Prediction

RelationalTransformer.predict returns sigmoid probabilities for a classification checkpoint and raw normalized values for a regression checkpoint. Pass activation="identity" to obtain classification logits.

from relational_transformers import RelationalTransformer

classifier = RelationalTransformer("RelativeDB/rt-j-fp16")
probabilities = classifier.predict(batch)

regressor = RelationalTransformer("RelativeDB/rt-j-fp16", task="regression")
normalized_values = regressor.predict(batch)

Activations

predict picks a default activation from the problem type and lets you override it:

Problem type	Default activation	Override
classification, binary, multilabel, clf	sigmoid	activation="identity" for logits
multiclass (fitted task heads)	softmax	activation="identity" for logits
regression, forecasting	identity	none needed

Ranking pipelines usually want logits, since sigmoid preserves order and calibration tooling expects the raw score. Set convert_to_numpy=False to keep results as torch tensors on the model’s device.

Predicting with a Fitted Task Head

After fit_head stores a head under a task name, route predictions through it:

model.fit_head(examples, task="issue_label", num_labels=5, problem_type="multiclass")
distribution = model.predict(batch, task_head="issue_label")

The head runs over frozen target features. The published scalar decoder stays available by omitting task_head.

Contextual Cell Embeddings

model.encode(context) returns one contextualized state per cell, shaped [batch, cells, 512] for the published models. Pass output_value="target_features" for the summed [batch, 512] target representation that task heads train on.

states = model.encode(batch)                                    # [B, S, 512]
features = model.encode(batch, output_value="target_features")  # [B, 512]

Model Outputs

Use forward when a consumer needs a lower-level output than predict exposes:

output = model.forward(batch, output="target_features")
features = output.features    # [batch, 512]

output	Field on ModelOutput	Shape	Meaning
target_scores	scores	[batch]	one scalar per context from the number decoder
token_scores	token_scores	[batch, cells]	the number-decoder score at every cell, in caller order
target_features	features	[batch, 512]	summed target state used for head tuning
target_scores_and_text	scores, target_text	[batch], [batch, 384]	scalar scores plus the raw text-decoder output at the target
embeddings	embeddings	[batch, cells, 512]	contextualized state for every cell

Every field comes back as a torch tensor on all backends. target_text is the raw decoder output; nothing normalizes it, so downstream nearest-neighbor lookups decide their own similarity convention.

Note

The ONNX and Triton backends serve target_scores only. Requesting any other output on those backends raises ValueError, and the meta backend raises on any inference call because it has no weights.