This is a working tool, not a write-up: 30 SAE (sparse autoencoder) features from the OpenVLA/LIBERO-Goal interpretability replication, waiting on a human call. For each one below: look at the activation plot + the episodes it fires on, and decide general (fires on a real recurring behavior across diverse episodes/tasks) or memorized (fires on a narrow, near-identical subset of episodes) — or skip if it's genuinely ambiguous.
Why this exists: the classifier that separates "general" from "memorized" SAE features has to be trained on human judgment — there's no way to derive it automatically. We already have one classifier (borrowed from the paper) for the original dictionary size, but we just trained a 4× bigger dictionary, and the paper's own protocol says you need a fresh batch of 30 human-labeled examples from that specific dictionary before refitting.
Read the per-task table first (this is the v3 fix): coverage alone (what fraction of episodes a feature touches) is misleading — many "general" candidates fire, at least weakly, in every episode, so the old "top-6 highest-activation episodes" view was a biased slice that often showed just 1–2 tasks even for a feature whose signal is genuinely spread. Each card now has a per-task breakdown: for all 10 LIBERO-Goal tasks, how many episodes it fires on and how strongly (mean peak when active, shown as a bar). Judge spread vs. concentration from that table. The general candidates were re-selected to reward features whose activation magnitude is spread across tasks (not just technically active everywhere); memorized candidates are narrow by construction.
Also worth knowing: the paper's fixed activation threshold (0.1) is miscalibrated for this dictionary — no feature in this layer ever crosses it, so the paper's "mean onset count" / "run length" columns would read exactly 0 and are omitted here. Judge from the per-task table and the plot shape.
0 / 30 labeled
Your picks are saved in this browser (localStorage) as you go — safe to close and come back. When you're done (or done enough), tap "Copy labels JSON" and send that back so it can be dropped into labels.json for 0.11 (refitting the classifier).