Each entry describes one MLP neuron of the language model Qwen3-32B, in the words of Goodfire's own interpretability pipeline: title, lede (their explanation after an agentic investigation) and judge_clean_note (their judge's note on whether each pole's strongest firings come down to one idea).
A neuron has two POLES: the positive pole (inputs that drive it strongly positive) and the negative pole (inputs that drive it strongly negative). Judge EACH POLE separately, using ONLY what the texts say (no outside beliefs; no benefit of the doubt in either direction). A positive pole and a negative pole doing different things is normal and is NOT a mixture: judge each pole on its own.
Verdict per pole:
- clean: the pole's strong firings come down to one main job;
- clean_with_residual: one main job, plus a residue the texts treat as minor (small offshoots, a few stray or unexplained examples, a weak branch);
- two_related: two distinct jobs of comparable weight that are related to each other;
- two_or_more_unrelated: two or more jobs of non-minor weight that are unrelated to each other;
- unresolved: the texts say the pole is largely not understood, or that a large share (around a quarter or more) of its strong firings is unexplained or fails to reproduce;
- not_described: the texts say nothing about this pole.
residue: for clean_with_residual, whether the minor residue is related to the main job, unexplained, or unrelated; otherwise 'none' (or the best fit if the verdict involves a residue).
evidence: a short quote (under 20 words) that decided it.
