Groundwork AI Studio
Build datasets and train language models from random weights, on your own machines.
Groundwork is a desktop studio for the whole small-model loop: build a dataset, train a language model from random weights, evaluate it, and export it. The transformer modules, optimizers, tokenizers, checkpoint format and sampler are written in the project itself; PyTorch supplies autograd and little else.
Training scales past one box. Machines on the same Tailscale network join a cluster and train together with DiLoCo, with coordinator failover, replicas and shard transfer. The web UI is served by a local API, so you can watch a run from any of your devices.
It also covers evals and decontamination, LoRA, RL and RLHF modules, and export to Hugging Face and GGUF.
Highlights
- From-scratch transformer, optimizer, tokenizer and sampler implementations
- Multi-machine DiLoCo training with coordinator failover
- Runs on CPU, NVIDIA CUDA and Apple MPS
- Evals, decontamination, LoRA / RL / RLHF, export to HF and GGUF
- Nightly end-to-end training runs in CI on macOS and Windows