Glossary

PFN Studio uses a small, deliberate vocabulary. Most of it comes straight from the PFN research literature; some pieces are studio-specific framing.

Core terms

Brain — a trained PFN. The artifact you end up with. In research papers this is “the model” or “the PFN”.

Prior — the recipe that generates synthetic practice tasks the brain trains on. A prior is a Python function that produces (X, y) (or (X, labels)) batches that look like the kind of problem the brain should solve. The brain never sees your real data during training — it learns from millions of synthetic tasks drawn from the prior, then generalises to your real data in-context at inference time.

Model — the brain’s neural architecture (blocks list + output heads + input shape). In basic mode, the wizard derives this from your selected capabilities and a recipe (Quick / Steady / Deep). In developer mode you can compose it block-by-block.

Eval — a scorer that runs after training to measure how well the brain learnt. Each eval has a metric (RMSE, AUROC, accuracy), a dataset, and zero or more baselines. The wizard’s “Goal” step picks evals; developer mode lets you author them directly.

Run — a training job. Pulls a prior + model + evals together with hyperparameters (lr, batch size, steps) and a compute target (local CPU or Vast.ai GPU). The run produces a trained brain on success.

Capability — plain-English description of what the brain can do — predict a number, forecast forward, show confidence, etc. The wizard’s primary unit of selection in step 1. See the Capabilities reference.

Speciality — an internal grouping of priors by tensor shape (regression_1d, regression_vard, classification, temporal_tf). Mostly invisible to basic-mode users; it’s how the wizard decides which priors can be mixed.

Wizard-specific terms

Mixture training — training a single brain on multiple priors at once so it learns multiple capabilities. Only works between priors with the same speciality (matching tensor shape). The wizard greys out cross-speciality combinations.

Specialist vs generalist — a specialist prior covers exactly one capability; a generalist covers several. The wizard’s auto-resolver prefers specialists when you ask for one capability — so picking Forecast forward alone gives you TabPFN-TS (forecast + confidence) rather than ifBO (forecast + 3 others). See Capabilities reference for the rules.

Paper-pinned — when you pick exactly one paper-backed prior, the wizard reproduces the paper’s study verbatim. Model spec, hyperparameters, eval suite, and prior overrides all come from the paper — the wizard’s recipe sliders are not applied. See Paper reproductions.

In-context learning — at inference time, the brain takes a small context set of examples and predicts for query rows without any further training. The defining feature of PFNs.

Training & compute

Quick taste — ~30 seconds, CPU. End-to-end smoke test; not enough for real learning.

Standard — ~5 minutes, CPU. Default for a small brain.

Long study — ~30 minutes, CPU. Larger brains start to shine here.

Marathon — a few hours, GPU. Needed for paper-scale brains. Runs on a Vast.ai rental.

Recipe — preset architecture size: Quick learner (2 layers, width 32), Steady thinker (3 / 64), Deep thinker (6 / 128). Steady is the default.

Marketplace

Listing — a publicly browsable prior, model, or brain. Each listing has a category, paper citation if applicable, and an install/fork CTA.

Install — copy a marketplace listing into one of your projects so you can train against it.

Fork — create a new project seeded from a marketplace listing, so you can tweak its parameters or retrain from scratch.

Compat check — when installing a model listing into a project, the marketplace checks whether the project has a prior in a matching category. A green chip says “matches your <X> prior”; an amber chip warns “install a <X> prior next.”

Data terms

Context — the example rows the brain reads in-context to learn the current task. Shape varies by speciality.

Query — the rows you want predictions for.

Reference set — synonym for context, more common in the PFN literature.

Dataset — a benchmark dataset registered in the workspace, used by evals. Cached per-workspace so it downloads once and reuses.

d_in — number of input features per row. The brain’s model spec pins this; the Try-it form validates rows against it.

Compute & deployment

Workspace / org — your team’s container. Holds projects, datasets, API tokens, Vast.ai credentials.

Vast.ai — the third-party GPU rental marketplace PFN Studio dispatches Marathon runs to. See Vast.ai setup.

API token — personal access token for the CLI (pfnstudio push) and direct API calls. See API tokens.

Share link — a public, no-auth URL that lets anyone Try the brain on their own data without an account. See Try it on your data.