Remove 'Why base-model gradients predict memorization' panel + dead helper code

app.py

@@ -40,14 +40,6 @@ DIST = {ds: {"mem": LOOKUP[(LOOKUP.dataset == ds) & LOOKUP.memorized],
 GREEN, RED = "#2e7d32", "#c62828"          # generation highlight
 MEMCLR, NONCLR = "#f58518", "#4c78a8"      # memorized (orange) vs not-memorized (blue)
 
-# Headline ablation result (paper: dedup, mean +/- std over 3 training seeds), demo datasets only.
-HEADLINE = {
-    "FineWeb":     {"text-only": (0.628, 0.011), "gradient-only": (0.994, 0.003), "all": (0.997, 0.002)},
-    "PG-19":       {"text-only": (0.640, 0.035), "gradient-only": (0.998, 0.001), "all": (0.996, 0.003)},
-    "The Stack":   {"text-only": (0.765, 0.001), "gradient-only": (0.965, 0.001), "all": (0.971, 0.001)},
-    "OpenWebMath": {"text-only": (0.837, 0.001), "gradient-only": (0.976, 0.002), "all": (0.992, 0.001)},
-}
-
 # ZeroGPU: place the base model on cuda at module level (a CUDA-emulation mode makes
 # this valid at startup; a real GPU is attached only inside @spaces.GPU functions).
 _ON_ZEROGPU = bool(os.environ.get("SPACES_ZERO_GPU"))
@@ -66,25 +58,6 @@ def _compute_features(text):
 
 
 # ------------------------------- graphics ----------------------------------
-def headline_figure():
-    tiers = ["text-only", "gradient-only", "all"]
-    colors = {"text-only": "#9aa0a6", "gradient-only": GREEN, "all": "#1f4e79"}
-    datasets = list(HEADLINE.keys())
-    x = np.arange(len(datasets)); w = 0.26
-    fig, ax = plt.subplots(figsize=(7.2, 3.6))
-    for i, t in enumerate(tiers):
-        vals = [HEADLINE[d][t][0] for d in datasets]
-        errs = [HEADLINE[d][t][1] for d in datasets]
-        ax.bar(x + (i - 1) * w, vals, w, yerr=errs, capsize=3, label=t, color=colors[t])
-    ax.set_xticks(x); ax.set_xticklabels(datasets)
-    ax.set_ylim(0.5, 1.02); ax.set_ylabel("ROC-AUC")
-    ax.set_title("Gradient-only features ≈ full model (mean ± std over 3 seeds)")
-    ax.axhline(0.5, color="k", lw=0.7, ls=":")
-    ax.legend(loc="lower right", ncol=3, fontsize=8, framealpha=0.9)
-    fig.tight_layout()
-    return fig
-
-
 def dist_figure(feats, dataset):
     """Six panels: where this example's feature values fall relative to the memorized
     and not-memorized populations of the chosen dataset. Model-free."""
@@ -187,19 +160,9 @@ fine-tuned model will memorize. **Explore** verifies this against ground truth o
 a fresh fine-tuning run; **Score your own text** applies the same classifier to a passage you provide.
 """
 
-WHY = """The classifier reads six features of the **base** model, measured **before** any fine-tuning.
-Three come from a single gradient step on the sequence (**gradient norm** and the **one-step loss change**);
-the rest are text and base-loss statistics. On the datasets shown here, the three **gradient-only** features
-alone reach ROC-AUC 0.965–0.998, essentially matching the full feature set, while text-only features lag far
-behind. (On conversational text, not shown here, the non-gradient features contribute more.)
-"""
-
 with gr.Blocks(title="Predicting Memorization Before Fine-Tuning") as demo:
     gr.Markdown(INTRO)
     with gr.Tab("Explore held-out examples"):
-        with gr.Accordion("Why base-model gradients predict memorization", open=False):
-            gr.Markdown(WHY)
-            gr.Plot(headline_figure())
         gr.Markdown("Pick a real Pythia-1.4B fine-tuning sequence and compare the classifier's "
                     "pre-fine-tuning risk score with what the fine-tuned model actually generated.")
         with gr.Row():