Do interpretable heads explain the task? Code, outputs and folds

Code, outputs and cross-validation folds for the manuscript "Do interpretable heads explain the task?Reliance, necessity and redundancy in Kolmogorov–Arnold networks" (M. Ataş, submitted). The study asks whether feature-level explanations of classification heads on pretrained backbones describethe task or only the fitted model. For the features an explanation ranks highest, it measures reliance (theloss of the fitted head when these features are removed) and necessity (the loss of a head retrained withoutthem), both relative to removing as many random features. Experiments cover two public diabetic retinopathydatasets (APTOS 2019 and a class-stratified half of DDR), three ImageNet-pretrained backbones(MobileNetV3-Large, EfficientNet-B3, ResNet-50), parameter-matched Kolmogorov–Arnold network (KAN), MLP andlinear heads, two explanation methods (partial-dependence range and Integrated Gradients), and threeinterventions on the stored features (decorrelated, rotated decorrelated and embedded-subspace bases). Contents- reliance-necessity-audit_v7.ipynb — the single Kaggle notebook that produces every result. Sections 1–12: main study; 13: redundancy intervention and linear head; 14: rotation control; 15: an exploratory pilot that is not used in the paper. The notebook resumes from stored outputs, so it can be run across several sessions.- audit_outputs_core.zip — the analysis outputs: duplicate-aware folds and duplicate lists, per-fold predictions, fitted heads, every retraining result, intervention results, the tables behind every number and figure, and manifest.csv, which maps each table and figure to its source file. The stored backbone features (folders features/ and frozen/, about 2 GB) are not included; the notebook regenerates them from the public datasets (Stage A and Stage B).- audit_outputs/rn_audit.py — the reliance–necessity test as a standalone module (numpy and pandas only). Models enter through two callables, so it applies to any model, ranking and training routine.- make_figure5.py, make_supp_figures.py — scripts that draw Figure 5 and Supplementary Figures S1–S5 from the archived tables. DataThe fundus images are not redistributed. APTOS 2019 is available from Kaggle (APTOS 2019 BlindnessDetection) and DDR from its original repository; the notebook reads them from these sources. Reproducing the resultsAttach the two datasets to the notebook on Kaggle (GPU T4 x2). To repeat the whole study from scratch, setthe RUN flags in Section 1 to True and run all cells; Stage A (fine-tuning) takes about 12–15 GPU hours intotal, and later stages run on the features it stores. The tables and figures of the paper can be checkeddirectly against the files in audit_outputs_core.zip. LicenseThe code (notebook, rn_audit.py and scripts) is released under the MIT License, and the outputs and foldsunder CC BY 4.0. Please cite the accompanying article when it is published.