unparallelled commited on
Commit
e0cce9b
·
verified ·
1 Parent(s): 856d2d1

Upload scripts/mk_scale_plot.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. scripts/mk_scale_plot.py +89 -0
scripts/mk_scale_plot.py ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """The scale sweep: retrieval vs YaRN scale, and the two-factor grid.
3
+ Every number read from the summary.csv files this session produced."""
4
+ import csv, os, glob
5
+ import matplotlib; matplotlib.use("Agg")
6
+ import matplotlib.pyplot as plt
7
+
8
+ S={"surface":"#fcfcfb","ink":"#0b0b0b","ink2":"#52514e","grid":"#e4e4e1"}
9
+ def style(ax):
10
+ ax.set_facecolor(S["surface"])
11
+ for s in ax.spines.values(): s.set_color(S["grid"])
12
+ ax.grid(True,color=S["grid"],lw=.7,alpha=.9); ax.tick_params(colors=S["ink2"],labelsize=9)
13
+
14
+ def read(path,model):
15
+ if not os.path.exists(path): return None
16
+ r=list(csv.reader(open(path)))
17
+ if len(r)<2: return None
18
+ hdr=r[0]
19
+ for row in r[1:]:
20
+ if row and row[0]=="niah_single" and model in hdr:
21
+ try: return float(row[hdr.index(model)])
22
+ except Exception: return None
23
+ return None
24
+
25
+ # (scale, prompt) -> dir
26
+ SRC={(4,1024):"eshort_cache16k_len1024",(8,1024):"eshort_cache32k_len1024",
27
+ (4,4096):"eshort_cache16k_len4096",(8,4096):"eshort_cache32k_len4096",
28
+ (16,4096):"e64k_len4096",(32,4096):"e128k_len4096",
29
+ (8,8192):"e32k_len8192",(8,16360):"e32k_len16360",(8,32744):"e32k_len32744",
30
+ (16,16360):"e64k_len16360",(16,32744):"e64k_len32744",(16,65512):"e64k_len65512",
31
+ (32,32744):"e128k_len32744",(32,65512):"e128k_len65512"}
32
+ PRIOR={(4,8192):84.0,(4,16360):40.0} # measured on the prior box, same haystack
33
+ base={}
34
+ for (sc,pl),d in SRC.items():
35
+ v=read(f"{d}/summary.csv","base680")
36
+ if v is not None: base[(sc,pl)]=v
37
+ base.update(PRIOR)
38
+
39
+ fig,(axA,axB,axC)=plt.subplots(1,3,figsize=(19.5,5.6),facecolor=S["surface"])
40
+ PL=[1024,4096,8192,16360,32744,65512]
41
+ cols=plt.cm.viridis([i/(len(PL)-1) for i in range(len(PL))])
42
+ for pl,c in zip(PL,cols):
43
+ pts=sorted((sc,v) for (sc,p),v in base.items() if p==pl)
44
+ if len(pts)<2: continue
45
+ axA.plot([p[0] for p in pts],[p[1] for p in pts],"o-",color=c,lw=2,ms=6,
46
+ label=f"{pl:,} tok")
47
+ style(axA); axA.set_xscale("log",base=2); axA.set_xticks([4,8,16,32])
48
+ axA.set_xticklabels(["x4","x8","x16","x32"]); axA.set_ylim(0,100)
49
+ axA.set_xlabel("YARN_SCALE (ORIG_MAXPOS held at 4096)",color=S["ink2"],fontsize=10)
50
+ axA.set_ylabel("niah_single accuracy (%)",color=S["ink2"],fontsize=10)
51
+ axA.set_title("Same untrained weights. Only the RoPE scale changes.",
52
+ color=S["ink"],fontsize=12,loc="left",pad=10)
53
+ axA.legend(frameon=False,fontsize=8.5,labelcolor=S["ink2"],loc="lower right")
54
+
55
+ for pl,c in zip(PL,cols):
56
+ pts=sorted((pl/sc,v) for (sc,p),v in base.items() if p==pl)
57
+ if not pts: continue
58
+ axB.plot([p[0] for p in pts],[p[1] for p in pts],"o-",color=c,lw=2,ms=6,label=f"{pl:,} tok")
59
+ style(axB); axB.set_xscale("log",base=2); axB.set_ylim(0,100)
60
+ axB.set_xlabel("virtual position of last token (real / scale)",color=S["ink2"],fontsize=10)
61
+ axB.set_ylabel("niah_single accuracy (%)",color=S["ink2"],fontsize=10)
62
+ axB.axvspan(2000,4200,color="#f3eceb",zorder=0)
63
+ axB.text(2900,6,"pretrain p90\n~2,721 tok",ha="center",fontsize=7.6,color=S["ink2"],style="italic")
64
+ axB.set_title("Collapses onto virtual position; spread = distractor load",
65
+ color=S["ink"],fontsize=12,loc="left",pad=10)
66
+ axB.legend(frameon=False,fontsize=8.5,labelcolor=S["ink2"],loc="lower left")
67
+ fig.text(.005,-.02,"680M base680, ZERO long-context training. niah_single N=100, greedy, "
68
+ "identical haystack (199,754,694 B) across every point. Wilson 95% CI is +/-9-13pp at N=100.",
69
+ fontsize=7.8,color=S["ink2"])
70
+
71
+ # ---- panel C: the 1.06B control -- config helps, but training still wins at 64k
72
+ PLC=[16360,32744,65512]
73
+ own=[70,50,31]; corr=[78,73,65]; cliff_own=[54,31,18]; cliff_corr=[66,70,65]
74
+ m680=[92,79,46]
75
+ x=range(len(PLC))
76
+ axC.plot(x,own,"o--",color="#b0b0aa",lw=1.6,ms=6,label="1.06B, its own config")
77
+ axC.plot(x,corr,"o-",color="#b3243b",lw=2.4,ms=7,label="1.06B, corrected config")
78
+ axC.plot(x,cliff_corr,"s--",color="#eb6834",lw=1.6,ms=6,label='1.06B "collapsed", corrected')
79
+ axC.plot(x,m680,"^-",color="#12508f",lw=2.4,ms=8,label="680M base, no long-ctx training")
80
+ style(axC); axC.set_xticks(list(x)); axC.set_xticklabels(["16k","32k","64k"])
81
+ axC.set_ylim(0,100); axC.set_xlabel("prompt length",color=S["ink2"],fontsize=10)
82
+ axC.set_ylabel("niah_single accuracy (%)",color=S["ink2"],fontsize=10)
83
+ axC.set_title("Control: training wins where it trained (64k)",color=S["ink"],fontsize=12,loc="left",pad=10)
84
+ axC.annotate("1.06B trained to 64k\nbeats the 680M here\nby 19pp",xy=(2,65),xytext=(1.05,30),
85
+ fontsize=8.2,color=S["ink2"],arrowprops=dict(arrowstyle="->",color=S["ink2"],lw=1))
86
+ axC.legend(frameon=False,fontsize=8,labelcolor=S["ink2"],loc="upper right")
87
+ fig.tight_layout(); fig.savefig("plots_p1/scale_sweep.png",dpi=155,bbox_inches="tight")
88
+ print(" wrote plots_p1/scale_sweep.png")
89
+ print(f" points plotted: {len(base)}")