From d52b7ece7d11930eae04b4237abc6fa67a815bd4 Mon Sep 17 00:00:00 2001
From: wassname <1103714+wassname@users.noreply.github.com>
Date: Fri, 10 Jul 2026 20:35:29 +0800
Subject: [PATCH] persona_topk: log j-thoughts as positive:/negative: (was
 pos>neg/neg>pos)

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
---
 jsteer/jacobian.py | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/jsteer/jacobian.py b/jsteer/jacobian.py
index 9e1929d..16da357 100644
--- a/jsteer/jacobian.py
+++ b/jsteer/jacobian.py
@@ -289,10 +289,13 @@ class Jacobian:
         diff = logits_pos - logits_neg                           # [vocab]
         top_pos = diff.topk(k)                                    # pos evokes > neg
         top_neg = (-diff).topk(k)                                 # neg evokes > pos
-        for name, sel in (("pos>neg", top_pos), ("neg>pos", top_neg)):
-            toks = [tok.decode([i]) for i in sel.indices.tolist()]
-            logger.info(f"persona_topk {name} top-{k}: {toks}")  # SHOULD be persona-
-            # specific words now, not generic \n/the; generic here => contrast is null.
+        toks_pos = [tok.decode([i]) for i in top_pos.indices.tolist()]
+        toks_neg = [tok.decode([i]) for i in top_neg.indices.tolist()]
+        # SHOULD be persona-specific words (positive vs negative affect here), not
+        # generic \n/the; generic on both sides means the contrast is null.
+        logger.info(f"j-thoughts (content of mental workspace, top-{k})\n"
+                    f"    positive: {toks_pos}\n"
+                    f"    negative: {toks_neg}")
         w = (W_U[top_pos.indices].float().mean(0)
              - W_U[top_neg.indices].float().mean(0)).cpu()
         cfg = JacobianPersonaTopkC(layers=self._steer_layers(layers))