vtok101-distr-attribution-baselines

vtok101-distr attribution scores (Function Bindings with distractors)

The attribution scores of "Elucidating the Design Space of LM Data Attribution" on Function Bindings with distractors, for the LoRA adapters in lamsheeper-data-attribution/Qwen3.5-4B-d0-vtok101-distr-lora-seeds. The CATT code release reads them with catt fetch results --task fb-distr and rebuilds the paper's figures and tables from them with catt report results.

Every real document has a decoy beside it: a shadow function <Axx> taught the same constant in the same template as <Bxx>. The pool is twice as large and queries only ask about real functions, so AUC keeps its meaning while Recall@N's chance line halves. metrics.json adds decoy_auc (a real document against the query's own decoys) and decoy_top1.

Layout

{run}/{group}/scores.{method}.{arm}.npz   one [query, train] score matrix
{run}/{group}/scores.npz                   the group's first method and arm
{run}/{group}/tokens.{method}.{loss}.npz  per-token terms [query, token]
{run}/{group}/tokens.meta.npz              token offsets, labels and p(correct)
{run}/{group}/metrics.json                 Recall@N, AUC, per-fact metrics, "method@arm"
{run}/{group}/config.json                  every argument, plus provenance