Soup

Why it's ranked here

Engineering care carries this more than any single novel idea: the same tool handles data cleaning, quantized training, evaluation gating, multi-format export and deployment, with every heavy capability split into its own optional install so a casual user never pulls down a full GPU toolchain by accident. The dependency manifest reads like it was written by people who got burned before, with real compatibility breakages named rather than guessed at. The headline memory-saving trick is honestly labeled experimental and opt-in rather than sold as the default, which is the kind of restraint that is easy to skip and rare to see.

What's good

The install layering is thoughtful: the base package carries no heavy machine-learning dependencies at all, and training, serving, the web dashboard and the agent-facing server are each their own add-on. That agent-facing server takes security seriously for something this niche: it checks its access token in constant time, blocks browser-driven requests to the local port with strict host checking, and scrubs every error message so a failure never leaks a stack trace or a file path back to the caller. The changelog's fixes read like they came from real, reproduced failures rather than hypothetical ones.

Tradeoffs

The command surface is enormous, spanning training, evaluation, deployment and governance, which makes it hard for an outside user to judge how deeply tested any one corner is compared with the well documented core loop. The project's own release notes admit a history of settings that were accepted, validated and then quietly ignored on one backend instead of erroring, a pattern only recently closed off. The headline low-memory training path is still marked experimental, and the maintainers themselves note that its last published benchmark predates a later correctness fix and has not been repeated since.

How to use it well

Reach for this when you want one tool to carry a model from a bare config file through training, evaluation and export without stitching together separate scripts, especially if you are experimenting on a small consumer graphics card and want the memory-saving path (treat it as experimental and confirm the numbers on your own hardware rather than trusting a published figure). Install only the light core if you need its data or configuration tooling and nothing more. Read the release notes before upgrading: this project ships real breaking changes to its configuration format, not just additive ones.

Observed

License

Apache-2.0 license, declared in the project manifest.

Packaging

Distributed as a Python package with a console-script entry point, installable via pip or pipx.

Install extras

The core install has no PyTorch dependency; the full training stack sits behind a separate optional extra.

Interfaces

Ships three distinct interfaces: a command-line tool, a local web dashboard, and a Model Context Protocol server.

MCP server

The Model Context Protocol server supports stdio transport locally and network transports (SSE, streamable HTTP); the network transports require bearer-token authentication and DNS-rebinding host checks.

Python support

Declared Python support is 3.10 through 3.12; newer Python majors are not declared as supported in the package manifest.

Install extras

Optional capabilities (training backends, serving, web UI, MCP server, Apple-Silicon support, evaluation tooling, and more) are each packaged as separate install extras.