catt-bench/route-attribution-baselines

route attribution scores (Rule Routings)

The attribution scores of "Elucidating the Design Space of LM Data Attribution" on Rule Routings, for the LoRA adapters in catt-bench/Qwen3.5-4B-route-ab-l8-lora-scale. The CATT code release reads them with catt fetch results --task rr and rebuilds the paper's figures and tables from them with catt report results.

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

run is route_ab_c{F}_{N}d_sd{seed} with F in {25, 50, 100}, N in {1, 3, 5, 10, 30, 50} and seeds 1001-1003, matching the adapter repository's subfolders. Not every group covers every cell. The trajectory groups (tracin, source, tracinfull) cover seeds 1001 and 1002, the seeds with a saved trajectory; bm25 reads no adapter and is scored at seed 1001 only.

group	Methods (file names)	catt attribute --group
ekfac	EK-FAC, Grad-Dot, Grad-Cos (if_ekfac, grad_dot, grad_sim)	flat
sketch	TRAK, TRAK without Q, LoGRA, LoGRA-Dot, TrackStar (trak, trak_nores, logra, logra_dot, trackstar)	sketch
trakjl	TRAK with a dense JL projection (trak_jl_*)	dense-trak
tracin	layerwise TracIn, TracIn-Cos (tracin, tracin_cos)	tracin-layerwise
source	SOURCE (source)	source
tracinfull	TracIn, TracIn-Cos on unprojected gradients (tracin_full, tracin_cos_full)	tracin-exact
bm25, repsim	BM25, RepSim	retrieval

bm25_prompt and repsim_mean are scored alongside but not in the paper.

Arms

The measurement (query) side is always the negative log-odds of the answer. The training side is one of:

Arm	Training-side score
ce	cross-entropy gradient, summed over tokens
ce_top3, ce_top10	mean of the 3 / 10 largest per-token CE terms
margin	negative log-odds gradient, averaged over supervised tokens
margin_sum	the same, summed
margin_docw	margin x (1 - p), the paper's "mean x Q"
margin_top3, margin_top10	mean of the 3 / 10 largest per-token margin terms

Every score is oriented so that higher means more helpful.

A scores file

scores          float32 [n_query, n_train]
train_uids      corpus uid of each column
train_func      the fact each training document describes
train_role      "constant" for a real document, "distractor" for a decoy
train_source    the fact a decoy shadows, empty otherwise
query_uids      query uid of each row
query_func      the fact each query asks about
query_correct   whether the model answers the query correctly
method, arm, convention, orientation, git_commit, git_dirty

Queries are capped at four per fact for the gradient methods; bm25 and repsim were scored on all five.

History

This layout dates from 2026-09-30. Before it, these scores sat under v2/ (the paper's code release pins revisions from that time), and the top level held an earlier grid, v1, together with the sweep's claim files. v1 was computed before a padding fix in the per-sample gradients and under other conventions; the paper does not use it. It is kept under v1/ with its original layout. The BGE and base-model embedding scores (v2/*/embed), which the paper does not use, were dropped. Revision b1e8b831871f166633329c398ad376547f28c567 is the last one with the old layout.