add OASM model from Hadidi et al. 2025

brainscore_language/models/oasm/__init__.py

@@ -0,0 +1,53 @@
+from brainscore_language import model_registry
+from .model import OASMSubject
+
+# 49 OASM variants: sigma=0 (pure identity) + sigma=0.1..4.8 (paper's search grid).
+model_registry['oasm-sigma0'] = lambda: OASMSubject(identifier='oasm-sigma0', sigma=0.0)
+model_registry['oasm-sigma0.1'] = lambda: OASMSubject(identifier='oasm-sigma0.1', sigma=0.1)
+model_registry['oasm-sigma0.2'] = lambda: OASMSubject(identifier='oasm-sigma0.2', sigma=0.2)
+model_registry['oasm-sigma0.3'] = lambda: OASMSubject(identifier='oasm-sigma0.3', sigma=0.3)
+model_registry['oasm-sigma0.4'] = lambda: OASMSubject(identifier='oasm-sigma0.4', sigma=0.4)
+model_registry['oasm-sigma0.5'] = lambda: OASMSubject(identifier='oasm-sigma0.5', sigma=0.5)
+model_registry['oasm-sigma0.6'] = lambda: OASMSubject(identifier='oasm-sigma0.6', sigma=0.6)
+model_registry['oasm-sigma0.7'] = lambda: OASMSubject(identifier='oasm-sigma0.7', sigma=0.7)
+model_registry['oasm-sigma0.8'] = lambda: OASMSubject(identifier='oasm-sigma0.8', sigma=0.8)
+model_registry['oasm-sigma0.9'] = lambda: OASMSubject(identifier='oasm-sigma0.9', sigma=0.9)
+model_registry['oasm-sigma1.0'] = lambda: OASMSubject(identifier='oasm-sigma1.0', sigma=1.0)
+model_registry['oasm-sigma1.1'] = lambda: OASMSubject(identifier='oasm-sigma1.1', sigma=1.1)
+model_registry['oasm-sigma1.2'] = lambda: OASMSubject(identifier='oasm-sigma1.2', sigma=1.2)
+model_registry['oasm-sigma1.3'] = lambda: OASMSubject(identifier='oasm-sigma1.3', sigma=1.3)
+model_registry['oasm-sigma1.4'] = lambda: OASMSubject(identifier='oasm-sigma1.4', sigma=1.4)
+model_registry['oasm-sigma1.5'] = lambda: OASMSubject(identifier='oasm-sigma1.5', sigma=1.5)
+model_registry['oasm-sigma1.6'] = lambda: OASMSubject(identifier='oasm-sigma1.6', sigma=1.6)
+model_registry['oasm-sigma1.7'] = lambda: OASMSubject(identifier='oasm-sigma1.7', sigma=1.7)
+model_registry['oasm-sigma1.8'] = lambda: OASMSubject(identifier='oasm-sigma1.8', sigma=1.8)
+model_registry['oasm-sigma1.9'] = lambda: OASMSubject(identifier='oasm-sigma1.9', sigma=1.9)
+model_registry['oasm-sigma2.0'] = lambda: OASMSubject(identifier='oasm-sigma2.0', sigma=2.0)
+model_registry['oasm-sigma2.1'] = lambda: OASMSubject(identifier='oasm-sigma2.1', sigma=2.1)
+model_registry['oasm-sigma2.2'] = lambda: OASMSubject(identifier='oasm-sigma2.2', sigma=2.2)
+model_registry['oasm-sigma2.3'] = lambda: OASMSubject(identifier='oasm-sigma2.3', sigma=2.3)
+model_registry['oasm-sigma2.4'] = lambda: OASMSubject(identifier='oasm-sigma2.4', sigma=2.4)
+model_registry['oasm-sigma2.5'] = lambda: OASMSubject(identifier='oasm-sigma2.5', sigma=2.5)
+model_registry['oasm-sigma2.6'] = lambda: OASMSubject(identifier='oasm-sigma2.6', sigma=2.6)
+model_registry['oasm-sigma2.7'] = lambda: OASMSubject(identifier='oasm-sigma2.7', sigma=2.7)
+model_registry['oasm-sigma2.8'] = lambda: OASMSubject(identifier='oasm-sigma2.8', sigma=2.8)
+model_registry['oasm-sigma2.9'] = lambda: OASMSubject(identifier='oasm-sigma2.9', sigma=2.9)
+model_registry['oasm-sigma3.0'] = lambda: OASMSubject(identifier='oasm-sigma3.0', sigma=3.0)
+model_registry['oasm-sigma3.1'] = lambda: OASMSubject(identifier='oasm-sigma3.1', sigma=3.1)
+model_registry['oasm-sigma3.2'] = lambda: OASMSubject(identifier='oasm-sigma3.2', sigma=3.2)
+model_registry['oasm-sigma3.3'] = lambda: OASMSubject(identifier='oasm-sigma3.3', sigma=3.3)
+model_registry['oasm-sigma3.4'] = lambda: OASMSubject(identifier='oasm-sigma3.4', sigma=3.4)
+model_registry['oasm-sigma3.5'] = lambda: OASMSubject(identifier='oasm-sigma3.5', sigma=3.5)
+model_registry['oasm-sigma3.6'] = lambda: OASMSubject(identifier='oasm-sigma3.6', sigma=3.6)
+model_registry['oasm-sigma3.7'] = lambda: OASMSubject(identifier='oasm-sigma3.7', sigma=3.7)
+model_registry['oasm-sigma3.8'] = lambda: OASMSubject(identifier='oasm-sigma3.8', sigma=3.8)
+model_registry['oasm-sigma3.9'] = lambda: OASMSubject(identifier='oasm-sigma3.9', sigma=3.9)
+model_registry['oasm-sigma4.0'] = lambda: OASMSubject(identifier='oasm-sigma4.0', sigma=4.0)
+model_registry['oasm-sigma4.1'] = lambda: OASMSubject(identifier='oasm-sigma4.1', sigma=4.1)
+model_registry['oasm-sigma4.2'] = lambda: OASMSubject(identifier='oasm-sigma4.2', sigma=4.2)
+model_registry['oasm-sigma4.3'] = lambda: OASMSubject(identifier='oasm-sigma4.3', sigma=4.3)
+model_registry['oasm-sigma4.4'] = lambda: OASMSubject(identifier='oasm-sigma4.4', sigma=4.4)
+model_registry['oasm-sigma4.5'] = lambda: OASMSubject(identifier='oasm-sigma4.5', sigma=4.5)
+model_registry['oasm-sigma4.6'] = lambda: OASMSubject(identifier='oasm-sigma4.6', sigma=4.6)
+model_registry['oasm-sigma4.7'] = lambda: OASMSubject(identifier='oasm-sigma4.7', sigma=4.7)
+model_registry['oasm-sigma4.8'] = lambda: OASMSubject(identifier='oasm-sigma4.8', sigma=4.8)

brainscore_language/models/oasm/metadata.yml

@@ -0,0 +1,15 @@
+models:
+  oasm:
+    architecture: DCNN
+    model_family: oasm
+    total_parameter_count: 1234567
+    trainable_parameter_count: 1234567
+    total_layers: 55
+    trainable_layers: 40
+    model_size_mb: 1202
+    training_dataset: null
+    task_specialization: null
+    brainscore_link: https://github.com/brain-score/language/tree/master/brainscore_language/models/oasm
+    huggingface_link: null
+    extra_notes: Temporary hardcoded metadata - will be replaced with actual generation
+    runnable: true

brainscore_language/models/oasm/model.py

@@ -0,0 +1,126 @@
+"""
+OASM: Orthogonal Autocorrelated Sequences Model.
+
+A confound baseline model following Hadidi et al. (2025) "Illusions of Alignment Between
+Large Language Models and Brains Emerge From Fragile Methods and Overlooked Confounds".
+
+The model constructs an N x N identity matrix (N = total stimuli) and applies Gaussian
+smoothing within each block (passage/sentence/story) along axis=1. This captures temporal
+autocorrelation -- the fact that brain responses to nearby stimuli are more similar --
+without encoding any linguistic content whatsoever.
+
+Reference:
+    Hadidi et al. (2025). bioRxiv. https://doi.org/10.1101/2025.03.09.642245
+    Code: https://github.com/ebrahimfeghhi/beyond-brainscore
+"""
+
+import copy
+import numpy as np
+from scipy.ndimage import gaussian_filter1d
+from typing import Union, List, Dict
+
+from brainscore_core.supported_data_standards.brainio.assemblies import NeuroidAssembly, merge_data_arrays
+from brainscore_language.artificial_subject import ArtificialSubject
+
+
+class OASMSubject(ArtificialSubject):
+    """
+    Orthogonal Autocorrelated Sequences Model (OASM) for brain-score evaluation.
+
+    For each block of K stimuli (passage, sentence, or story), constructs a K x K identity
+    matrix and applies ``scipy.ndimage.gaussian_filter1d`` along axis=1 with the given sigma.
+    Each block is placed at a unique offset in a fixed-size feature space, maintaining
+    between-block orthogonality while introducing within-block temporal autocorrelation.
+
+    This is mathematically equivalent to the paper's full N x N construction, adapted for
+    brain-score's per-block ``digest_text`` calling convention.
+
+    :param identifier: Unique model identifier (e.g., 'oasm-sigma1.0').
+    :param sigma: Gaussian smoothing width. Must be >= 0. sigma=0 gives pure identity.
+    :param max_features: Fixed dimensionality of the feature space. Must be >= total
+        number of stimuli across all blocks in a benchmark run. Default 2000.
+    """
+
+    def __init__(self, identifier: str, sigma: float, max_features: int = 2000):
+        self._identifier = identifier
+        self._sigma = sigma
+        self._max_features = max_features
+        self._neural_recordings: list = []
+        self._offset: int = 0
+
+    def identifier(self) -> str:
+        return self._identifier
+
+    def start_behavioral_task(self, task: ArtificialSubject.Task):
+        raise NotImplementedError("OASM encodes no linguistic content and cannot perform behavioral tasks")
+
+    def start_neural_recording(self, recording_target: ArtificialSubject.RecordingTarget,
+                               recording_type: ArtificialSubject.RecordingType):
+        self._neural_recordings.append((recording_target, recording_type))
+        self._offset = 0
+
+    def digest_text(self, text: Union[str, List[str]]) -> Dict[str, NeuroidAssembly]:
+        assert len(self._neural_recordings) > 0, "Must call start_neural_recording before digest_text"
+
+        if isinstance(text, str):
+            text = [text]
+        if isinstance(text, np.ndarray):
+            text = list(text)
+
+        block_size = len(text)
+
+        if self._offset + block_size > self._max_features:
+            raise ValueError(
+                f"Cumulative stimulus count ({self._offset + block_size}) exceeds "
+                f"max_features ({self._max_features}). Increase max_features."
+            )
+
+        # Build block: K x K identity, smoothed along axis=1
+        block_features = np.eye(block_size, dtype=np.float64)
+        if self._sigma > 0 and block_size > 1:
+            block_features = gaussian_filter1d(block_features, sigma=self._sigma, axis=1)
+
+        # Embed into D-dimensional space at the current offset
+        features = np.zeros((block_size, self._max_features), dtype=np.float64)
+        features[:, self._offset:self._offset + block_size] = block_features
+        self._offset += block_size
+
+        # Package each stimulus as a NeuroidAssembly
+        output = {'behavior': [], 'neural': []}
+        for part_number, stimulus in enumerate(text):
+            stimuli_coords = {
+                'stimulus': ('presentation', [stimulus]),
+                'part_number': ('presentation', [part_number]),
+            }
+            neural_assembly = self._package_representations(features[part_number], stimuli_coords=stimuli_coords)
+            output['neural'].append(neural_assembly)
+
+        output['neural'] = merge_data_arrays(output['neural']).sortby('part_number')
+        return output
+
+    def _package_representations(self, representation_values: np.ndarray,
+                                 stimuli_coords: dict) -> NeuroidAssembly:
+        """Package a feature vector as a NeuroidAssembly matching the brain-score interface."""
+        layer_name = f'oasm_sigma{self._sigma}'
+        num_units = len(representation_values)
+
+        neuroid_coords = {
+            'layer': ('neuroid', [layer_name] * num_units),
+            'neuron_number_in_layer': ('neuroid', np.arange(num_units)),
+            'neuroid_id': ('neuroid', [f'{layer_name}--{i}' for i in range(num_units)]),
+        }
+
+        layer_representations = NeuroidAssembly(
+            [representation_values],
+            coords={**stimuli_coords, **neuroid_coords},
+            dims=['presentation', 'neuroid'])
+
+        representations = []
+        for recording_target, recording_type in self._neural_recordings:
+            current_representations = copy.deepcopy(layer_representations)
+            current_representations['recording_target'] = 'neuroid', [recording_target] * num_units
+            current_representations['recording_type'] = 'neuroid', [recording_type] * num_units
+            current_representations = type(current_representations)(current_representations)
+            representations.append(current_representations)
+        representations = merge_data_arrays(representations)
+        return representations