You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
TabICL classification predict returns numeric labels without return_proba, strings with it #45
The type of predictions in a classification predict depends on the backend and on return_proba:
Model
return_proba: true
return_proba: false
tabicl-v2
["1.0"] (strings)
[1.0] (numbers)
tabpfn-3.5
["1.0"]
["1.0"]
Seen on the prod A40 deployment through the gateway on 2026-10-20, fitting train_y: [0, 1, 2] and predicting one row. A client that compares predictions with classes, or switches return_proba on and off, gets different types for the same answer. The gateway docs now warn about it (VectorInstitute/aieng-proxy-gateway#471).
Cause
TabICLBackend._predict (src/tabctx/backends/tabicl.py:530) returns model.predict(X_test).tolist() when return_proba is false, which keeps the label dtype (float for JSON numbers). With return_proba it derives labels from predict_proba as str(c) for c in model.classes_. The TabPFN backend (src/tabctx/backends/tabpfn.py:186) always takes the second route.
The TabICL route also goes against the backend protocol in backends/base.py, which says to derive predictions from probabilities rather than calling predict().
Fix
Make TabICL do what TabPFN does: always call predict_proba, take the argmax over [str(c) for c in model.classes_], and return probabilities only when asked. Regression stays numeric. Then add a test over both backends that classification predictions are strings drawn from classes, whatever return_proba is, and drop the exception from the gateway docs.
Labels also come back in float form ("1.0" for a JSON 1), because the JSON train_y is parsed as floats. That's a separate question: whether to keep integer labels as integers.
Activity
fix(predict): classification labels are strings whatever return_proba…
fix(predict): classification labels are strings whatever return_proba…