Upload scripts/mk_recipe.py with huggingface_hub

scripts/mk_recipe.py

@@ -0,0 +1,52 @@
+#!/usr/bin/env python3
+"""The recipe figure: compression helps at INFERENCE, destroys at TRAINING."""
+import matplotlib; matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+S={"surface":"#fcfcfb","ink":"#0b0b0b","ink2":"#52514e","grid":"#e4e4e1"}
+def sty(ax):
+    ax.set_facecolor(S["surface"])
+    for s in ax.spines.values(): s.set_color(S["grid"])
+    ax.grid(True,color=S["grid"],lw=.7,alpha=.9); ax.tick_params(colors=S["ink2"],labelsize=9)
+fig,(a,b,c)=plt.subplots(1,3,figsize=(17.4,5.4),facecolor=S["surface"])
+
+# A: inference-side -- compression helps
+sc=[8,16,32,64]; v=[34,54,79,79]; v2=[None,32,46,75]
+a.plot(sc,v,"o-",color="#12508f",lw=2.4,ms=8,label="32,744 tokens")
+a.plot(sc[1:],v2[1:],"s-",color="#4a8fd4",lw=2.4,ms=8,label="65,512 tokens")
+sty(a); a.set_xscale("log",base=2); a.set_xticks(sc); a.set_xticklabels([f"x{s}" for s in sc])
+a.set_ylim(0,100); a.set_xlabel("YARN_SCALE at INFERENCE",color=S["ink2"],fontsize=10)
+a.set_ylabel("niah_single (%)",color=S["ink2"],fontsize=10)
+a.set_title("At INFERENCE, compression HELPS",color=S["ink"],fontsize=12,loc="left",pad=10)
+a.legend(frameon=False,fontsize=9,labelcolor=S["ink2"],loc="lower right")
+a.text(9,12,"content moves into angles\nthe model saw in pretraining",fontsize=8,color=S["ink2"],style="italic")
+
+# B: training-side -- compression destroys
+labels=["pretrain\nT=4096 x1","p1final train\nT=16k x4\n(invariant)","t32 train\nT=32k x32\n(invariant BROKEN)"]
+dens=[1,4,32]; res=[None,90,51]
+bars=b.bar(range(3),dens,color=["#b0b0aa","#12508f","#b3243b"])
+sty(b); b.set_yscale("log",base=2); b.set_xticks(range(3)); b.set_xticklabels(labels,fontsize=8.4)
+b.set_ylabel("tokens per virtual position unit",color=S["ink2"],fontsize=10)
+b.set_title("At TRAINING, compression DESTROYS",color=S["ink"],fontsize=12,loc="left",pad=10)
+for i,(d,r) in enumerate(zip(dens,res)):
+    b.text(i,d*1.25,f"{d}x",ha="center",fontsize=10,color=S["ink2"])
+    if r: b.text(i,1.05,f"scores {r}",ha="center",fontsize=9.5,color="white",
+                 bbox=dict(fc="#12508f" if r>70 else "#b3243b",ec="none",pad=2.6))
+
+# C: the recipe
+c.axis("off"); c.set_facecolor(S["surface"])
+c.text(.02,.95,"THE RECIPE",fontsize=13,color=S["ink"],weight="bold",va="top")
+c.text(.02,.845,"TRAIN with the invariant intact:\n    SCALE x ORIG == CP_T\n"
+                "  -> uses the FULL virtual range 0..4095\n  -> minimum compression for that length\n"
+                "  -> positional sensitivity preserved",
+       fontsize=10,color=S["ink2"],va="top",family="monospace")
+c.text(.02,.545,"INFER at a HIGHER scale:\n    SCALE ~= prompt_len / 1500\n"
+                "  -> lands content near virtual 1000-2000\n  -> worth up to +52pp, free",
+       fontsize=10,color=S["ink2"],va="top",family="monospace")
+c.text(.02,.31,"p1final does exactly this:\n  trained T=16,384 x4 (invariant OK)\n  evaluated at x64  ->  95 / 90 / 82",
+       fontsize=10,color="#12508f",va="top",family="monospace")
+c.text(.02,.13,"What I did wrong:\n  trained T=32,768 x32 (invariant broken)\n  90 -> 51, collapsed by step 800",
+       fontsize=10,color="#b3243b",va="top",family="monospace")
+fig.text(.005,.005,"680M, niah_single, N=100, identical haystack. Compression = tokens per virtual position unit = SCALE.",
+         fontsize=8,color=S["ink2"])
+fig.tight_layout(); fig.savefig("results_2026-09-16/recipe.png",dpi=150,bbox_inches="tight")
+print("  wrote results_2026-09-16/recipe.png")