About
About
A workbench for language model experiments, built on the idea that a small model you own and understand beats a large one you rent and can't inspect.
Why this exists
General-purpose models are remarkable, and for most one-off questions they are the right tool. But a lot of real work is narrow and repetitive: the same kind of document, the same kind of decision, the same output format, thousands of times over. For that shape of problem, a small model tuned on your own data can be accurate enough, dramatically cheaper to run, and because you hold the weights, something you can deploy wherever you like.
Getting there is the awkward part. The techniques are published and the open models are available, but stitching together data preparation, training orchestration, honest evaluation, and serving is a real infrastructure project before you have learned anything about your task. Ekcron exists to remove that gap: bring data, run the experiment, see whether it worked.
How it approaches the problem
Three principles shape the platform:
- Evaluation is not optional. A tuned model is only interesting relative to a baseline. A split is held out before training, and results are always reported against both the held-out data and the untuned model, including when the answer is that nothing improved.
- You keep what you make. Datasets, adapters, and exported builds can be downloaded and run on your own hardware. Hosted inference is a convenience, never a lock-in mechanism.
- Small and reproducible over large and mysterious. Every run records its dataset version, configuration, and metrics, so a result you liked can be found again and repeated.
What this is, plainly
Ekcron is an independent, experimental engineering project, built and maintained by one person in a personal capacity. It is not a company, not a registered business entity, and not a commercial service. Nothing is sold and no fee is charged.
It runs as is, with no warranties and no guarantees of uptime, retention, or confidentiality. Please do not upload confidential or sensitive data, and do not depend on it for production work. Whatever you build here is yours: the project claims no ownership of your data, your trained models, or their output. The terms and privacy notice set this out in full.
The platform is under active development. The pipeline described on this site, from data preparation through to deployment and export, is what it does today. Capabilities land incrementally, and anything genuinely usable shows up in the app rather than only on this page.
Open source
The platform is developed in the open. The full source — the training pipeline, data generation, and the evaluation harness — lives on GitHub, so you can read exactly how your data is handled and how every number on a results page was produced. If a claim on this site and the code ever disagree, the code is the truth.
Contributions are welcome in whatever size: an issue describing something confusing, a bug report with a reproduction, or a pull request. If you're unsure whether something is worth reporting, it is.
Get in touch
Questions, feedback, and bug reports are all welcome. The contact page has the details, or email hello@ekcron.com. Public code and issues live on GitHub.