dilyabareeva · dilyabareeva · May 8, 2026 · May 4, 2026 · May 4, 2026 · May 4, 2026
diff --git a/.gitignore b/.gitignore
@@ -63,15 +63,15 @@ outputs/
 cache/
 /.env
 bin/*
-docs/source/**
+docs/source/docs_api/**
 eval_results/*
 eval_bench/**
 
 assets/demo/*
 bench_out/*
 fig_1_images/*
 tutorials/*
-scripts/bench_out/*
+scriptsbench_out/*
 CLAUDE.md
 .vscode/
 quanda_benchmark_tutorial_cache/*
@@ -88,4 +88,8 @@ tmp_*/*
 scripts/paper_tex/**
 scratch/*
 slurm/get_logs.sh
-scripts/delete_hf.py
+scripts/delete_hf.py
+scripts/plot_results_LOCAL.py
+scripts/**/plot_*_LOCAL.sh
+scripts/PLOT_ALL_LOCAL.sh
+slurm_LOCAL/*
diff --git a/CHANGELOG.md b/CHANGELOG.md
diff --git a/CODE_OF_CONDUCT.md b/CODE_OF_CONDUCT.md
@@ -60,7 +60,7 @@ representative at an online or offline event.
 
 Instances of abusive, harassing, or otherwise unacceptable behavior may be
 reported to the community leaders responsible for enforcement at
-dilyabareeva@gmail.com.
+[dilyabareeva@gmail.com](mailto:dilyabareeva@gmail.com).
 All complaints will be reviewed and investigated promptly and fairly.
 
 All community leaders are obligated to respect the privacy and security of the

diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
@@ -84,6 +84,12 @@ Currently, the following markers are available to filter tests:
 - benchmarks: Benchmark modules
 - global_ranking: global_ranking modules
 - self_influence: self_influence methods of explainers
+- tasks: task modules
+- integration: integration tests
+- slow: tests marked as slow (excluded by default; run with `pytest -m slow`)
+- production_bench: production benchmark sanity checks, run only when explicitly specified
+
+The authoritative list lives in `pytest.ini`.
 
 Ideally, all contributions should include tests to ensure correctness.
 
@@ -187,7 +193,7 @@ To contribute a new benchmark you generally do not need to override these four c
 1. **A subclass of** `Benchmark` under the appropriate `quanda/benchmarks/{downstream_eval,heuristics,ground_truth}/` subdirectory. Subclasses customize behavior via:
 
    - `__init__` — accept any benchmark-specific fields beyond what the base `__init__` already stores (`model`, `train_dataset`, `eval_dataset`, `checkpoints`, `checkpoints_load_func`, `device`, `val_dataset`, `use_predictions`).
-   - `_extra_kwargs_from_config(cls, config, train_dataset, eval_dataset, metadata_dir, load_meta_from_disk)` — extract any subclass-specific kwargs from the YAML and return them as a dict; they get passed into `__init__` by `from_config`.
+   - `_extra_kwargs_from_config(cls, config, train_dataset, eval_dataset, metadata_dir, load_fresh)` — extract any subclass-specific kwargs from the YAML and return them as a dict; they get passed into `__init__` by `from_config`.
    - `_compute_and_save_indices(self, config, batch_size)` — only override if your benchmark needs to cache extra metadata on the train pass (filtered eval indices, ranking caches, etc.).
    - `evaluate(self, explainer_cls, expl_kwargs, batch_size)` — runs the explainer over `eval_dataset`, feeds the attributions to the corresponding `Metric` via `update`/`compute`, and returns the result dict (must contain `"score"`).
 

diff --git a/LICENSE b/LICENSE
@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2024 Dilyara Bareeva, Galip Ümit Yolcu
+Copyright (c) 2026 Dilyara Bareeva, Galip Ümit Yolcu, Anna Hedström, Niklas Schmolenski, Thomas Wiegand, Wojciech Samek, Sebastian Lapuschkin
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

diff --git a/README.md b/README.md
diff --git a/config/eval/awa2_resnet50.yaml b/config/eval/awa2_resnet50.yaml
@@ -12,7 +12,7 @@ hydra:
     n_jobs: 1
 
 bench: awa2_class_detection
-root_dir: ${cluster_or_local:/data/cluster/users/bareeva/quanda_output_new2,/data2/bareeva/Projects/quanda/cluster_output_new2}
+root_dir: bench_out
 cache_dir: ${root_dir}/eval_bench/awa2
 results_dir: ${root_dir}/eval_results/awa2
 

diff --git a/config/eval/bert_qnli.yaml b/config/eval/bert_qnli.yaml
@@ -12,7 +12,7 @@ hydra:
     n_jobs: 1
 
 bench: qnli_class_detection
-root_dir: ${cluster_or_local:/data/cluster/users/bareeva/quanda_output_new2,/data2/bareeva/Projects/quanda/cluster_output_new2}
+root_dir: bench_out
 cache_dir: ${root_dir}/eval_bench/qnli
 results_dir: ${root_dir}/eval_results/qnli
 device: cuda:0

diff --git a/config/eval/cifar_resnet9.yaml b/config/eval/cifar_resnet9.yaml
@@ -12,7 +12,7 @@ hydra:
     n_jobs: 1
 
 bench: cifar_class_detection
-root_dir: ${cluster_or_local:/data/cluster/users/bareeva/quanda_output_new2,/data2/bareeva/Projects/quanda/cluster_output_new2}
+root_dir: bench_out
 cache_dir: ${root_dir}/eval_bench/cifar
 results_dir: ${root_dir}/eval_results/cifar
 

diff --git a/config/eval/gpt2_trex.yaml b/config/eval/gpt2_trex.yaml
@@ -12,7 +12,7 @@ hydra:
     n_jobs: 1
 
 bench: gpt2_trex_openwebtext_ft_mrr
-root_dir: ${cluster_or_local:/data/cluster/users/bareeva/quanda_output_new2,/data2/bareeva/Projects/quanda/cluster_output_new2}
+root_dir: bench_out
 cache_dir: ${root_dir}/eval_bench/gpt2_trex
 results_dir: ${root_dir}/eval_results/gpt2_trex
 device: cuda:0

diff --git a/config/eval/mnist_lenet.yaml b/config/eval/mnist_lenet.yaml
@@ -12,7 +12,7 @@ hydra:
     n_jobs: 1
 
 bench: mnist_class_detection
-root_dir: ${cluster_or_local:/data/cluster/users/bareeva/quanda_output_new2,/data2/bareeva/Projects/quanda/cluster_output_new2}
+root_dir: bench_out
 cache_dir: ${root_dir}/eval_bench/mnist
 results_dir: ${root_dir}/eval_results/mnist
 

diff --git a/docs/Makefile b/docs/Makefile
@@ -22,6 +22,7 @@ help:
 
 rst:
 	@sphinx-apidoc -o source/docs_api ../quanda --module-first --separate --force
+	@sed -i '/^\.\. automodule:: quanda$$/,/^$$/ { /:members:/d; /:undoc-members:/d; }' source/docs_api/quanda.rst
 
 
 clean:

diff --git a/docs/source/_static/components-darkmode.png b/docs/source/_static/components-darkmode.png
diff --git a/docs/source/background.rst b/docs/source/background.rst
@@ -1,5 +1,5 @@
 What is Training Data Attribution?
-==========
+==================================
 
 The interpretability of neural network decisions is an active area of research which has seen a variety of approaches over time. Most of the initial focus was on feature attribution methods, which highlight features in the input space that are responsible for a specific prediction (`Simonyan et al., 2014 <https://arxiv.org/abs/1312.6034>`_; `Bach et al., 2015 <https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0130140>`_; `Lundberg and Lee, 2017 <https://proceedings.neurips.cc/paper_files/paper/2017/file/8a20a8621978632d76c43dfd28b67767-Paper.pdf>`_). These methods were often criticized for being unreliable and difficult to understand (`Adebayo et al., 2018 <https://proceedings.neurips.cc/paper_files/paper/2018/file/294a8ed24b1ad22ec2e7efea049b8737-Paper.pdf>`_; `Ghorbani et al., 2019 <https://ojs.aaai.org/index.php/AAAI/article/view/4252>`_). In response, researchers explored new directions, such as concept-based (`Poeta et al., 2023 <https://arxiv.org/abs/2312.12936>`_) and mechanistic interpretability (`Bereska and Gavves <https://openreview.net/forum?id=ePUVetPKu6>`_) methods. Recently, **Training Data Attribution** (TDA) has gained attention as a promising approach for enhancing the interpretability of neural networks.
 

diff --git a/docs/source/conf.py b/docs/source/conf.py
@@ -15,7 +15,7 @@
 project = "quanda"
 copyright = f"{str(datetime.utcnow().year)}, Dilyara Bareeva, Galip Ümit Yolcu"
 author = "Dilyara Bareeva, Galip Ümit Yolcu"
-release = "05.10.2024"
+release = "05.05.2026"
 
 # -- General configuration ---------------------------------------------------
 # https://www.sphinx-doc.org/en/master/usage/configuration.html#general-configuration
@@ -30,6 +30,9 @@
 ]
 source_suffix = [".rst", ".md"]
 autosummary_generate = True
+numpydoc_class_members_toctree = False
+numpydoc_show_class_members = False
+suppress_warnings = ["ref.python"]
 exclude_patterns = ["_build", "Thumbs.db", ".DS_Store"]
 rst_epilog = """
 .. |quanda| raw:: html

diff --git a/docs/source/contributing.rst b/docs/source/contributing.rst
@@ -10,7 +10,8 @@ In this guide, you will get a summary of the main components of
 
 If you have any questions regarding the codebase, please `open an
 issue <https://github.com/dilyabareeva/quanda/issues/new/choose>`__ or write us
-at dilyabareeva@gmail.com or galip.uemit.yolcu@hhi.fraunhofer.de.
+at `dilyabareeva@gmail.com <mailto:dilyabareeva@gmail.com>`__ or
+`galip.uemit.yolcu@hhi.fraunhofer.de <mailto:galip.uemit.yolcu@hhi.fraunhofer.de>`__.
 
 Table of Contents
 -----------------
@@ -122,6 +123,12 @@ Currently, the following markers are available to filter tests:
 -  benchmarks: Benchmark modules
 -  global_ranking: global_ranking modules
 -  self_influence: self_influence methods of explainers
+-  tasks: task modules
+-  integration: integration tests
+-  slow: tests marked as slow (excluded by default; run with ``pytest -m slow``)
+-  production_bench: production benchmark sanity checks, run only when explicitly specified
+
+The authoritative list lives in ``pytest.ini``.
 
 Ideally, all contributions should include tests to ensure correctness.
 
@@ -330,7 +337,7 @@ these four classmethods . What you should provide is:
       ``checkpoints_load_func``, ``device``, ``val_dataset``,
       ``use_predictions``).
    -  ``_extra_kwargs_from_config(cls, config, train_dataset,
-      eval_dataset, metadata_dir, load_meta_from_disk)`` — extract any
+      eval_dataset, metadata_dir, load_fresh)`` — extract any
       subclass-specific kwargs from the YAML and return them as a dict;
       they get passed into ``__init__`` by ``from_config``.
    -  ``_compute_and_save_indices(self, config, batch_size)`` — only

diff --git a/docs/source/explainers.rst b/docs/source/explainers.rst
@@ -0,0 +1,154 @@
+Explainer Wrappers
+==================
+
+|quanda| ships wrappers around several existing TDA libraries, exposing them
+through a single :doc:`Explainer <docs_api/quanda.explainers.base>` interface.
+The tables below list every wrapper class and cite the paper that introduced
+the underlying method. All wrapper sources live under
+``quanda/explainers/wrappers/``.
+
+All wrappers can be imported directly from ``quanda.explainers.wrappers``,
+for example:
+
+.. code:: python
+
+   from quanda.explainers.wrappers import (
+       CaptumSimilarity,
+       TRAK,
+       Kronfluence,
+       RepresenterPoints,
+       DattriIFExplicit,
+   )
+
+Captum
+------
+Wrappers around the influence methods provided by `Captum
+<https://github.com/pytorch/captum/tree/master>`_. Source:
+``quanda/explainers/wrappers/captum_influence.py``.
+
+.. list-table::
+   :header-rows: 1
+   :widths: 25 75
+
+   * - Wrapper
+     - Reference
+   * - ``CaptumSimilarity``
+     - Similarity between test and training samples in the representation
+       space of a chosen layer. See `Captum's SimilarityInfluence docs
+       <https://captum.ai/api/influence.html#similarityinfluence>`__.
+   * - ``CaptumArnoldi``
+     - Schioppa et al., 2022. *Scaling Up Influence Functions.*
+       `arXiv:2112.03052 <https://arxiv.org/abs/2112.03052>`__
+   * - ``CaptumTracInCP``
+     - Pruthi et al., 2020. *Estimating Training Data Influence by Tracing
+       Gradient Descent.* `NeurIPS 2020
+       <https://proceedings.neurips.cc/paper/2020/hash/e6385d39ec9394f2f3a354d9d2b88eec-Abstract.html>`__
+   * - ``CaptumTracInCPFast``
+     - Pruthi et al., 2020. *Estimating Training Data Influence by Tracing
+       Gradient Descent.* `NeurIPS 2020
+       <https://proceedings.neurips.cc/paper/2020/hash/e6385d39ec9394f2f3a354d9d2b88eec-Abstract.html>`__
+   * - ``CaptumTracInCPFastRandProj``
+     - Pruthi et al., 2020. *Estimating Training Data Influence by Tracing
+       Gradient Descent.* `NeurIPS 2020
+       <https://proceedings.neurips.cc/paper/2020/hash/e6385d39ec9394f2f3a354d9d2b88eec-Abstract.html>`__
+
+Representer Point Selection
+---------------------------
+Source: ``quanda/explainers/wrappers/representer_points.py``.
+
+.. list-table::
+   :header-rows: 1
+   :widths: 25 75
+
+   * - Wrapper
+     - Reference
+   * - ``RepresenterPoints``
+     - Yeh et al., 2018. *Representer Point Selection for Explaining Deep
+       Neural Networks.* `NeurIPS 2018
+       <https://proceedings.neurips.cc/paper/2018/hash/8a7129b8f3edd95b7d969dfc2c8e9d9d-Abstract.html>`__
+       — original implementation: `chihkuanyeh/Representer_Point_Selection
+       <https://github.com/chihkuanyeh/Representer_Point_Selection>`__
+
+TRAK
+----
+Source: ``quanda/explainers/wrappers/trak_wrapper.py``.
+
+.. list-table::
+   :header-rows: 1
+   :widths: 25 75
+
+   * - Wrapper
+     - Reference
+   * - ``TRAK``
+     - Park et al., 2023. *TRAK: Attributing Model Behavior at Scale.*
+       `ICML 2023 <https://proceedings.mlr.press/v202/park23c.html>`__
+       — original implementation: `MadryLab/trak
+       <https://github.com/MadryLab/trak>`__
+
+Kronfluence
+-----------
+Source: ``quanda/explainers/wrappers/kronfluence.py``.
+
+.. list-table::
+   :header-rows: 1
+   :widths: 25 75
+
+   * - Wrapper
+     - Reference
+   * - ``Kronfluence (incl. EK-FAC)``
+     - Grosse et al., 2023. *Studying Large Language Model Generalization with
+       Influence Functions.* `arXiv:2308.03296
+       <https://arxiv.org/abs/2308.03296>`__
+       — original implementation: `pomonam/kronfluence
+       <https://github.com/pomonam/kronfluence>`__
+
+Dattri
+------
+Wrappers around the unified TDA family provided by `Dattri
+<https://github.com/TRAIS-Lab/dattri>`_ (Deng et al., 2024,
+`arXiv:2410.04555 <https://arxiv.org/abs/2410.04555>`__). Source:
+``quanda/explainers/wrappers/dattri_influence.py``.
+
+.. list-table::
+   :header-rows: 1
+   :widths: 25 75
+
+   * - Wrapper
+     - Reference
+   * - ``DattriIFExplicit``
+     - Koh and Liang, 2017. *Understanding Black-box Predictions via Influence
+       Functions.* `ICML 2017
+       <https://proceedings.mlr.press/v70/koh17a.html>`__
+   * - ``DattriIFCG``
+     - Koh and Liang, 2017 (conjugate-gradient solver). `ICML 2017
+       <https://proceedings.mlr.press/v70/koh17a.html>`__
+   * - ``DattriIFLiSSA``
+     - Agarwal et al., 2017. *Second-Order Stochastic Optimization for Machine
+       Learning in Linear Time.* `JMLR 2017
+       <https://www.jmlr.org/papers/v18/16-491.html>`__
+   * - ``DattriIFDataInf``
+     - Kwon et al., 2024. *DataInf: Efficiently Estimating Data Influence in
+       LoRA-tuned LLMs and Diffusion Models.* `ICLR 2024
+       <https://openreview.net/forum?id=9m02ib92Wz>`__
+   * - ``DattriArnoldi``
+     - Schioppa et al., 2022. *Scaling Up Influence Functions.*
+       `arXiv:2112.03052 <https://arxiv.org/abs/2112.03052>`__
+   * - ``DattriEKFAC``
+     - Grosse et al., 2023. *Studying Large Language Model Generalization with
+       Influence Functions.* `arXiv:2308.03296
+       <https://arxiv.org/abs/2308.03296>`__
+   * - ``DattriTracInCP``
+     - Pruthi et al., 2020. *Estimating Training Data Influence by Tracing
+       Gradient Descent.* `NeurIPS 2020
+       <https://proceedings.neurips.cc/paper/2020/hash/e6385d39ec9394f2f3a354d9d2b88eec-Abstract.html>`__
+   * - ``DattriGradDot``
+     - Charpiat et al., 2019. *Input Similarity from the Neural Network
+       Perspective.* `NeurIPS 2019
+       <https://proceedings.neurips.cc/paper_files/paper/2019/hash/c61f571dbd2fb949d3fe5ae1608dd48b-Abstract.html>`__
+   * - ``DattriGradCos``
+     - Charpiat et al., 2019. *Input Similarity from the Neural Network
+       Perspective.* `NeurIPS 2019
+       <https://proceedings.neurips.cc/paper_files/paper/2019/hash/c61f571dbd2fb949d3fe5ae1608dd48b-Abstract.html>`__
+   * - ``DattriTRAK``
+     - Park et al., 2023. *TRAK: Attributing Model Behavior at Scale.*
+       `ICML 2023 <https://proceedings.mlr.press/v202/park23c.html>`__
diff --git a/docs/source/how_to_evaluate.rst b/docs/source/how_to_evaluate.rst
@@ -1,5 +1,5 @@
 How to Assess the Quality of Attributions?
-==========
+==========================================
 
 Evaluation of interpretability approaches is a challenging task, as it is often difficult to define a ground truth for interpretability. Although there are various demonstrations of TDA’s potential for interpretability and practical applications, the critical question of how TDA methods should be effectively evaluated remains open. While methods based on estimating counterfactual retraining effects have a well-defined ground truth, this ground truth is computationally demanding and is not feasibly computable for large scale experiments. To address these shortcomings, several approaches have been proposed by the community, which can be categorized into three groups:
 
@@ -17,7 +17,7 @@ As some of the methods are designed to approximate LOO effects, ground truth can
 
     <details><summary><b><big>Downstream Task Evaluators</big></b></summary>
 
-To remedy the challenges associated with ground truth evaluation, the literature proposes to assess the utility of a TDA method within the context of an end-task. The most commonly used evaluation criteria is Mislabeling Detection (`Koh and Liang, 2017 <https://proceedings.mlr.press/v70/koh17a.html>`_; `Yeh et al., 2018 <https://proceedings.neurips.cc/paper/2018/hash/8a7129b8f3edd95b7d969dfc2c8e9d9d-Abstract.html>`_; `Pruthi et al., 2020 <https://proceedings.neurips.cc/paper/2020/hash/e6385d39ec9394f2f3a354d9d2b88eec-Abstract.html>`_) which compares different TDA methods in terms of their usefulness for detecting mislabeled samples after training the network on a dataset of which the labels are deliberately poisoned. Other examples could be detecting backdoor attacks (`Karthikeyan et al., 2021 <https://arxiv.org/abs/2111.04683>`_; `Yolcu et al., 2024 <https://arxiv.org/abs/2402.12118>`_) or predicting the model decision from its attributions (`Hanawa et al., 2021 <https://openreview.net/forum?id=9uvhpyQwzM_>`_).
+To remedy the challenges associated with ground truth evaluation, the literature proposes to assess the utility of a TDA method within the context of an end-task. The most commonly used evaluation criteria is Mislabeling Detection (`Koh and Liang, 2017 <https://proceedings.mlr.press/v70/koh17a.html>`_; `Yeh et al., 2018 <https://proceedings.neurips.cc/paper/2018/hash/8a7129b8f3edd95b7d969dfc2c8e9d9d-Abstract.html>`_; `Pruthi et al., 2020 <https://proceedings.neurips.cc/paper/2020/hash/e6385d39ec9394f2f3a354d9d2b88eec-Abstract.html>`_) which compares different TDA methods in terms of their usefulness for detecting mislabeled samples after training the network on a dataset of which the labels are deliberately poisoned. Other examples could be detecting backdoor attacks (`Karthikeyan et al., 2021 <https://arxiv.org/abs/2111.04683>`_; `Yolcu et al., 2025 <https://openreview.net/forum?id=qfx81N884A>`_) or predicting the model decision from its attributions (`Hanawa et al., 2021 <https://openreview.net/forum?id=9uvhpyQwzM_>`_).
 
 .. raw:: html
-Original file line number
+Diff line change
@@ Expand Up / @@ -22,6 +22,7 @@ help: @@
     rst:
     	@sphinx-apidoc -o source/docs_api ../quanda --module-first --separate --force
+    	@sed -i '/^\.\. automodule:: quanda$$/,/^$$/ { /:members:/d; /:undoc-members:/d; }' source/docs_api/quanda.rst
     clean:
@@ Expand Down @@