Purpose: Transformer Lab is aimed at machine learning researchers, integrating scattered AI tools into one interface, allowing models to be trained, evaluated, and scaled locally, on a local server, or in the cloud. The Personal Edition is suitable for single-machine use, with data staying on the local machine; the Team Edition is aimed at GPU clusters.
Main capabilities: Supports downloading and running models such as Llama 3, DeepSeek, Mistral, Qwen, and Phi; inference engines cover MLX, vLLM, Ollama, and HuggingFace Transformers; conversion is possible among HuggingFace, GGUF, and MLX formats. Training methods include full fine-tuning, LoRA/QLoRA, RLHF (DPO, ORPO, SIMPO), and reward modeling, with support for Apple Silicon, NVIDIA, and AMD hardware. There is also text-to-image, image-to-image, and inpainting, ControlNets and IP-Adapters, the ability to train custom LoRA adapters, and automatic dataset annotation with WD14. For evaluation, it provides LLM-as-a-Judge, the EleutherAI LM Evaluation Harness benchmark, and red teaming. The Team Edition can uniformly submit jobs to Slurm clusters or SkyPilot clouds (AWS, GCP, Azure), and supports centralized experiment tracking, model registry, artifact management, one-click Jupyter/VSCode/SSH, as well as checkpoint recovery and automatic recovery of preemptible instances. Extension methods include a Python plugin system, the Lab SDK (import lab), and a CLI.
How to use: After installing the Personal Edition, open a browser and visit http://localhost:8338 . The Team Edition requires configuration according to the Teams Install Guide, and setting the compute provider in Team Settings, then using the CLI (lab) or Web UI to queue tasks on the cluster.
Limitations: Official materials do not provide pricing, specific installation commands, system requirements, or complete steps; this information is not currently available. The project uses the AGPL-3.0 license; see LICENSE for details.
