Upload tables/scripts/make_numbers.py with huggingface_hub

tables/scripts/make_numbers.py

@@ -1,9 +1,9 @@
 """Emit the numbers the body quotes, as LaTeX macros, from the run outputs.
 
-Two files, both \input by paper.tex:
+One file, \input by paper.tex:
 
-    sections/numbers.tex           scalar macros (\PCTONE, \DIRECTIONPARA, ...)
-    sections/table_null_body.tex   the rows of Table 1
+    sections/numbers.tex   scalar macros (\PCTONE, \DIRECTIONPARA, ...) and
+                           \TABBODY, the rows of Table 2
 
 Anything here is a number the reviewer can trace: it is read from the CSVs the
 runs wrote, never typed. Re-run after any rerun and diff the two files.
@@ -171,6 +171,12 @@ def main():
             per_step.append((lay, math.floor(min(sh)), math.ceil(max(sh))))
     L.append("\\newcommand{\\SHAREBYSTEP}{" + ", ".join(
         f"${lo}\\text{{--}}{hi}\\%$ at layer {lay}" for lay, lo, hi in per_step) + "}")
+    # The body quotes the envelope over every step and layer; the per-layer
+    # breakdown above is kept for the appendix.
+    if per_step:
+        blo = min(lo for _, lo, _ in per_step)
+        bhi = max(hi for _, _, hi in per_step)
+        L.append(f"\\newcommand{{\\SHAREBAND}}{{${blo}\\text{{--}}{bhi}\\%$}}")
     print("null share across all seven steps: " + ", ".join(
         f"L{lay} {lo}-{hi}%" for lay, lo, hi in per_step))
 
@@ -179,7 +185,7 @@ def main():
     L.append(f"\\newcommand{{\\NMSEMAX}}{{{f['nmse']:.4f}}}")
     L.append(f"\\newcommand{{\\FRACRECMIN}}{{{f['frac_rec']:.3f}}}")
     L.append(f"\\newcommand{{\\DEADMAX}}{{{f['dead']:.1f}}}")
-    L.append(f"\\newcommand{{\\SEPRATIO}}{{{min(ratios):.0f}$ to ${max(ratios):.0f}}}")
+    L.append(f"\\newcommand{{\\SEPRATIO}}{{{math.floor(min(ratios))}$ to ${math.ceil(max(ratios))}}}")
     # NSEEDS describes every row of Table 1, so it has to be the range across
     # layers, not layer 18's counts. L6 and L12 carry 3 null chains, L18 has 4;
     # printing "4/4" would overstate replication for two of the three rows.
@@ -188,25 +194,34 @@ def main():
     fmt = lambda x: f"${min(x)}$" if len(x) == 1 else f"${min(x)}$ to ${max(x)}$"
     L.append(f"\\newcommand{{\\NSEEDS}}{{{fmt(reals)} real and {fmt(nulls)} null "
              "chains per layer}")
+    rerun_dead = {}
+    for lay in LAYERS:
+        csvp = DRIFT / f"real_arm_L{lay}" / "real_arm_replicate.csv"
+        if csvp.exists():
+            v = pd.read_csv(csvp)["dead_frac"] * 100
+            rerun_dead[lay] = (math.floor(v.min()), math.ceil(v.max()))
+    if 18 in rerun_dead and 12 in rerun_dead:
+        L.append("\\newcommand{\\RERUNDEAD}{$%d$ to $%d\\%%$ at layer 18 and $%d$ to $%d\\%%$ "
+                 "at layer 12}" % (*rerun_dead[18], *rerun_dead[12]))
+        print(f"rerun dead fractions: L18 {rerun_dead[18]}, L12 {rerun_dead[12]}")
+
     q = resid[(18, "1")]
     L.append(f"\\newcommand{{\\RESIDL}}{{${q.mean():.4f}$ "
              f"$[{q.min():.4f},{q.max():.4f}]$}}")
-    L.append(f"\\newcommand{{\\RESIDRATIO}}{{{ratios[LAYERS.index(18)*2]:.0f}}}")
+    L.append(f"\\newcommand{{\\RESIDRATIO}}{{{math.floor(ratios[LAYERS.index(18)*2])}}}")
 
     d = direction(18)
     if d and 7 in d:
         end, one = d[7], d.get(1, d[7])
         L.append("\\newcommand{\\DIRECTIONPARA}{%\n"
-                 f"At layer 18, over {end['n_real']} real and {end['n_null']} null chains, "
-                 "the seven-step endpoints come apart completely "
-                 "(Figure~\\ref{fig:main}d). Two real chains agree with each other at a mean "
-                 f"per-latent decoder cosine of {rng(end['rr'])} and two null chains at "
-                 f"{rng(end['nn'])}, but a real chain agrees with a null chain at only "
-                 f"{rng(end['rn'])}. All ${len(end['rr'])+len(end['nn'])}$ within-arm pairs sit "
-                 f"above all ${len(end['rn'])}$ cross-arm pairs, and they still do after a "
-                 "single step. So fine-tuning leaves a reproducible signature in the "
-                 "\\emph{direction} of the move, which the magnitude gives no way to isolate "
-                 "because retraining supplies most of it.}")
+                 f"At layer 18, over {end['n_real']} real and {end['n_null']} null chains, two "
+                 f"real endpoints agree at a mean matched-latent cosine of "
+                 f"${end['rr'].mean():.3f}$, two null endpoints at ${end['nn'].mean():.3f}$, and "
+                 f"a real endpoint with a null endpoint at only ${end['rn'].mean():.3f}$. Every "
+                 "within-arm pair therefore exceeds every real--null pair, both after the first "
+                 "step and after the seventh (Figure~\\ref{fig:main}b). Changes in the model "
+                 "activations affect the direction of SAE movement, even though retraining "
+                 "accounts for most of its magnitude.}")
         print(f"direction: step7 rr={end['rr'].mean():.4f} nn={end['nn'].mean():.4f} "
               f"rn={end['rn'].mean():.4f} separated={end['separated']}")
     else: