Data Attribution
A methodological framework for quantifying and analyzing the contribution of each training sample to a given model prediction. Common techniques include influence functions, gradient similarity, and leave-one-out methods. It is used for model debugging, auditing, and copyright tracing.
Timeline (last 90 days)
Existing language model data attribution evaluations suffer from variable coupling, that is, the evaluation set simultaneously varies multiple dataset properties such as redundant documents, lexical distractors and text length, making it impossible to pinpoint the specific reason an attribution algorithm fails.
Data attribution aims to quantify the contribution of each sample in the training data to a given prediction of the model, and common methods include influence functions, gradient similarity (such as TracIn) and leave-one-out.
Standard data attribution methods average attribution scores over the entire document, which dilutes the key signal.
Counterfactual retraining is a downstream metric for evaluating data attribution quality: after removing the key documents identified by attribution, the model is retrained and the degree of change in the target prediction is observed, and this metric is closer to practical application than recall.
The signal this study conveys is that the performance bottleneck of data attribution may lie in peripheral designs such as the loss function and aggregation strategy, rather than in the estimator itself.
All Facts (5)
Existing language model data attribution evaluations suffer from variable coupling, that is, the evaluation set simultaneously varies multiple dataset properties such as redundant documents, lexical distractors and text length, making it impossible to pinpoint the specific reason an attribution algorithm fails.
50%UnverifiedData attribution aims to quantify the contribution of each sample in the training data to a given prediction of the model, and common methods include influence functions, gradient similarity (such as TracIn) and leave-one-out.
50%UnverifiedStandard data attribution methods average attribution scores over the entire document, which dilutes the key signal.
50%UnverifiedCounterfactual retraining is a downstream metric for evaluating data attribution quality: after removing the key documents identified by attribution, the model is retrained and the degree of change in the target prediction is observed, and this metric is closer to practical application than recall.
50%UnverifiedThe signal this study conveys is that the performance bottleneck of data attribution may lie in peripheral designs such as the loss function and aggregation strategy, rather than in the estimator itself.
50%