Transformer Lab

Transformer Lab is an open-source machine learning platform that provides a unified interface for training, evaluating, and scaling AI models. It serves as an alternative to fragmented AI tooling by consolidating model management and compute orchestration into a single desktop application. The software is available for individuals working on local hardware and for research teams managing larger GPU clusters, providing a consistent experience across different scales of infrastructure.

The application runs natively on macOS with Apple Silicon, Linux, and Windows via WSL2. Users can install the software via a curl script and access the interface through a web browser. It supports a wide range of foundation models, including Llama 3, Mistral, and DeepSeek, and integrates with various inference engines such as MLX, vLLM, and Ollama to handle diverse model architectures.

Key features

• Support for full fine-tuning, LoRA, QLoRA, and RLHF methods
• Integration with Slurm and SkyPilot for GPU cluster orchestration
• Conversion between HuggingFace, GGUF, and MLX model formats
• Text-to-Image and Inpainting using Stable Diffusion and Flux
• Built-in benchmarks via EleutherAI LM Evaluation Harness
• Automated red teaming for prompt injection and PII leakage
• Python plugin architecture and SDK for custom training scripts
• Text-to-Speech generation and fine-tuning using Kokoro and Bark

The platform is designed specifically for ML researchers and hobbyists who require a centralized environment for model development. It includes a model registry, artifact management, and centralized experiment tracking for collaborative environments. For remote work, it provides one-click access to Jupyter, VSCode, and SSH sessions on remote nodes, allowing researchers to move from local prototyping to remote execution without changing tools. The architecture allows users to define hyperparameter sweeps in YAML to automate grid searches across local hardware or cloud providers like AWS, GCP, and Azure.

Additionally, the software includes tools for dataset management, such as auto-captioning images using WD14 taggers for diffusion model training. The inclusion of a command line interface allows power users to submit tasks and monitor jobs without using the browser interface.

Transformer Lab positions itself as a control plane for AI research, bridging the gap between local experimentation and large-scale cluster deployment.