Computing Predictions¶
Prediction¶
model.predict() accepts a RelationalBatch, a canonical tensor mapping, a
simple all-text cell-vector array, or a sequence of one-context inputs.
Contextual Cell Embeddings¶
model.encode(context) returns one contextual state per cell. Use
output_value="target_features" to return the summed target representation
used by task heads.
Classification and Regression¶
Classification applies sigmoid by default. Multiclass task heads apply
softmax. Regression and forecasting return raw scalar outputs. Override with
activation="identity" whenever calibration or ranking consumes logits.
Batching¶
Variable-length contexts are padded during sequence collation. All contexts must use the same text embedding width and model contract.
Ablation¶
Call batch.ablate(positions) to mask caller-selected cells without changing
stable node or parent IDs. Compare the resulting prediction with the original;
the library never chooses an ablation on the user’s behalf.