Bridge — Orchestrate heavy Hugging Face and TRL training workloads (SFTTrainer/PEFT LoRA) from TypeScript using real-time child process stdout streaming and stdin JSON IPC.
• 🧪 Build-Time Schema Verification — Automated type-checking that matches TypeScript training configurations with Python dataclasses via introspection to prevent API drift.
• 🤖 LLM Orchestration Agents — Use pre-configured agents to guide hyperparameter tuning, dataset preparation, and model evaluation.
• ☁️ Cloud Run Deployment — Run models and inference endpoints on demand using containerized builds.
• 🔁 Seamless Inference API — Serve models as REST endpoints for text generation, embedding, or classification.
• 🧠 Local + Cloud Support — Develop locally, then deploy to GCP with one command.

modlable/
├── frontend/
│   ├── modalble/            # The frontend Angular/Ionic App
│
├── backend/
│   ├── src/                 # Genkit + Cloud Functions/Cloud Run for agents
│   │   ├── trainer/         # Node-to-Python SFT Training Bridge
│   │   │   ├── types.ts     # TypeScript interfaces for configurations
│   │   │   ├── NodeSFTTrainer.ts # Wrapper orchestration class
│   │   │   ├── bridge.py    # Python runner executing TRL SFTTrainer
│   │   │   └── verify_types.py # Introspection type verification layer
│   │   └── index.ts         # Agent logic and flows
│   ├── Dockerfile           # Dual-runtime container environment
│   ├── requirements.txt     # Python training packages (torch, trl, transformers)
│   └── package.json         # Node project dependencies
├── docker-compose.yaml      # Multi-container local execution mapping
├── .env.example             # Example environment configuration
└── README.md                # You are here

Before getting started, ensure you have:

• Node.js ≥ 18.x
• npm or yarn
• Python ≥ 3.10 and pip (for local Python bridge development)
• Docker (for containerization)
• Google Cloud SDK (gcloud) configured with billing and permissions
• A Google Cloud Project with Cloud Run, Artifact Registry, and Cloud Build enabled

1. Clone the repository

   git clone https://github.com/yourusername/modlable.git
   cd modlable

2. Install dependencies

   npm install

3. Set up environment variables

   cp .env.example .env
   # Edit the .env file with your API keys and Cloud project info

Run the API locally for testing:

npm run dev

To start a local training session:

npm run train

Run inference on a local model:

npm run infer "Once upon a time"

1. Build and push Docker image

   gcloud builds submit --tag gcr.io/[PROJECT_ID]/modlable

2. Deploy to Cloud Run

   gcloud run deploy modlable \
       --image gcr.io/[PROJECT_ID]/modlable \
       --platform managed \
       --region [REGION] \
       --allow-unauthenticated

3. Access your endpoint

   https://modlable-[REGION]-a.run.app/infer
   

Example Request:

curl -X POST https://modlable-[REGION]-a.run.app/infer \
  -H "Content-Type: application/json" \
  -d '{"prompt": "The future of AI is"}'

Contributions are welcome!

1. Fork the repo
2. Create a new feature branch
3. Submit a pull request with a clear description

MIT License © 2025 — Modlable Contributors

• Web dashboard for model visualization and management
• Support for multimodal training data (text + image)
• Integration with Hugging Face model zoo
• Cross-runtime inference (Node, Browser, Edge)

Made with ❤️ by the Modlable Team Empowering developers to build LMs anywhere.