Research map / Label-efficient and robust learning / Recognition and long-tail

quanda: An Interpretability Toolkit for Training Data Attribution Evaluation

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Prototype and concept dictionary models

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces quanda, a PyTorch toolkit that unifies explainers (wrappers for TDA methods), evaluation metrics, and precomputed benchmark suites for training data attribution evaluation.

Problem

Training data attribution (TDA) evaluation is scattered across papers, relies on inconsistent assumptions, and is rarely reproducible, which makes fair comparison of TDA methods difficult.

Approach

A modular Python library with a uniform explainer interface wrapping existing TDA implementations (Captum, TRAK), stateful metrics across ground-truth, downstream, and heuristic categories, and benchmark classes that generate or download controlled setups for systematic comparison.

Key results

• Library provides explainer wrappers for TDA methods including Captum-based SimilarityExplainer, Arnoldi IF, TracInCP variants, Representer Points, and TRAK.
• Test coverage reaches 90% at submission time with pytest, mypy, black, isort and flake8 in CI.
• Demonstration on Tiny ImageNet with ResNet-18 (43% and 56% top-1 accuracy on the two modified datasets) compares RP, IF-Arnoldi, TracIn (last layer), TRAK, and a random baseline across metrics.

Surprising findings

• Ground-truth counterfactual values are reported to be dominated by noise in practical deep learning settings due to training stochasticity.

Limitations

• Currently supports only image classification and post-hoc attribution of a single model.
• TRAK implemented with a single model instance, reducing attribution quality relative to the original multi-model method.
• Arnoldi IF and TracIn use only final-layer parameters for feasibility.

Experimental setup

Datasets: Tiny ImageNet ImageNet-Sketch ImageNet

Benchmarks: Tiny ImageNet controlled setup (quanda metrics)

Metrics: LDS class detection subclass detection mislabeling detection AUC shortcut detection AUPRC mixed datasets AUPRC model randomization top-k cardinality

Baselines: Representer Points Influence Functions with Arnoldi iterations TracInCP TRAK random explainer

Builds on: Captum TRAK Quantus OpenXAI Linear Datamodeling Score

Closest papers by meaning

• Final-Model-Only Data Attribution with a Unifying View of Gradient-Based Methods (NeurIPS 2025 · similarity 0.90)
• $\texttt{dattri}$: A Library for Efficient Data Attribution (NeurIPS 2024 · similarity 0.89)
• DATE-LM: Benchmarking Data Attribution Evaluation for Large Language Models (NeurIPS 2025 · similarity 0.87)
• Fast Data Attribution for Text-to-Image Models (NeurIPS 2025 · similarity 0.87)
• Taming Hyperparameter Sensitivity in Data Attribution: Practical Selection Without Costly Retraining (NeurIPS 2025 · similarity 0.87)
• Training Data Attribution in Diffusion Models via Mirrored Unlearning and Noise-Consistent Skew (NeurIPS 2026 · similarity 0.86)