FAQ

Privacy#

Does anything I send leave my computer?#

No. Models run on your computer, and requests, answers and history stay in the studio's data folder. The studio sends nothing about your decisions anywhere and has no telemetry.

It does use the internet for three things, all to fetch software: installing the engine (Python, PyTorch and the model libraries), downloading the models you choose from Hugging Face, and reading each model's file list, size and likes from Hugging Face at startup (cached for six hours, and the studio works offline without it). Loading and running a model never touches the network.

Where is my history kept, and how do I delete it?#

In studio.db in the data folder. Delete single decisions or everything matching a filter on the History page, or with DELETE /v1/studio/decisions/{id} and the bulk endpoints. Decisions older than 30 days are removed automatically; you can change the period or keep them forever.

How do I stop the studio keeping decisions?#

At whichever level suits you:

• One call: send "store": false in the body, or the header X-Basal-Store: 0. TypeSafe's own servers ignore the field, so the same code runs against both.
• One template: set the template's storage to none.
• Everything: on the History page, choose Nothing for what history keeps, or send {"history": {"store": "none"}} with PATCH /v1/studio/settings.

"answers_only" is the middle ground: it keeps the answers, model and timing, but not the situation, variables or media. Variables marked sensitive are never written to disk at any level. Details: History and privacy.

Hardware#

Do I need a GPU?#

No. Every model except Jev-Omni runs on the processor, and small models such as Laya and Julia 1 answer in about a second there. A GPU makes the 4B to 12B models practical.

Intel Macs are not supported, because current PyTorch releases do not support them.

How much memory does a model need?#

The memory each model needs once loaded, as the Models page shows it:

Several models can be loaded at once. A model too large for your device is marked Too large here on the Models page and the loader says why instead of loading it; the System page shows what each loaded model uses.

Do I need a GPU to train a model?#

Yes. Teaching a model on the Train page needs an NVIDIA RTX 30 series or newer (including the GB10), or an Apple M2 or newer; Intel Arc and Core Ultra graphics, AMD on Linux, older NVIDIA cards and the M1 work as an experimental option. Trained models then run anywhere their original model runs, including on the processor.

Models#

Which model should I start with?#

Each model comes with examples made for it: choosing a model in the Playground shows what it is good at. Decision models explains how they differ.

Are the answers the same as Jev's?#

No. These are open models from different makers, trained differently from TypeSafe's Jev; the studio runs them behind Jev's API. The Models page shows each model's published results next to Jev's on the same benchmarks, labelled as the publishers' own figures. Probabilities are comparable across models in shape, but each model's calibration differs; the Evaluate page measures it on your own examples.

What licence do the models have?#

Each model keeps its own licence. The studio's catalog lists all eleven under Apache-2.0; each model's page in the app links to its Hugging Face card, which is the authority. Weights come from each publisher's Hugging Face repository and are not redistributed by the studio.

Training#

Do my examples leave my computer?#

No. Training runs on this computer's GPU; the file you add is kept in the studio's data folder with the training job, and nothing is uploaded.

How many examples do I need?#

At least 30 answered questions, and a few hundred for a result you can rely on: the studio sets about a third of them aside to test the result, and with fewer than 300 only large improvements can be told apart from luck. Every answer should appear at least ten to twenty times. Teach a model your own decisions has the details.

Can training make the model worse?#

Not one you get to use. The trained model is tested on examples it never saw and on general questions it never trains on, and it is added only if it is better on the first without a clear drop on the second; otherwise the original stays and the page says why. The original model is never changed. Fine-tuning and its safeguards lists the checks and the results on every model.

Can I use a trained model on another computer?#

Yes. Export on its row of the Train page saves it as one .zip, a few to a few hundred MB; Import a trained model on the other computer's Train page adds it. The model it was trained from must be downloaded there too.

Using it from code#

Does my code written for Jev work?#

Yes, by changing the base URL to http://127.0.0.1:8420. The studio serves TypeSafe's POST /v1/systemone and GET /v1/models with the same request and response shapes, and the official SDKs, typesafe-sdk for Python and @typesafe-ai/sdk for JavaScript, pass every conformance check unchanged. OpenRouter's and Vercel AI Gateway's formats are served too. Responses are exactly TypeSafe's shape unless you ask for more with X-Basal-Extensions: 1. See Move from Jev or a gateway.

What is the difference between /v1/systemone and /v1/studio?#

/v1/systemone is stateless and TypeSafe-compatible: you send the situation and questions every time. /v1/studio adds what a local studio can offer: templates you call by name with variables, versions, a decision object that says whether to act, history you can query, feedback and test examples. Both record decisions in history. See the API overview.

Can it serve other machines?#

Yes. Start it with --host 0.0.0.0 and set BASAL_API_KEY; callers then send Authorization: Bearer <key>. Without a key, other machines can make decisions but cannot reach history, templates or settings. The studio speaks plain HTTP and has one workspace; to serve beyond a trusted network, put it behind a reverse proxy that adds HTTPS.

Can a web page call it from the browser?#

Only pages you allow. Other websites open in your browser cannot send the studio requests or change anything. List your own web app's origin in BASAL_CORS_ORIGINS when starting the studio.

The app#

How do I update?#

Install the new version over the old one: download it from the latest release, or run the one-line install command again, then quit and reopen the app. The engine, settings, models, templates and history are kept. The app does not update itself yet.

How do I uninstall it?#

Remove the app as you would any other. The engine and your data are in the app's application-data folder (ai.bud.decisionstudio, listed on Data and storage); delete it to remove them too. Models are in the Hugging Face cache, ~/.cache/huggingface/hub, which other tools may share; delete them on the Models page before uninstalling, or remove their folders from the cache.

Something is not working.#

Troubleshooting lists the studio's messages with their causes and fixes. If yours is not there, report a problem.