Download and cache pretrained TabPFN weights

brulee_tab_pfn() needs pretrained weights that are not shipped with the package. tab_pfn_download_weights() downloads them from Prior Labs' Hugging Face repositories into the local cache. tab_pfn_weights_available() reports whether the cache already holds them.

Usage

tab_pfn_download_weights(
  version = NULL,
  task = c("classification", "regression"),
  files = c("default", "all"),
  cache_dir = tabpfn_cache_dir()
)

tab_pfn_weights_available(
  version = NULL,
  task = c("classification", "regression"),
  cache_dir = tabpfn_cache_dir()
)

Arguments

version

A model version from tab_pfn_versions(). A number such as 3.5 is also accepted. The default is the newest version.

task

The task(s), one or both of "classification" and "regression". Both by default. (v3.5 uses one checkpoint for both.)

files

"default" downloads each task's default checkpoint; "all" also downloads the alternative checkpoints.

cache_dir

The root of the local weight cache.

Value

tab_pfn_download_weights() invisibly returns the paths of the checkpoints. tab_pfn_weights_available() returns a single logical: TRUE when the checkpoints for every task in task are cached.

Details

Files already cached (by brulee, or by the Python tabpfn package) are checked and skipped, so re-running does not download again. Attaching brulee never downloads the weights. If brulee_tab_pfn() is run before they are cached, it asks to download them in an interactive session and errors, pointing here, otherwise.

tab_pfn_weights_available() only looks for files of the expected size; brulee_tab_pfn() verifies their checksums when it first loads them.

Cache location

The weights are stored in a tabpfn directory of brulee's per-user cache, tools::R_user_dir()("brulee", "cache"). Set the brulee.tabpfn_cache_dir option to use another directory. Weights that the Python tabpfn package has already downloaded (into its own cache, or the directory in the TABPFN_MODEL_CACHE_DIR environment variable) are used from there without copying, after their checksum is verified. Use tab_pfn_clear_cache() to delete the weights brulee downloaded.

Notes about where the weights come from (downloaded, already cached, or reused from the Python package) appear only in interactive sessions, at most once a day for each file.

License

The weights are released by Prior Labs under a non-commercial license that must be accepted once before the first download. The first download asks you to log in to Prior Labs, accept the license, and paste your API key. The key is stored in ~/.cache/tabpfn/auth_token, the same file the Python tabpfn package uses. In non-interactive sessions, set the TABPFN_TOKEN environment variable to your API key instead. See brulee_tab_pfn() for the full setup.

Examples

if (FALSE) {
tab_pfn_download_weights("v3.5")
tab_pfn_weights_available("v3.5")
tab_pfn_weights_available("v3", task = "regression")
}