Elucidating the Design Space of LM Data Attribution
Averaging scores across whole training documents can bury the specific tokens that reveal where a model learned a fact.
The researchers introduce CATT, a benchmark built from fictitious knowledge that gives them clear ground truth for tracing facts to training data. Changing one dataset property at a time shows that document redundancy and lexical distractors consistently worsen attribution. Using negative log-odds gradients and aggregating only the highest-scoring tokens improves ground-truth recall by 35–72 percentage points across multiple baselines on CATT. The gradient correction also improves counterfactual retraining performance for every method tested on FTRACE-TREX, by up to 4.9 times.
ArXiv · AI/CL/LG's note
The researchers introduce CATT, a benchmark built from fictitious knowledge that gives them clear ground truth for tracing facts to training data. Changing one dataset property at a time shows that document redundancy and lexical distractors consistently worsen attribution. Using negative log-odds gradients and aggregating only the highest-scoring tokens improves ground-truth recall by 35–72 percentage points across multiple baselines on CATT. The gradient correction also improves counterfactual retraining performance for every method tested on FTRACE-TREX, by up to 4.9 times.
ArXiv · AI/CL/LG's note
score 5