Without a target
Anomalies, gaps, synthetic rows and segments, with foundation models and no labels. They run with the open source SDK on your machine today, and each one compares the foundation models with the classic method on the same data.
Run them on your data
pip install zerograd: local, no API key, nothing leaves your machine.
import zerograd
from zerograd import ZeroGrad
zg = ZeroGrad() # local engine
zerograd.list_tasks() # every task and which models run itAnomaly detection
Every row scored by how unusual it is, the top ones flagged, and the columns that make each one unusual.
For. Unusual transactions, accounting entries, sensor readings.
TabPFN v2, TabICLv2, FoMo-0D · against Isolation Forest, LOF and kNNDocs →
found = zg.anomalies(transactions)
found.top(10) # rows, values and why
zg.anomalies(transactions, labels="fraud")Imputation
A championship on known cells we hide, then the champion fills the real gaps, from the rest of each row.
For. Incomes missing from applications, incomplete product catalogs.
TabPFN v2, TabICLv2 · against mean/mode, KNN and iterative imputationDocs →
filled = zg.impute(customers)
filled.evaluation # error per method
filled.data # gaps filledSynthetic data
New rows with your columns and relations, scored by how hard they are to tell apart from yours.
For. Test data for pipelines, examples to share, more of a rare case.
TabPFN v2, TabICLv2 · against sampling each column on its ownDocs →
fake = zg.synthesize(customers, n=5_000)
fake.evaluation # detection AUC, copies
fake.dataClustering
Segments with no labels, and what sets each one apart from the rest.
For. Customer segments, merchant types, usage patterns.
ZEUS · against k-means and a Gaussian mixtureDocs →
groups = zg.cluster(customers) # or k=4
groups.profiles # what sets each apart
groups.labelsEvery model here allows commercial use: TabPFN v2 (Built with PriorLabs-TabPFN), TabICLv2 (BSD-3), FoMo-0D (MIT) and ZEUS (Apache 2.0). On the platform these tasks are open to design partners first.