Models
One API for
tabular foundation models.
Each pretrained prior-fitted network below answers through the same two calls. capabilities lists the tasks it serves and the payload each takes. predict takes { task, inputs } and returns the answer. Switch models by changing the slug, not your code.
What goes in inputs differs per model. --input sample runs the sample input pfnstudio ships for it and prints the file's path: copy it, put your data in, pass --input your.json. The first run asks before installing PyTorch and the model's packages.
$ pip install pfnstudio
$ pfnstudio run <model> --input <your file>
# ex:
$ pfnstudio run ceptron-xs-discovery --input sample
Choose this for a first causal map of a cross-sectional / i.i.d. table (rows are independent samples, not a time series): which variables cause which, in one forward pass. It can also impute missing values. Inference-only.
Choose this to forecast your own time series without training. Inference-only.
Choose this to forecast your own time series without training. Inference-only.
Choose this to forecast your own time series without training, using Google's TimesFM. Inference-only.
Choose this to forecast your own time series without training. Inference-only.
Choose this for a first look at which variables of a time series drive which, with lags. Inference-only. Read the probabilities, not only the list: on very small graphs it tends to mark one link too many.
Choose this when the signal for a prediction lives across several related tables joined by foreign keys, and you want a fast zero-shot baseline. Provide the tables, their relationships, and the target (table + id + label column). Inference-only.
Choose this to test causal-effect estimation on your own observational data: covariates, a treatment/decision column, and an outcome. It returns a per-unit effect (CATE) and an average effect (ATE) in-context. Inference-only.
Choose this to continue pretraining or fine-tune an interventional / treatment-effect model on your own data. The target model must be architecture-compatible with Do-PFN (or supply a keymap); the trainer verifies the weights load before continuing.