Bud Decision Studio 0.3.0

Teach a model your own decisions, and twelve fixes, one of them for privacy.

New: the Train page

• Teach any of the eleven models from a spreadsheet of your past decisions, on your own computer's GPU. The studio works out the questions from the file, recommends a model with a time estimate, trains it, and shows the result on examples it never saw, with one it used to get wrong.
• Nothing worse is ever kept. Each training replays general decisions so the model keeps what it knew. The trained model is kept only if it is better on your held-out examples, no worse on general questions, not answering everything the same way, and identical once saved and reloaded.
• Trained models appear with the others in the Playground, Models, Evaluate and your code. Each one is a small file of changes on disk and needs no extra memory. Export and Import a trained model move one between computers.
• Training is safe to leave running. One training at a time, in its own process, within a memory limit, at low priority. It waits when other programs need the memory, can be paused and continued, and restarts itself if the GPU stops responding.
• The training API is under /api/training and /api/finetunes.
• CLM 8B gives the same answer to the same request every time. Before, an answer could move by up to 0.06 depending on what was cached.
• Intern-Decision 4B and Lev load straight onto the GPU, without a temporary second copy of the weights in memory (about 9 GB on an NVIDIA GB10).

Training has been verified on an NVIDIA GB10 (Linux), with accepted runs on all eleven models. It has not yet been run on Windows, Apple Silicon or discrete NVIDIA cards; AMD and Intel GPUs are experimental and off by default (Settings, "Allow experimental training").

Fixes

• A sensitive file is never kept (privacy). An image, audio or video variable marked sensitive was stored with the decision: its contents, its file name and its content hash. The model now reads it and History keeps only a keyed hash, as for every other sensitive value. If you used sensitive file variables on 0.2.x, those earlier decisions still hold the files; delete or redact them in History.
• Background decisions no longer report a false save failure. Each one came back with a history_write_failed warning although it was saved.
• A template's default model is no longer shown as unable to run it. A variable that may hold more text than a model reads is now a note, not a blocker.
• A version that narrows a variable is marked breaking, because it can refuse values callers already send.
• A variable with a default is never required.
• Cancelling a background decision stops the model load it started, unless another request is waiting for that model or you loaded it yourself.
• Evaluate recommends an act threshold only within 50% to 99%, the range the Playground's slider can show.
• Save as template in a fresh window saves the model the Playground uses, instead of no default model.
• The template id and alias fields are checked in the browser again.
• Leaving the Playground straight after opening it no longer throws an error.
• A second studio started with BASAL_NO_DOWNLOADS=1 explains why it cannot download, instead of failing with an error 500.
• The environment check reports a GPU that is out of memory as that, not as "PyTorch failed to import".

Documentation

The repository now has a full documentation site in site/docs/: installation, a user manual for every page, concepts, guides, the API reference and developer docs. Open site/docs/index.html, or serve the site/ folder.

Update

Install 0.3.0 over the app: download the file for your computer below, or run the one-line command again (curl -fsSL https://raw.githubusercontent.com/BudEcosystem/Bud-Decision-Engine/main/get.sh | sh on macOS and Linux, irm https://raw.githubusercontent.com/BudEcosystem/Bud-Decision-Engine/main/get.ps1 | iex in PowerShell on Windows), then quit and reopen the app. Your engine, settings, models, templates and history are kept.