bead 0.1.0__py3-none-any.whl

bead/active_learning/models/categorical.py

@@ -0,0 +1,943 @@
+"""Model for categorical tasks (unordered N-class classification)."""
+
+from __future__ import annotations
+
+import tempfile
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.nn as nn
+from transformers import AutoModel, AutoTokenizer, TrainingArguments
+
+from bead.active_learning.config import VarianceComponents
+from bead.active_learning.models.base import ActiveLearningModel, ModelPrediction
+from bead.active_learning.models.random_effects import RandomEffectsManager
+from bead.active_learning.trainers.data_collator import MixedEffectsDataCollator
+from bead.active_learning.trainers.dataset_utils import items_to_dataset
+from bead.active_learning.trainers.metrics import compute_multiclass_metrics
+from bead.active_learning.trainers.model_wrapper import EncoderClassifierWrapper
+from bead.config.active_learning import CategoricalModelConfig
+from bead.items.item import Item
+from bead.items.item_template import ItemTemplate, TaskType
+
+__all__ = ["CategoricalModel"]
+
+
+class CategoricalModel(ActiveLearningModel):
+    """Model for categorical tasks with N unordered categories.
+
+    Supports N-class classification (N ≥ 2) using any HuggingFace transformer
+    model. Provides two encoding strategies: single encoder (concatenate
+    categories) or dual encoder (separate embeddings).
+
+    Parameters
+    ----------
+    config : CategoricalModelConfig
+        Configuration object containing all model parameters.
+
+    Attributes
+    ----------
+    config : CategoricalModelConfig
+        Model configuration.
+    tokenizer : AutoTokenizer
+        Transformer tokenizer.
+    encoder : AutoModel
+        Transformer encoder model.
+    classifier_head : nn.Sequential
+        Classification head (fixed effects head).
+    num_classes : int | None
+        Number of classes (inferred from training data).
+    category_names : list[str] | None
+        Category names (e.g., ["entailment", "neutral", "contradiction"]).
+    random_effects : RandomEffectsManager
+        Manager for participant-level random effects.
+    variance_history : list[VarianceComponents]
+        Variance component estimates over training (for diagnostics).
+    _is_fitted : bool
+        Whether model has been trained.
+
+    Examples
+    --------
+    >>> from uuid import uuid4
+    >>> from bead.items.item import Item
+    >>> from bead.config.active_learning import CategoricalModelConfig
+    >>> items = [
+    ...     Item(
+    ...         item_template_id=uuid4(),
+    ...         rendered_elements={"premise": "sent A", "hypothesis": "sent B"}
+    ...     )
+    ...     for _ in range(10)
+    ... ]
+    >>> labels = ["entailment"] * 5 + ["contradiction"] * 5
+    >>> config = CategoricalModelConfig(  # doctest: +SKIP
+    ...     num_epochs=1, batch_size=2, device="cpu"
+    ... )
+    >>> model = CategoricalModel(config=config)  # doctest: +SKIP
+    >>> metrics = model.train(items, labels, participant_ids=None)  # doctest: +SKIP
+    >>> predictions = model.predict(items[:3], participant_ids=None)  # doctest: +SKIP
+    """
+
+    def __init__(
+        self,
+        config: CategoricalModelConfig | None = None,
+    ) -> None:
+        """Initialize categorical model.
+
+        Parameters
+        ----------
+        config : CategoricalModelConfig | None
+            Configuration object. If None, uses default configuration.
+        """
+        self.config = config or CategoricalModelConfig()
+
+        # Validate mixed_effects configuration
+        super().__init__(self.config)
+
+        self.tokenizer = AutoTokenizer.from_pretrained(self.config.model_name)
+        self.encoder = AutoModel.from_pretrained(self.config.model_name)
+
+        self.num_classes: int | None = None
+        self.category_names: list[str] | None = None
+        self.classifier_head: nn.Sequential | None = None
+        self._is_fitted = False
+
+        # Initialize random effects manager
+        self.random_effects: RandomEffectsManager | None = None
+        self.variance_history: list[VarianceComponents] = []
+
+        self.encoder.to(self.config.device)
+
+    @property
+    def supported_task_types(self) -> list[TaskType]:
+        """Get supported task types.
+
+        Returns
+        -------
+        list[TaskType]
+            List containing "categorical".
+        """
+        return ["categorical"]
+
+    def validate_item_compatibility(
+        self, item: Item, item_template: ItemTemplate
+    ) -> None:
+        """Validate item is compatible with categorical model.
+
+        Parameters
+        ----------
+        item : Item
+            Item to validate.
+        item_template : ItemTemplate
+            Template the item was constructed from.
+
+        Raises
+        ------
+        ValueError
+            If task_type is not "categorical".
+        """
+        if item_template.task_type != "categorical":
+            raise ValueError(
+                f"Expected task_type 'categorical', got '{item_template.task_type}'"
+            )
+
+    def _initialize_classifier(self, num_classes: int) -> None:
+        """Initialize classification head for given number of classes.
+
+        Parameters
+        ----------
+        num_classes : int
+            Number of output classes.
+        """
+        hidden_size = self.encoder.config.hidden_size
+
+        if self.config.encoder_mode == "dual_encoder":
+            input_size = hidden_size * num_classes
+        else:
+            input_size = hidden_size
+
+        self.classifier_head = nn.Sequential(
+            nn.Linear(input_size, 256),
+            nn.ReLU(),
+            nn.Dropout(0.1),
+            nn.Linear(256, num_classes),
+        )
+        self.classifier_head.to(self.config.device)
+
+    def _encode_single(self, texts: list[str]) -> torch.Tensor:
+        """Encode texts using single encoder strategy.
+
+        Concatenates all category texts with [SEP] tokens and encodes once.
+
+        Parameters
+        ----------
+        texts : list[str]
+            List of concatenated category texts for each item.
+
+        Returns
+        -------
+        torch.Tensor
+            Encoded representations of shape (batch_size, hidden_size).
+        """
+        encodings = self.tokenizer(
+            texts,
+            padding=True,
+            truncation=True,
+            max_length=self.config.max_length,
+            return_tensors="pt",
+        )
+        encodings = {k: v.to(self.config.device) for k, v in encodings.items()}
+
+        outputs = self.encoder(**encodings)
+        return outputs.last_hidden_state[:, 0, :]
+
+    def _encode_dual(self, categories_per_item: list[list[str]]) -> torch.Tensor:
+        """Encode texts using dual encoder strategy.
+
+        Encodes each category separately and concatenates embeddings.
+
+        Parameters
+        ----------
+        categories_per_item : list[list[str]]
+            List of category lists. Each inner list contains category texts
+            for one item.
+
+        Returns
+        -------
+        torch.Tensor
+            Concatenated encodings of shape (batch_size, hidden_size * num_categories).
+        """
+        all_embeddings = []
+
+        for categories in categories_per_item:
+            category_embeddings = []
+            for category_text in categories:
+                encodings = self.tokenizer(
+                    [category_text],
+                    padding=True,
+                    truncation=True,
+                    max_length=self.config.max_length,
+                    return_tensors="pt",
+                )
+                encodings = {k: v.to(self.config.device) for k, v in encodings.items()}
+
+                outputs = self.encoder(**encodings)
+                cls_embedding = outputs.last_hidden_state[0, 0, :]
+                category_embeddings.append(cls_embedding)
+
+            concatenated = torch.cat(category_embeddings, dim=0)
+            all_embeddings.append(concatenated)
+
+        return torch.stack(all_embeddings)
+
+    def _prepare_inputs(self, items: list[Item]) -> torch.Tensor:
+        """Prepare inputs for encoding based on encoder mode.
+
+        For categorical tasks, concatenates all rendered elements.
+
+        Parameters
+        ----------
+        items : list[Item]
+            Items to encode.
+
+        Returns
+        -------
+        torch.Tensor
+            Encoded representations.
+        """
+        if self.category_names is None:
+            raise ValueError("Model not initialized. Call train() first.")
+
+        if self.config.encoder_mode == "single_encoder":
+            texts = []
+            for item in items:
+                # Concatenate all rendered elements
+                all_text = " ".join(item.rendered_elements.values())
+                texts.append(all_text)
+            return self._encode_single(texts)
+        else:
+            categories_per_item = []
+            for item in items:
+                category_texts = list(item.rendered_elements.values())
+                categories_per_item.append(category_texts)
+            return self._encode_dual(categories_per_item)
+
+    def _validate_labels(self, labels: list[str]) -> None:
+        """Validate that all labels are valid category names.
+
+        Parameters
+        ----------
+        labels : list[str]
+            Labels to validate.
+
+        Raises
+        ------
+        ValueError
+            If any label is not in category_names.
+        """
+        if self.category_names is None:
+            raise ValueError("category_names not initialized")
+
+        valid_labels = set(self.category_names)
+        invalid = [label for label in labels if label not in valid_labels]
+        if invalid:
+            raise ValueError(
+                f"Invalid labels found: {set(invalid)}. "
+                f"Labels must be one of {valid_labels}."
+            )
+
+    def _prepare_training_data(
+        self,
+        items: list[Item],
+        labels: list[str],
+        participant_ids: list[str],
+        validation_items: list[Item] | None,
+        validation_labels: list[str] | None,
+    ) -> tuple[list[Item], list[int], list[str], list[Item] | None, list[int] | None]:
+        """Prepare training data for categorical model.
+
+        Parameters
+        ----------
+        items : list[Item]
+            Training items.
+        labels : list[str]
+            Training labels.
+        participant_ids : list[str]
+            Normalized participant IDs.
+        validation_items : list[Item] | None
+            Validation items.
+        validation_labels : list[str] | None
+            Validation labels.
+
+        Returns
+        -------
+        tuple[list[Item], list[int], list[str], list[Item] | None, list[int] | None]
+            Prepared items, numeric labels, participant_ids, validation_items,
+            numeric validation_labels.
+        """
+        unique_labels = sorted(set(labels))
+        self.num_classes = len(unique_labels)
+        self.category_names = unique_labels
+
+        self._validate_labels(labels)
+        self._initialize_classifier(self.num_classes)
+
+        label_to_idx = {label: idx for idx, label in enumerate(self.category_names)}
+        y_numeric = [label_to_idx[label] for label in labels]
+
+        # Convert validation labels if provided
+        val_y_numeric = None
+        if validation_items is not None and validation_labels is not None:
+            self._validate_labels(validation_labels)
+            if len(validation_items) != len(validation_labels):
+                raise ValueError(
+                    f"Number of validation items ({len(validation_items)}) "
+                    f"must match number of validation labels ({len(validation_labels)})"
+                )
+            val_y_numeric = [label_to_idx[label] for label in validation_labels]
+
+        return items, y_numeric, participant_ids, validation_items, val_y_numeric
+
+    def _initialize_random_effects(self, n_classes: int) -> None:
+        """Initialize random effects manager.
+
+        Parameters
+        ----------
+        n_classes : int
+            Number of classes.
+        """
+        self.random_effects = RandomEffectsManager(
+            self.config.mixed_effects, n_classes=n_classes
+        )
+
+    def _do_training(
+        self,
+        items: list[Item],
+        labels_numeric: list[int],
+        participant_ids: list[str],
+        validation_items: list[Item] | None,
+        validation_labels_numeric: list[int] | None,
+    ) -> dict[str, float]:
+        """Perform categorical model training.
+
+        Parameters
+        ----------
+        items : list[Item]
+            Training items.
+        labels_numeric : list[int]
+            Numeric labels (class indices).
+        participant_ids : list[str]
+            Participant IDs.
+        validation_items : list[Item] | None
+            Validation items.
+        validation_labels_numeric : list[int] | None
+            Numeric validation labels.
+
+        Returns
+        -------
+        dict[str, float]
+            Training metrics.
+        """
+        # Convert validation_labels_numeric back to string labels for validation metrics
+        validation_labels = None
+        if validation_items is not None and validation_labels_numeric is not None:
+            validation_labels = [
+                self.category_names[label_idx]
+                for label_idx in validation_labels_numeric
+            ]
+
+        # Use HuggingFace Trainer for fixed and random_intercepts modes
+        if self.config.mixed_effects.mode in ("fixed", "random_intercepts"):
+            metrics = self._train_with_huggingface_trainer(
+                items=items,
+                y_numeric=labels_numeric,
+                participant_ids=participant_ids,
+                validation_items=validation_items,
+                validation_labels=validation_labels,
+            )
+        else:
+            # Use custom loop for random_slopes mode
+            metrics = self._train_with_custom_loop(
+                items=items,
+                y_numeric=labels_numeric,
+                participant_ids=participant_ids,
+                validation_items=validation_items,
+                validation_labels=validation_labels,
+            )
+
+        # Add validation accuracy if validation data provided and not already computed
+        if (
+            validation_items is not None
+            and validation_labels is not None
+            and "val_accuracy" not in metrics
+        ):
+            # Validation with placeholder participant_ids for mixed effects
+            if self.config.mixed_effects.mode == "fixed":
+                val_participant_ids = ["_fixed_"] * len(validation_items)
+            else:
+                val_participant_ids = ["_validation_"] * len(validation_items)
+            val_predictions = self._do_predict(validation_items, val_participant_ids)
+            val_pred_labels = [p.predicted_class for p in val_predictions]
+            val_acc = sum(
+                pred == true
+                for pred, true in zip(val_pred_labels, validation_labels, strict=True)
+            ) / len(validation_labels)
+            metrics["val_accuracy"] = val_acc
+
+        return metrics
+
+    def _train_with_huggingface_trainer(
+        self,
+        items: list[Item],
+        y_numeric: list[int],
+        participant_ids: list[str],
+        validation_items: list[Item] | None,
+        validation_labels: list[str] | None,
+    ) -> dict[str, float]:
+        """Train using HuggingFace Trainer with mixed effects support.
+
+        Parameters
+        ----------
+        items : list[Item]
+            Training items.
+        y_numeric : list[int]
+            Numeric labels (class indices).
+        participant_ids : list[str]
+            Participant IDs.
+        validation_items : list[Item] | None
+            Validation items.
+        validation_labels : list[str] | None
+            Validation labels.
+
+        Returns
+        -------
+        dict[str, float]
+            Training metrics.
+        """
+        # Convert items to HuggingFace Dataset
+        train_dataset = items_to_dataset(
+            items=items,
+            labels=y_numeric,
+            participant_ids=participant_ids,
+            tokenizer=self.tokenizer,
+            max_length=self.config.max_length,
+        )
+
+        # Create validation dataset if provided
+        eval_dataset = None
+        if validation_items is not None and validation_labels is not None:
+            label_to_idx = {label: idx for idx, label in enumerate(self.category_names)}
+            val_y_numeric = [label_to_idx[label] for label in validation_labels]
+            val_participant_ids = (
+                ["_validation_"] * len(validation_items)
+                if self.config.mixed_effects.mode != "fixed"
+                else ["_fixed_"] * len(validation_items)
+            )
+            eval_dataset = items_to_dataset(
+                items=validation_items,
+                labels=val_y_numeric,
+                participant_ids=val_participant_ids,
+                tokenizer=self.tokenizer,
+                max_length=self.config.max_length,
+            )
+
+        # Create wrapper model for Trainer
+        wrapped_model = EncoderClassifierWrapper(
+            encoder=self.encoder, classifier_head=self.classifier_head
+        )
+
+        # Create data collator
+        data_collator = MixedEffectsDataCollator(tokenizer=self.tokenizer)
+
+        # Create metrics computation function
+        def compute_metrics_fn(eval_pred: object) -> dict[str, float]:
+            return compute_multiclass_metrics(eval_pred, num_labels=self.num_classes)
+
+        # Create training arguments with checkpointing
+        with tempfile.TemporaryDirectory() as tmpdir:
+            checkpoint_dir = Path(tmpdir) / "checkpoints"
+            checkpoint_dir.mkdir(parents=True, exist_ok=True)
+
+            training_args = TrainingArguments(
+                output_dir=str(checkpoint_dir),
+                num_train_epochs=self.config.num_epochs,
+                per_device_train_batch_size=self.config.batch_size,
+                per_device_eval_batch_size=self.config.batch_size,
+                learning_rate=self.config.learning_rate,
+                logging_steps=10,
+                eval_strategy="epoch" if eval_dataset is not None else "no",
+                save_strategy="epoch",
+                save_total_limit=1,
+                load_best_model_at_end=False,
+                report_to="none",
+                remove_unused_columns=False,
+                use_cpu=self.config.device == "cpu",
+            )
+
+            # Import here to avoid circular import
+            from bead.active_learning.trainers.mixed_effects import (  # noqa: PLC0415
+                MixedEffectsTrainer,
+            )
+
+            # Create trainer
+            trainer = MixedEffectsTrainer(
+                model=wrapped_model,
+                args=training_args,
+                train_dataset=train_dataset,
+                eval_dataset=eval_dataset,
+                data_collator=data_collator,
+                tokenizer=self.tokenizer,
+                random_effects_manager=self.random_effects,
+                compute_metrics=compute_metrics_fn,
+            )
+
+            # Train
+            train_result = trainer.train()
+
+            # Get training metrics
+            train_metrics = trainer.evaluate(eval_dataset=train_dataset)
+            metrics: dict[str, float] = {
+                "train_loss": float(train_result.training_loss),
+                "train_accuracy": train_metrics.get("eval_accuracy", 0.0),
+                "train_precision": train_metrics.get("eval_precision", 0.0),
+                "train_recall": train_metrics.get("eval_recall", 0.0),
+                "train_f1": train_metrics.get("eval_f1", 0.0),
+            }
+
+            # Get validation metrics if eval_dataset was provided
+            if eval_dataset is not None:
+                val_metrics = trainer.evaluate(eval_dataset=eval_dataset)
+                metrics.update(
+                    {
+                        "val_accuracy": val_metrics.get("eval_accuracy", 0.0),
+                        "val_precision": val_metrics.get("eval_precision", 0.0),
+                        "val_recall": val_metrics.get("eval_recall", 0.0),
+                        "val_f1": val_metrics.get("eval_f1", 0.0),
+                    }
+                )
+
+        # Estimate variance components
+        if self.config.mixed_effects.estimate_variance_components:
+            var_comps = self.random_effects.estimate_variance_components()
+            if var_comps:
+                var_comp = var_comps.get("mu") or var_comps.get("slopes")
+                if var_comp:
+                    self.variance_history.append(var_comp)
+                    metrics["participant_variance"] = var_comp.variance
+                    metrics["n_participants"] = var_comp.n_groups
+
+        self._is_fitted = True
+
+        return metrics
+
+    def _train_with_custom_loop(
+        self,
+        items: list[Item],
+        y_numeric: list[int],
+        participant_ids: list[str],
+        validation_items: list[Item] | None,
+        validation_labels: list[str] | None,
+    ) -> dict[str, float]:
+        """Train using custom training loop (for random_slopes mode).
+
+        Parameters
+        ----------
+        items : list[Item]
+            Training items.
+        y_numeric : list[int]
+            Numeric labels (class indices).
+        participant_ids : list[str]
+            Participant IDs.
+        validation_items : list[Item] | None
+            Validation items.
+        validation_labels : list[str] | None
+            Validation labels.
+
+        Returns
+        -------
+        dict[str, float]
+            Training metrics.
+        """
+        # Convert to tensor
+        y = torch.tensor(y_numeric, dtype=torch.long, device=self.config.device)
+
+        # Build optimizer parameters
+        params_to_optimize = list(self.encoder.parameters()) + list(
+            self.classifier_head.parameters()
+        )
+
+        # Add random effects parameters (for random_slopes)
+        if self.config.mixed_effects.mode == "random_slopes":
+            for head in self.random_effects.slopes.values():
+                params_to_optimize.extend(head.parameters())
+
+        optimizer = torch.optim.AdamW(params_to_optimize, lr=self.config.learning_rate)
+        criterion = nn.CrossEntropyLoss()
+
+        self.encoder.train()
+        self.classifier_head.train()
+
+        for _epoch in range(self.config.num_epochs):
+            n_batches = (
+                len(items) + self.config.batch_size - 1
+            ) // self.config.batch_size
+            epoch_loss = 0.0
+            epoch_correct = 0
+
+            for i in range(n_batches):
+                start_idx = i * self.config.batch_size
+                end_idx = min(start_idx + self.config.batch_size, len(items))
+
+                batch_items = items[start_idx:end_idx]
+                batch_labels = y[start_idx:end_idx]
+                batch_participant_ids = participant_ids[start_idx:end_idx]
+
+                embeddings = self._prepare_inputs(batch_items)
+
+                # Forward pass depends on mixed effects mode
+                if self.config.mixed_effects.mode == "fixed":
+                    # Standard forward pass
+                    logits = self.classifier_head(embeddings)
+
+                elif self.config.mixed_effects.mode == "random_intercepts":
+                    # Fixed head + per-participant bias
+                    logits = self.classifier_head(embeddings)
+                    for j, pid in enumerate(batch_participant_ids):
+                        bias = self.random_effects.get_intercepts(
+                            pid,
+                            n_classes=self.num_classes,
+                            param_name="mu",
+                            create_if_missing=True,
+                        )
+                        logits[j] = logits[j] + bias
+
+                elif self.config.mixed_effects.mode == "random_slopes":
+                    # Per-participant head
+                    logits_list = []
+                    for j, pid in enumerate(batch_participant_ids):
+                        participant_head = self.random_effects.get_slopes(
+                            pid,
+                            fixed_head=self.classifier_head,
+                            create_if_missing=True,
+                        )
+                        logits_j = participant_head(embeddings[j : j + 1])
+                        logits_list.append(logits_j)
+                    logits = torch.cat(logits_list, dim=0)
+
+                # Data loss + prior regularization
+                loss_ce = criterion(logits, batch_labels)
+                loss_prior = self.random_effects.compute_prior_loss()
+                loss = loss_ce + loss_prior
+
+                optimizer.zero_grad()
+                loss.backward()
+                optimizer.step()
+
+                epoch_loss += loss.item()
+                predictions = torch.argmax(logits, dim=1)
+                epoch_correct += (predictions == batch_labels).sum().item()
+
+            epoch_acc = epoch_correct / len(items)
+            epoch_loss = epoch_loss / n_batches
+
+        self._is_fitted = True
+
+        metrics: dict[str, float] = {
+            "train_accuracy": epoch_acc,
+            "train_loss": epoch_loss,
+        }
+
+        # Estimate variance components
+        if self.config.mixed_effects.estimate_variance_components:
+            var_comps = self.random_effects.estimate_variance_components()
+            if var_comps:
+                var_comp = var_comps.get("mu") or var_comps.get("slopes")
+                if var_comp:
+                    self.variance_history.append(var_comp)
+                    metrics["participant_variance"] = var_comp.variance
+                    metrics["n_participants"] = var_comp.n_groups
+
+        if validation_items is not None and validation_labels is not None:
+            self._validate_labels(validation_labels)
+
+            if len(validation_items) != len(validation_labels):
+                raise ValueError(
+                    f"Number of validation items ({len(validation_items)}) "
+                    f"must match number of validation labels ({len(validation_labels)})"
+                )
+
+            # Validation with placeholder participant_ids for mixed effects
+            if self.config.mixed_effects.mode == "fixed":
+                val_predictions = self.predict(validation_items, participant_ids=None)
+            else:
+                val_participant_ids = ["_validation_"] * len(validation_items)
+                val_predictions = self.predict(
+                    validation_items, participant_ids=val_participant_ids
+                )
+            val_pred_labels = [p.predicted_class for p in val_predictions]
+            val_acc = sum(
+                pred == true
+                for pred, true in zip(val_pred_labels, validation_labels, strict=True)
+            ) / len(validation_labels)
+            metrics["val_accuracy"] = val_acc
+
+        return metrics
+
+    def _do_predict(
+        self, items: list[Item], participant_ids: list[str]
+    ) -> list[ModelPrediction]:
+        """Perform categorical model prediction.
+
+        Parameters
+        ----------
+        items : list[Item]
+            Items to predict.
+        participant_ids : list[str]
+            Normalized participant IDs.
+
+        Returns
+        -------
+        list[ModelPrediction]
+            Predictions.
+        """
+        self.encoder.eval()
+        self.classifier_head.eval()
+
+        with torch.no_grad():
+            embeddings = self._prepare_inputs(items)
+
+            # Forward pass depends on mixed effects mode
+            if self.config.mixed_effects.mode == "fixed":
+                logits = self.classifier_head(embeddings)
+
+            elif self.config.mixed_effects.mode == "random_intercepts":
+                logits = self.classifier_head(embeddings)
+                for i, pid in enumerate(participant_ids):
+                    # Unknown participants: use prior mean (zero bias)
+                    bias = self.random_effects.get_intercepts(
+                        pid,
+                        n_classes=self.num_classes,
+                        param_name="mu",
+                        create_if_missing=False,
+                    )
+                    logits[i] = logits[i] + bias
+
+            elif self.config.mixed_effects.mode == "random_slopes":
+                logits_list = []
+                for i, pid in enumerate(participant_ids):
+                    # Unknown participants: use fixed head
+                    participant_head = self.random_effects.get_slopes(
+                        pid, fixed_head=self.classifier_head, create_if_missing=False
+                    )
+                    logits_i = participant_head(embeddings[i : i + 1])
+                    logits_list.append(logits_i)
+                logits = torch.cat(logits_list, dim=0)
+
+            proba = torch.softmax(logits, dim=1).cpu().numpy()
+            pred_classes = torch.argmax(logits, dim=1).cpu().numpy()
+
+        predictions = []
+        for i, item in enumerate(items):
+            pred_label = self.category_names[pred_classes[i]]
+            prob_dict = {
+                cat: float(proba[i, idx]) for idx, cat in enumerate(self.category_names)
+            }
+            predictions.append(
+                ModelPrediction(
+                    item_id=str(item.id),
+                    probabilities=prob_dict,
+                    predicted_class=pred_label,
+                    confidence=float(proba[i, pred_classes[i]]),
+                )
+            )
+
+        return predictions
+
+    def _do_predict_proba(
+        self, items: list[Item], participant_ids: list[str]
+    ) -> np.ndarray:
+        """Perform categorical model probability prediction.
+
+        Parameters
+        ----------
+        items : list[Item]
+            Items to predict.
+        participant_ids : list[str]
+            Normalized participant IDs.
+
+        Returns
+        -------
+        np.ndarray
+            Probability array of shape (n_items, n_classes).
+        """
+        self.encoder.eval()
+        self.classifier_head.eval()
+
+        with torch.no_grad():
+            embeddings = self._prepare_inputs(items)
+
+            # Forward pass depends on mixed effects mode
+            if self.config.mixed_effects.mode == "fixed":
+                logits = self.classifier_head(embeddings)
+
+            elif self.config.mixed_effects.mode == "random_intercepts":
+                logits = self.classifier_head(embeddings)
+                for i, pid in enumerate(participant_ids):
+                    bias = self.random_effects.get_intercepts(
+                        pid,
+                        n_classes=self.num_classes,
+                        param_name="mu",
+                        create_if_missing=False,
+                    )
+                    logits[i] = logits[i] + bias
+
+            elif self.config.mixed_effects.mode == "random_slopes":
+                logits_list = []
+                for i, pid in enumerate(participant_ids):
+                    participant_head = self.random_effects.get_slopes(
+                        pid, fixed_head=self.classifier_head, create_if_missing=False
+                    )
+                    logits_i = participant_head(embeddings[i : i + 1])
+                    logits_list.append(logits_i)
+                logits = torch.cat(logits_list, dim=0)
+
+            proba = torch.softmax(logits, dim=1).cpu().numpy()
+
+        return proba
+
+    def _get_save_state(self) -> dict[str, object]:
+        """Get model-specific state to save.
+
+        Returns
+        -------
+        dict[str, object]
+            State dictionary.
+        """
+        return {
+            "num_classes": self.num_classes,
+            "category_names": self.category_names,
+        }
+
+    def _save_model_components(self, save_path: Path) -> None:
+        """Save model-specific components.
+
+        Parameters
+        ----------
+        save_path : Path
+            Directory to save to.
+        """
+        self.encoder.save_pretrained(save_path / "encoder")
+        self.tokenizer.save_pretrained(save_path / "encoder")
+
+        torch.save(
+            self.classifier_head.state_dict(),
+            save_path / "classifier_head.pt",
+        )
+
+    def _restore_training_state(self, config_dict: dict[str, object]) -> None:
+        """Restore model-specific training state.
+
+        Parameters
+        ----------
+        config_dict : dict[str, object]
+            Configuration dictionary with training state.
+        """
+        self.num_classes = config_dict.pop("num_classes")
+        self.category_names = config_dict.pop("category_names")
+
+    def _load_model_components(self, load_path: Path) -> None:
+        """Load model-specific components.
+
+        Parameters
+        ----------
+        load_path : Path
+            Directory to load from.
+        """
+        # Load config.json to reconstruct config
+        with open(load_path / "config.json") as f:
+            import json  # noqa: PLC0415
+
+            config_dict = json.load(f)
+
+        # Reconstruct MixedEffectsConfig if needed
+        if "mixed_effects" in config_dict and isinstance(
+            config_dict["mixed_effects"], dict
+        ):
+            from bead.active_learning.config import MixedEffectsConfig  # noqa: PLC0415
+
+            config_dict["mixed_effects"] = MixedEffectsConfig(
+                **config_dict["mixed_effects"]
+            )
+
+        self.config = CategoricalModelConfig(**config_dict)
+
+        self.encoder = AutoModel.from_pretrained(load_path / "encoder")
+        self.tokenizer = AutoTokenizer.from_pretrained(load_path / "encoder")
+
+        self._initialize_classifier(self.num_classes)
+        self.classifier_head.load_state_dict(
+            torch.load(
+                load_path / "classifier_head.pt", map_location=self.config.device
+            )
+        )
+        self.classifier_head.to(self.config.device)
+
+    def _get_random_effects_fixed_head(self) -> torch.nn.Module | None:
+        """Get fixed head for random effects loading.
+
+        Returns
+        -------
+        nn.Module | None
+            Fixed head module.
+        """
+        return self.classifier_head
+
+    def _get_n_classes_for_random_effects(self) -> int:
+        """Get number of classes for random effects initialization.
+
+        Returns
+        -------
+        int
+            Number of classes.
+        """
+        return self.num_classes