interpreto 0.5.0.dev0__tar.gz → 0.5.0.dev1__tar.gz

{interpreto-0.5.0.dev0 → interpreto-0.5.0.dev1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: interpreto
-Version: 0.5.0.dev0
+Version: 0.5.0.dev1
 Summary: Interpretability toolbox for LLMs
 Author: FOR Team
 Author-email: fanny.jourdan@irt-saintexupery.com
@@ -57,7 +57,7 @@ License-File: LICENSE
 Requires-Dist: transformers>=4.22.0
 Requires-Dist: nltk
 Requires-Dist: torch>=2.0
-Requires-Dist: nnsight<0.6.0,>=0.5.1
+Requires-Dist: nnsight<0.8.0,>=0.7.0
 Requires-Dist: jaxtyping<=0.2.36
 Requires-Dist: beartype
 Requires-Dist: mknotebooks
@@ -120,6 +120,7 @@ Dynamic: license-file
 <p align="center">
   <a href="https://for-sight-ai.github.io/interpreto/"><strong>📚 Explore Interpreto docs &gt;&gt;</strong></a><br />
   <a href="https://for-sight-ai.github.io/interpreto-demo/"><strong>🖼️ Checkout our explanation gallery &gt;&gt;</strong></a>
+  <a href="https://arxiv.org/abs/2512.09730"><strong>📜 Read our paper &gt;&gt;</strong></a>
 </p>
 
 ## 🚀 Quick Start
@@ -163,41 +164,51 @@ They all work seamlessly for both classification (`...ForSequenceClassification`
 
 Concept-based explanations aim to provide high-level interpretations of latent model representations.
 
-Interpreto generalizes these methods through four core steps:
+We propose both supervised (probes and CAVs) and unsupervised (dictionary learning) approaches.
+
+Interpreto generalizes these methods through four core steps, the two first are common between both approaches:
 
 1. Split a model in two and obtain a dataset of activations
-2. Concept Discovery (e.g., from latent embeddings)
-3. Concept Interpretation (mapping discovered concepts to human-understandable elements)
-4. Concept-to-Output Attribution (assessing concept relevance to model outputs)
+2. Learn concepts (e.g., from latent embeddings)
+3. Interpret concepts (mapping discovered concepts to human-understandable elements)
+4. Estimate concepts importance (assessing concept relevance to model outputs)
 
 **1. Split a model in two and obtain a dataset of activations:** (mainly via [`nnsight`](https://github.com/ndif-team/nnsight)):
 
 Choose any layer in any HuggingFace language model with our `ModelWithSplitPoints` based on `nnsight`. Then pass a dataset through it to obtain a dataset of activations.
 
-**2. Dictionary Learning for Concept Discovery** (mainly via [`overcomplete`](https://github.com/KempnerInstitute/overcomplete)):
+**2. (supervised) Train probe** with the [`ProbeExplainer`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/)
+
+We differentiate two families of probes:
+
+- Linear probes: [`LinearRegressionProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.LinearRegressionProbe), [`LogisticRegressionProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.LogisticRegressionProbe), [`LinearSVMProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.LinearSVMProbe), [`MeansDiffProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.MeansDiffProbe)
+- Centroid-based probes: [`CosineCentroidProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.CosineCentroidProbe), [`DotProductCentroidProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.DotProductCentroidProbe), [`SqL2CentroidProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.SqL2CentroidProbe), [`SVDDCentroidProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.SVDDCentroidProbe), [`DiagonalMahalanobisCentroidProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.DiagonalMahalanobisCentroidProbe)
+
+Both can be tuned with `bias_calibrator` and `normalization` parameters.
+
+**2. (unsupervised) Dictionary Learning for Concept Discovery** (mainly via [`overcomplete`](https://github.com/KempnerInstitute/overcomplete)):
 
 - Interpret neurons directly via [`NeuronsAsConcepts`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/neurons_as_concepts/)
 - [`NMF`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/optim/#interpreto.concepts.NMFConcepts), [`Semi-NMF`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/optim/#interpreto.concepts.SemiNMFConcepts), [`ConvexNMF`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/optim/#interpreto.concepts.ConvexNMFConcepts)
 - [`ICA`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/optim/#interpreto.concepts.ICAConcepts), [`SVD`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/optim/#interpreto.concepts.SVDConcepts), [`PCA`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/optim/#interpreto.concepts.PCAConcepts), [`KMeans`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/optim/#interpreto.concepts.KMeansConcepts)
 - SAE variants: [`Vanilla SAE`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/sae/#interpreto.concepts.VanillaSAEConcepts), [`TopK SAE`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/sae/#interpreto.concepts.TopKSAEConcepts), [`JumpReLU SAE`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/sae/#interpreto.concepts.JumpReLUSAEConcepts), [`BatchTopK SAE`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/sae/#interpreto.concepts.BatchTopKSAEConcepts)
 
-**3. Available Concept Interpretation Techniques:**
+**3. (unsupervised) Available Concept Interpretation Techniques:**
 
 - Top-k tokens from tokenizer vocabulary via [`TopKInputs`](https://for-sight-ai.github.io/interpreto/api/concepts/concepts_interpretations/#interpreto.concepts.interpretations.TopKInputs) and `use_vocab=True`
 - Top-k tokens/words/sentences/samples from specific datasets via [`TopKInputs`](https://for-sight-ai.github.io/interpreto/api/concepts/concepts_interpretations/#interpreto.concepts.interpretations.TopKInputs)
 - Label concepts via LLMs with [`LLMLabels`](https://for-sight-ai.github.io/interpreto/api/concepts/concepts_interpretations/#interpreto.concepts.interpretations.LLMLabels) ([Bills et al. 2023](https://openai.com/index/language-models-can-explain-neurons-in-language-models/))
+- Input-to-concept attribution from dataset examples ([Concept Attributions](https://for-sight-ai.github.io/interpreto/api/concepts/interpretations/concept_attributions/)) ([Jourdan et al. 2023](https://aclanthology.org/2023.findings-acl.317/))
 
 <details><summary>Concept Interpretation Techniques Added in the future:</summary>
 
-- Input-to-concept attribution from dataset examples ([Jourdan et al. 2023](https://aclanthology.org/2023.findings-acl.317/))
-- Theme prediction via LLMs from top-k tokens/sentences
 - Aligning concepts with human labels ([Sajjad et al. 2022](https://aclanthology.org/2022.naacl-main.225/))
 - Word cloud visualizations of concepts ([Dalvi et al. 2022](https://arxiv.org/abs/2205.07237))
 - VocabProj & TokenChange ([Gur-Arieh et al. 2025](https://arxiv.org/abs/2501.08319))
 
 </details>
 
-**4. Concept-to-Output Attribution:**
+**4. (unsupervised) Concept-to-Output Attribution:**
 
 Estimate the contribution of each concept to the model output.
 
@@ -207,7 +218,6 @@ Can be obtained with any concept-based explainer via [`MethodConcepts.concept_ou
 
 Thanks to this generalization encompassing all concept-based methods and our highly flexible architecture, we can easily obtain a large number of concept-based methods:
 
-- CAV and TCAV: [Kim et al. 2018, Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)](http://proceedings.mlr.press/v80/kim18d.html)
 - ConceptSHAP: [Yeh et al. 2020, On Completeness-aware Concept-Based Explanations in Deep Neural Networks](https://proceedings.neurips.cc/paper/2020/hash/ecb287ff763c169694f682af52c1f309-Abstract.html)
 - COCKATIEL: [Jourdan et al. 2023, COCKATIEL: COntinuous Concept ranKed ATtribution with Interpretable ELements for explaining neural net classifiers on NLP](https://aclanthology.org/2023.findings-acl.317/)
 - Yun et al. 2021, [Transformer visualization via dictionary learning: contextualized embedding as a linear superposition of transformer factors](https://arxiv.org/abs/2103.15949)
@@ -264,7 +274,7 @@ Interpreto 🪄 is a project of the [FOR](https://www.irt-saintexupery.com/fr/fo
 
 ## 🗞️ Citation
 
-If you use Interpreto 🪄 as part of your workflow in a scientific publication, please consider citing 🗞️ our paper:
+If you use Interpreto 🪄 as part of your workflow in a scientific publication, please consider citing 🗞️ [our paper](https://arxiv.org/abs/2512.09730):
 
 ```bibtex
 @article{poche2025interpreto,

{interpreto-0.5.0.dev0 → interpreto-0.5.0.dev1}/README.md

@@ -15,6 +15,7 @@
 <p align="center">
   <a href="https://for-sight-ai.github.io/interpreto/"><strong>📚 Explore Interpreto docs &gt;&gt;</strong></a><br />
   <a href="https://for-sight-ai.github.io/interpreto-demo/"><strong>🖼️ Checkout our explanation gallery &gt;&gt;</strong></a>
+  <a href="https://arxiv.org/abs/2512.09730"><strong>📜 Read our paper &gt;&gt;</strong></a>
 </p>
 
 ## 🚀 Quick Start
@@ -58,41 +59,51 @@ They all work seamlessly for both classification (`...ForSequenceClassification`
 
 Concept-based explanations aim to provide high-level interpretations of latent model representations.
 
-Interpreto generalizes these methods through four core steps:
+We propose both supervised (probes and CAVs) and unsupervised (dictionary learning) approaches.
+
+Interpreto generalizes these methods through four core steps, the two first are common between both approaches:
 
 1. Split a model in two and obtain a dataset of activations
-2. Concept Discovery (e.g., from latent embeddings)
-3. Concept Interpretation (mapping discovered concepts to human-understandable elements)
-4. Concept-to-Output Attribution (assessing concept relevance to model outputs)
+2. Learn concepts (e.g., from latent embeddings)
+3. Interpret concepts (mapping discovered concepts to human-understandable elements)
+4. Estimate concepts importance (assessing concept relevance to model outputs)
 
 **1. Split a model in two and obtain a dataset of activations:** (mainly via [`nnsight`](https://github.com/ndif-team/nnsight)):
 
 Choose any layer in any HuggingFace language model with our `ModelWithSplitPoints` based on `nnsight`. Then pass a dataset through it to obtain a dataset of activations.
 
-**2. Dictionary Learning for Concept Discovery** (mainly via [`overcomplete`](https://github.com/KempnerInstitute/overcomplete)):
+**2. (supervised) Train probe** with the [`ProbeExplainer`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/)
+
+We differentiate two families of probes:
+
+- Linear probes: [`LinearRegressionProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.LinearRegressionProbe), [`LogisticRegressionProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.LogisticRegressionProbe), [`LinearSVMProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.LinearSVMProbe), [`MeansDiffProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.MeansDiffProbe)
+- Centroid-based probes: [`CosineCentroidProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.CosineCentroidProbe), [`DotProductCentroidProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.DotProductCentroidProbe), [`SqL2CentroidProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.SqL2CentroidProbe), [`SVDDCentroidProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.SVDDCentroidProbe), [`DiagonalMahalanobisCentroidProbe`](https://for-sight-ai.github.io/interpreto/api/concepts/probes/#interpreto.concepts.probes.DiagonalMahalanobisCentroidProbe)
+
+Both can be tuned with `bias_calibrator` and `normalization` parameters.
+
+**2. (unsupervised) Dictionary Learning for Concept Discovery** (mainly via [`overcomplete`](https://github.com/KempnerInstitute/overcomplete)):
 
 - Interpret neurons directly via [`NeuronsAsConcepts`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/neurons_as_concepts/)
 - [`NMF`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/optim/#interpreto.concepts.NMFConcepts), [`Semi-NMF`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/optim/#interpreto.concepts.SemiNMFConcepts), [`ConvexNMF`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/optim/#interpreto.concepts.ConvexNMFConcepts)
 - [`ICA`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/optim/#interpreto.concepts.ICAConcepts), [`SVD`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/optim/#interpreto.concepts.SVDConcepts), [`PCA`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/optim/#interpreto.concepts.PCAConcepts), [`KMeans`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/optim/#interpreto.concepts.KMeansConcepts)
 - SAE variants: [`Vanilla SAE`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/sae/#interpreto.concepts.VanillaSAEConcepts), [`TopK SAE`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/sae/#interpreto.concepts.TopKSAEConcepts), [`JumpReLU SAE`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/sae/#interpreto.concepts.JumpReLUSAEConcepts), [`BatchTopK SAE`](https://for-sight-ai.github.io/interpreto/api/concepts/methods/sae/#interpreto.concepts.BatchTopKSAEConcepts)
 
-**3. Available Concept Interpretation Techniques:**
+**3. (unsupervised) Available Concept Interpretation Techniques:**
 
 - Top-k tokens from tokenizer vocabulary via [`TopKInputs`](https://for-sight-ai.github.io/interpreto/api/concepts/concepts_interpretations/#interpreto.concepts.interpretations.TopKInputs) and `use_vocab=True`
 - Top-k tokens/words/sentences/samples from specific datasets via [`TopKInputs`](https://for-sight-ai.github.io/interpreto/api/concepts/concepts_interpretations/#interpreto.concepts.interpretations.TopKInputs)
 - Label concepts via LLMs with [`LLMLabels`](https://for-sight-ai.github.io/interpreto/api/concepts/concepts_interpretations/#interpreto.concepts.interpretations.LLMLabels) ([Bills et al. 2023](https://openai.com/index/language-models-can-explain-neurons-in-language-models/))
+- Input-to-concept attribution from dataset examples ([Concept Attributions](https://for-sight-ai.github.io/interpreto/api/concepts/interpretations/concept_attributions/)) ([Jourdan et al. 2023](https://aclanthology.org/2023.findings-acl.317/))
 
 <details><summary>Concept Interpretation Techniques Added in the future:</summary>
 
-- Input-to-concept attribution from dataset examples ([Jourdan et al. 2023](https://aclanthology.org/2023.findings-acl.317/))
-- Theme prediction via LLMs from top-k tokens/sentences
 - Aligning concepts with human labels ([Sajjad et al. 2022](https://aclanthology.org/2022.naacl-main.225/))
 - Word cloud visualizations of concepts ([Dalvi et al. 2022](https://arxiv.org/abs/2205.07237))
 - VocabProj & TokenChange ([Gur-Arieh et al. 2025](https://arxiv.org/abs/2501.08319))
 
 </details>
 
-**4. Concept-to-Output Attribution:**
+**4. (unsupervised) Concept-to-Output Attribution:**
 
 Estimate the contribution of each concept to the model output.
 
@@ -102,7 +113,6 @@ Can be obtained with any concept-based explainer via [`MethodConcepts.concept_ou
 
 Thanks to this generalization encompassing all concept-based methods and our highly flexible architecture, we can easily obtain a large number of concept-based methods:
 
-- CAV and TCAV: [Kim et al. 2018, Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)](http://proceedings.mlr.press/v80/kim18d.html)
 - ConceptSHAP: [Yeh et al. 2020, On Completeness-aware Concept-Based Explanations in Deep Neural Networks](https://proceedings.neurips.cc/paper/2020/hash/ecb287ff763c169694f682af52c1f309-Abstract.html)
 - COCKATIEL: [Jourdan et al. 2023, COCKATIEL: COntinuous Concept ranKed ATtribution with Interpretable ELements for explaining neural net classifiers on NLP](https://aclanthology.org/2023.findings-acl.317/)
 - Yun et al. 2021, [Transformer visualization via dictionary learning: contextualized embedding as a linear superposition of transformer factors](https://arxiv.org/abs/2103.15949)
@@ -159,7 +169,7 @@ Interpreto 🪄 is a project of the [FOR](https://www.irt-saintexupery.com/fr/fo
 
 ## 🗞️ Citation
 
-If you use Interpreto 🪄 as part of your workflow in a scientific publication, please consider citing 🗞️ our paper:
+If you use Interpreto 🪄 as part of your workflow in a scientific publication, please consider citing 🗞️ [our paper](https://arxiv.org/abs/2512.09730):
 
 ```bibtex
 @article{poche2025interpreto,

{interpreto-0.5.0.dev0 → interpreto-0.5.0.dev1}/interpreto/__init__.py

@@ -42,7 +42,7 @@ from .attributions import (
 from .commons import (
     Granularity,
 )
-from .model_wrapping import ModelWithSplitPoints
+from .model_wrapping import ModelWithSplitPoints, SplitSequenceClassification
 from .visualizations import (
     AttributionVisualization,
     plot_attributions,
@@ -71,6 +71,7 @@ __all__ = [
     "SquareGrad",
     "Saliency",
     "SmoothGrad",
+    "SplitSequenceClassification",
     "Sobol",
     "VarGrad",
     "get_version",

interpreto-0.5.0.dev1/interpreto/_vendor/overcomplete/base.pyi

@@ -0,0 +1,16 @@
+from abc import ABC
+
+import torch
+from torch import nn
+
+class BaseDictionaryLearning(ABC, nn.Module):
+    nb_concepts: int
+    device: str | torch.device
+    fitted: bool
+
+    def __init__(self, nb_concepts: int, device: str | torch.device = "cpu") -> None: ...
+    def encode(self, x: torch.Tensor) -> torch.Tensor: ...
+    def decode(self, z: torch.Tensor) -> torch.Tensor: ...
+    def fit(self, x: torch.Tensor) -> None: ...
+    def get_dictionary(self) -> torch.Tensor: ...
+    def to(self, device: str | torch.device) -> "BaseDictionaryLearning": ...  # type: ignore[override]

{interpreto-0.5.0.dev0 → interpreto-0.5.0.dev1}/interpreto/attributions/base.py

@@ -45,13 +45,17 @@ from interpreto.attributions.perturbations.base import Perturbator
 from interpreto.commons import Granularity
 from interpreto.commons.generator_tools import split_iterator
 from interpreto.commons.granularity import GranularityAggregationStrategy
+from interpreto.concepts.base import ModelForInputsToConcepts
 from interpreto.model_wrapping.classification_inference_wrapper import ClassificationInferenceWrapper
 from interpreto.model_wrapping.generation_inference_wrapper import GenerationInferenceWrapper
 from interpreto.model_wrapping.inference_wrapper import InferenceModes, InferenceWrapper
+from interpreto.model_wrapping.inputs_to_concepts_inference_wrapper import InputsToConceptsInferenceWrapper
 from interpreto.typing import ClassificationTarget, GeneratedTarget, ModelInputs, SingleAttribution, TensorMapping
 
 
-def setup_token_ids(model: PreTrainedModel, tokenizer: PreTrainedTokenizer, require_mask_token: bool = True) -> int:
+def setup_token_ids(
+    model: PreTrainedModel | ModelForInputsToConcepts, tokenizer: PreTrainedTokenizer, require_mask_token: bool = True
+) -> int:
     """
     Setup the tokenizer and the model with the appropriate token IDs, for padding and masking.
 
@@ -139,6 +143,7 @@ class ModelTask(Enum):
 
     CLASSIFICATION = "classification"
     GENERATION = "generation"
+    CONCEPTS = "concepts"
 
 
 def clone_tensor_mapping(tm: TensorMapping, detach: bool = False) -> TensorMapping:
@@ -232,7 +237,7 @@ class AttributionExplainer:
 
     def __init__(
         self,
-        model: PreTrainedModel,
+        model: PreTrainedModel | ModelForInputsToConcepts,
         tokenizer: PreTrainedTokenizer,
         batch_size: int = 4,
         perturbator: Perturbator | None = None,
@@ -248,7 +253,7 @@ class AttributionExplainer:
         Initializes the AttributionExplainer.
 
         Args:
-            model (PreTrainedModel): The model to be explained.
+            model (PreTrainedModel | ModelForInputsToConcepts): The model to be explained.
             tokenizer (PreTrainedTokenizer): The tokenizer associated with the model.
             batch_size (int): The batch size used for model inference.
             perturbator (Perturbator, optional): Instance used to generate input perturbations.
@@ -276,7 +281,7 @@ class AttributionExplainer:
         self.tokenizer = tokenizer
 
         self.inference_wrapper = self._associated_inference_wrapper(
-            model,
+            model,  # type: ignore
             gradients=use_gradient,
             input_x_gradient=input_x_gradient,
             batch_size=batch_size,
@@ -846,13 +851,99 @@ class GenerationAttributionExplainer(AttributionExplainer):
         return ModelTask.GENERATION, contribution
 
 
+class InputsToConceptsAttributionsExplainer(AttributionExplainer):
+    """Attribution explainer for input-to-concept models.
+
+    This explainer computes how much each input token contributes to each concept
+    activation. It bridges the attribution framework with the concept framework:
+    once a concept explainer is fitted, its ``inputs_to_concepts`` property returns
+    a model that can be passed to any perturbation-based attribution method.
+
+    The result is a per-token attribution for each concept, revealing which parts
+    of the input are most responsible for activating a given concept.
+
+    Note:
+        Only perturbation-based methods (Lime, KernelShap, Occlusion, Sobol) are
+        supported. Gradient-based methods are incompatible because the
+        ``ModelForInputsToConcepts`` is based on `nnsight` and which make differentiation complex.
+
+    Example:
+        ```python
+        from interpreto import Occlusion, SplitSequenceClassification
+        from interpreto.concepts import SemiNMFConcepts
+
+        split_model = SplitSequenceClassification("model_id", device_map="cuda")
+        concept_explainer = SemiNMFConcepts(split_model, nb_concepts=20)
+        concept_explainer.fit(activations)
+
+        explainer = Occlusion(concept_explainer.inputs_to_concepts, split_model.tokenizer)
+        results = explainer.explain("Some input text.", targets=torch.arange(5))
+        ```
+    """
+
+    _associated_inference_wrapper = InputsToConceptsInferenceWrapper
+    inference_wrapper: InputsToConceptsInferenceWrapper
+
+    def process_inputs_to_explain_and_targets(  # type: ignore
+        self,
+        model_inputs: ModelInputs,
+        targets: Iterable[int] | None = None,
+    ) -> tuple[list[TensorMapping], list[Int[torch.Tensor, "t"]]]:
+        """
+        Processes the inputs and targets for explanation.
+
+        This method must be implemented by subclasses.
+
+        Args:
+            model_inputs (ModelInputs):
+                The inputs to the model.
+            targets (Optional[Iterable[int]]):
+                The targets to be explained.
+                If None, all concepts are explained.
+
+        Returns:
+            processed_inputs (list[TensorMapping]):
+                The processed inputs.
+            processed_targets (list[Int[torch.Tensor, "t"]]):
+                The processed targets.
+        """
+        sanitized_targets: list[Int[torch.Tensor, "t"]]
+        if targets is None:
+            # explain all concepts
+            input_wise_targets = torch.arange(self.inference_wrapper.model.nb_concepts)  # type: ignore
+            sanitized_targets = [input_wise_targets] * len(model_inputs)  # type: ignore
+        else:
+            # targets are concept indices, shared across all inputs
+            if isinstance(targets, torch.Tensor):
+                input_wise_targets = targets.long()
+            else:
+                input_wise_targets = torch.tensor(list(targets), dtype=torch.long)
+            sanitized_targets = [input_wise_targets] * len(model_inputs)  # type: ignore
+        return model_inputs, sanitized_targets  # type: ignore
+
+    def post_processing(self, contribution: Float[torch.Tensor, "t l"]):
+        """
+        Concepts specific post-processing of the attribution scores.
+
+        No post-processing is required for concept attributions.
+
+        Args:
+            contribution (Float[torch.Tensor, "t l"]): The contribution values.
+
+        Returns:
+            model_task (ModelTask): The model task.
+            contribution (Float[torch.Tensor, "t l"]): The post-processed contribution values.
+        """
+        return ModelTask.CONCEPTS, contribution
+
+
 class FactoryGeneratedMeta(type):
     """
     Metaclass to distinguish classes generated by the MultitaskExplainerMixin.
     """
 
 
-class MultitaskExplainerMixin(AttributionExplainer):
+class MultitaskExplainerMixin:
     """
     Mixin class to generate the appropriate Explainer based on the model type.
     """
@@ -866,6 +957,11 @@ class MultitaskExplainerMixin(AttributionExplainer):
         if model.__class__.__name__.endswith("ForCausalLM") or model.__class__.__name__.endswith("LMHeadModel"):
             t = FactoryGeneratedMeta("Generation" + cls.__name__, (cls, GenerationAttributionExplainer), {})
             return t.__new__(t, model, *args, **kwargs)  # type: ignore
+        if model.__class__.__name__.endswith("ForInputsToConcepts"):
+            t = FactoryGeneratedMeta(
+                "InputsToConcepts" + cls.__name__, (cls, InputsToConceptsAttributionsExplainer), {}
+            )
+            return t.__new__(t, model, *args, **kwargs)  # type: ignore
         raise NotImplementedError(
             "Model type not supported for Explainer. Use a ModelForSequenceClassification, a ModelForCausalLM model or a LMHeadModel model."
         )

{interpreto-0.5.0.dev0 → interpreto-0.5.0.dev1}/interpreto/attributions/methods/kernel_shap.py

@@ -40,6 +40,7 @@ from interpreto.attributions.aggregations.linear_regression_aggregation import (
 from interpreto.attributions.base import AttributionExplainer, MultitaskExplainerMixin, setup_token_ids
 from interpreto.attributions.perturbations.shap_perturbation import ShapTokenPerturbator
 from interpreto.commons.granularity import Granularity, GranularityAggregationStrategy
+from interpreto.concepts.base import ModelForInputsToConcepts
 from interpreto.model_wrapping.inference_wrapper import InferenceModes
 
 
@@ -68,7 +69,7 @@ class KernelShap(MultitaskExplainerMixin, AttributionExplainer):
 
     def __init__(
         self,
-        model: PreTrainedModel,
+        model: PreTrainedModel | ModelForInputsToConcepts,
         tokenizer: PreTrainedTokenizer,
         batch_size: int = 4,
         granularity: Granularity = Granularity.WORD,
@@ -81,7 +82,7 @@ class KernelShap(MultitaskExplainerMixin, AttributionExplainer):
         Initialize the attribution method.
 
         Args:
-            model (PreTrainedModel): model to explain
+            model (PreTrainedModel | ModelForInputsToConcepts): model to explain
             tokenizer (PreTrainedTokenizer): Hugging Face tokenizer associated with the model
             batch_size (int): batch size for the attribution method
             granularity (Granularity, optional): The level of granularity for the explanation.

{interpreto-0.5.0.dev0 → interpreto-0.5.0.dev1}/interpreto/attributions/methods/lime.py

@@ -43,6 +43,7 @@ from interpreto.attributions.aggregations.linear_regression_aggregation import (
 from interpreto.attributions.base import AttributionExplainer, InferenceModes, MultitaskExplainerMixin, setup_token_ids
 from interpreto.attributions.perturbations.random_perturbation import RandomMaskedTokenPerturbator
 from interpreto.commons import Granularity, GranularityAggregationStrategy
+from interpreto.concepts.base import ModelForInputsToConcepts
 
 
 class Lime(MultitaskExplainerMixin, AttributionExplainer):
@@ -72,7 +73,7 @@ class Lime(MultitaskExplainerMixin, AttributionExplainer):
 
     def __init__(
         self,
-        model: PreTrainedModel,
+        model: PreTrainedModel | ModelForInputsToConcepts,
         tokenizer: PreTrainedTokenizer,
         batch_size: int = 4,
         granularity: Granularity = Granularity.WORD,
@@ -88,7 +89,7 @@ class Lime(MultitaskExplainerMixin, AttributionExplainer):
         Initialize the attribution method.
 
         Args:
-            model (PreTrainedModel): model to explain
+            model (PreTrainedModel | ModelForInputsToConcepts): model to explain
             tokenizer (PreTrainedTokenizer): Hugging Face tokenizer associated with the model
             batch_size (int): batch size for the attribution method
             granularity (Granularity, optional): The level of granularity for the explanation.

{interpreto-0.5.0.dev0 → interpreto-0.5.0.dev1}/interpreto/attributions/methods/occlusion.py

@@ -29,10 +29,9 @@ Occlusion attribution method
 from __future__ import annotations
 
 from collections.abc import Callable
-from typing import Any
 
 import torch
-from transformers import PreTrainedTokenizer
+from transformers import PreTrainedModel, PreTrainedTokenizer
 
 from interpreto.attributions.aggregations.base import OcclusionAggregator
 from interpreto.attributions.base import (
@@ -42,6 +41,7 @@ from interpreto.attributions.base import (
 )
 from interpreto.attributions.perturbations import OcclusionPerturbator
 from interpreto.commons.granularity import Granularity, GranularityAggregationStrategy
+from interpreto.concepts.base import ModelForInputsToConcepts
 from interpreto.model_wrapping.inference_wrapper import InferenceModes
 
 
@@ -68,7 +68,7 @@ class Occlusion(MultitaskExplainerMixin, AttributionExplainer):
 
     def __init__(
         self,
-        model: Any,
+        model: PreTrainedModel | ModelForInputsToConcepts,
         tokenizer: PreTrainedTokenizer,
         batch_size: int = 4,
         granularity: Granularity = Granularity.WORD,
@@ -80,7 +80,7 @@ class Occlusion(MultitaskExplainerMixin, AttributionExplainer):
         Initialize the attribution method.
 
         Args:
-            model (PreTrainedModel): model to explain
+            model (PreTrainedModel | ModelForInputsToConcepts): model to explain
             tokenizer (PreTrainedTokenizer): Hugging Face tokenizer associated with the model
             batch_size (int): batch size for the attribution method
             granularity (Granularity, optional): The level of granularity for the explanation.

{interpreto-0.5.0.dev0 → interpreto-0.5.0.dev1}/interpreto/attributions/methods/sobol_attribution.py

@@ -41,6 +41,7 @@ from interpreto.attributions.perturbations.sobol_perturbation import (
     SobolTokenPerturbator,
 )
 from interpreto.commons.granularity import Granularity, GranularityAggregationStrategy
+from interpreto.concepts.base import ModelForInputsToConcepts
 
 
 class Sobol(MultitaskExplainerMixin, AttributionExplainer):
@@ -73,7 +74,7 @@ class Sobol(MultitaskExplainerMixin, AttributionExplainer):
 
     def __init__(
         self,
-        model: PreTrainedModel,
+        model: PreTrainedModel | ModelForInputsToConcepts,
         tokenizer: PreTrainedTokenizer,
         batch_size: int = 4,
         granularity: Granularity = Granularity.WORD,
@@ -88,7 +89,7 @@ class Sobol(MultitaskExplainerMixin, AttributionExplainer):
         Initialize the attribution method.
 
         Args:
-            model (PreTrainedModel): model to explain
+            model (PreTrainedModel | ModelForInputsToConcepts): model to explain
             tokenizer (PreTrainedTokenizer): Hugging Face tokenizer associated with the model
             batch_size (int): batch size for the attribution method
             granularity (Granularity, optional): The level of granularity for the explanation.

{interpreto-0.5.0.dev0 → interpreto-0.5.0.dev1}/interpreto/concepts/__init__.py

@@ -43,6 +43,18 @@ from .methods import (
     TopKSAEConcepts,
     VanillaSAEConcepts,
 )
+from .probes import (
+    CosineCentroidProbe,
+    DiagonalMahalanobisCentroidProbe,
+    DotProductCentroidProbe,
+    LinearRegressionProbe,
+    LinearSVMProbe,
+    LogisticRegressionProbe,
+    MeansDiffProbe,
+    ProbeExplainer,
+    SqL2CentroidProbe,
+    SVDDCentroidProbe,
+)
 
 __all__ = [
     "BatchTopKSAEConcepts",
@@ -50,19 +62,29 @@ __all__ = [
     "ConceptAutoEncoderExplainer",
     "ConceptEncoderExplainer",
     "ConvexNMFConcepts",
+    "CosineCentroidProbe",
+    "DiagonalMahalanobisCentroidProbe",
     "DictionaryLearningConcepts",
+    "DotProductCentroidProbe",
     "ICAConcepts",
     "JumpReLUSAEConcepts",
     "KMeansConcepts",
     "LLMLabels",
+    "LinearRegressionProbe",
+    "LinearSVMProbe",
+    "LogisticRegressionProbe",
+    "MeansDiffProbe",
     "MpSAEConcepts",
     "NeuronsAsConcepts",
     "NMFConcepts",
     "PCAConcepts",
+    "ProbeExplainer",
     "SAELossClasses",
     "SemiNMFConcepts",
     "SparsePCAConcepts",
+    "SqL2CentroidProbe",
     "SVDConcepts",
+    "SVDDCentroidProbe",
     "TopKInputs",
     "TopKSAEConcepts",
     "VanillaSAEConcepts",