Eriksanchezolvera commited on
Commit
7420c83
·
verified ·
1 Parent(s): 86415de

Revision 5: negative-result model card, report rev. 5 (PDF/MD/figure), CITATION.cff, iterations 020-021 code, logs, sealed tests and checkpoints

Browse files
020/cuerdas020-sealed.log ADDED
@@ -0,0 +1,114 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [{"stream_name":"stdout","time":3.388341406,"data":" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 44.1/44.1 kB 1.6 MB/s eta 0:00:00\n"}
2
+ ,{"stream_name":"stdout","time":4.420337612,"data":" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 10.0/10.0 MB 79.0 MB/s eta 0:00:00\n"}
3
+ ,{"stream_name":"stdout","time":4.454446033,"data":" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 566.4/566.4 kB 28.4 MB/s eta 0:00:00\n"}
4
+ ,{"stream_name":"stdout","time":4.530811111,"data":" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 3.0/3.0 MB 48.9 MB/s eta 0:00:00\n"}
5
+ ,{"stream_name":"stderr","time":17.960055603,"data":"ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.\n"}
6
+ ,{"stream_name":"stderr","time":17.960079725,"data":"litellm 1.85.7 requires tokenizers\u003c1.0,\u003e=0.21.0, but you have tokenizers 0.20.3 which is incompatible.\n"}
7
+ ,{"stream_name":"stderr","time":17.960084532,"data":"gradio 6.26.0 requires huggingface-hub\u003c2.0,\u003e=1.16.0, but you have huggingface-hub 0.36.2 which is incompatible.\n"}
8
+ ,{"stream_name":"stderr","time":17.960088477,"data":"diffusers 0.40.0 requires huggingface-hub\u003c2.0,\u003e=1.23.0, but you have huggingface-hub 0.36.2 which is incompatible.\n"}
9
+ ,{"stream_name":"stdout","time":65.837648993,"data":"{\"env\": {\"python\": \"3.13.15\", \"torch\": \"2.11.0+cu128\", \"transformers\": \"4.46.3\", \"device\": \"Tesla T4\"}, \"sealed_sha256\": \"4c7737f66364a44786b4d9acb4686a6f31829d78cb5914f19eca1df74fe73393\"}\n"}
10
+ ,{"stream_name":"stdout","time":68.610543429,"data":"{\"route\": \"text_template\", \"set\": \"sealed_id\", \"n\": 200, \"acc\": 0.995, \"wilson95\": [97.2, 99.9], \"positions_per_event\": 35.0}\n"}
11
+ ,{"stream_name":"stdout","time":72.844882846,"data":"{\"route\": \"text_template\", \"set\": \"sealed_ood\", \"n\": 400, \"acc\": 0.9925, \"wilson95\": [97.8, 99.7], \"positions_per_event\": 35.0}\n"}
12
+ ,{"stream_name":"stdout","time":74.885158342,"data":"{\"route\": \"text_template\", \"set\": \"sealed_1_new\", \"n\": 192, \"acc\": 0.9635416666666666, \"wilson95\": [92.7, 98.2], \"positions_per_event\": 35.0}\n"}
13
+ ,{"stream_name":"stdout","time":76.948103187,"data":"{\"route\": \"text_template\", \"set\": \"sealed_2_new\", \"n\": 192, \"acc\": 0.984375, \"wilson95\": [95.5, 99.5], \"positions_per_event\": 35.0}\n"}
14
+ ,{"stream_name":"stdout","time":79.000557751,"data":"{\"route\": \"text_template\", \"set\": \"sealed_4_new\", \"n\": 192, \"acc\": 0.9427083333333334, \"wilson95\": [90.0, 96.8], \"positions_per_event\": 35.0}\n"}
15
+ ,{"stream_name":"stdout","time":81.022912645,"data":"{\"route\": \"text_kv\", \"set\": \"sealed_id\", \"n\": 200, \"acc\": 0.065, \"wilson95\": [3.8, 10.8], \"positions_per_event\": 31.0, \"diff_vs_template\": -93.0, \"boot95\": [-96.5, -89.0]}\n"}
16
+ ,{"stream_name":"stdout","time":85.009245483,"data":"{\"route\": \"text_kv\", \"set\": \"sealed_ood\", \"n\": 400, \"acc\": 0.025, \"wilson95\": [1.4, 4.5], \"positions_per_event\": 31.0, \"diff_vs_template\": -96.75, \"boot95\": [-98.25, -95.0]}\n"}
17
+ ,{"stream_name":"stdout","time":86.926660201,"data":"{\"route\": \"text_kv\", \"set\": \"sealed_1_new\", \"n\": 192, \"acc\": 0.010416666666666666, \"wilson95\": [0.3, 3.7], \"positions_per_event\": 31.0, \"diff_vs_template\": -95.31, \"boot95\": [-98.44, -91.67]}\n"}
18
+ ,{"stream_name":"stdout","time":88.836022647,"data":"{\"route\": \"text_kv\", \"set\": \"sealed_2_new\", \"n\": 192, \"acc\": 0.052083333333333336, \"wilson95\": [2.9, 9.3], \"positions_per_event\": 31.0, \"diff_vs_template\": -93.23, \"boot95\": [-96.35, -89.58]}\n"}
19
+ ,{"stream_name":"stdout","time":90.758205317,"data":"{\"route\": \"text_kv\", \"set\": \"sealed_4_new\", \"n\": 192, \"acc\": 0.08333333333333333, \"wilson95\": [5.2, 13.1], \"positions_per_event\": 31.0, \"diff_vs_template\": -85.94, \"boot95\": [-90.62, -80.73]}\n"}
20
+ ,{"stream_name":"stdout","time":92.80039212,"data":"{\"route\": \"text_abbrev\", \"set\": \"sealed_id\", \"n\": 200, \"acc\": 0.035, \"wilson95\": [1.7, 7.0], \"positions_per_event\": 27.67, \"diff_vs_template\": -96.0, \"boot95\": [-98.5, -93.0]}\n"}
21
+ ,{"stream_name":"stdout","time":96.629012409,"data":"{\"route\": \"text_abbrev\", \"set\": \"sealed_ood\", \"n\": 400, \"acc\": 0.0, \"wilson95\": [0.0, 1.0], \"positions_per_event\": 27.63, \"diff_vs_template\": -99.25, \"boot95\": [-100.0, -98.25]}\n"}
22
+ ,{"stream_name":"stdout","time":98.586372404,"data":"{\"route\": \"text_abbrev\", \"set\": \"sealed_1_new\", \"n\": 192, \"acc\": 0.0, \"wilson95\": [0.0, 2.0], \"positions_per_event\": 29.1, \"diff_vs_template\": -96.35, \"boot95\": [-98.96, -93.23]}\n"}
23
+ ,{"stream_name":"stdout","time":100.684886105,"data":"{\"route\": \"text_abbrev\", \"set\": \"sealed_2_new\", \"n\": 192, \"acc\": 0.0, \"wilson95\": [0.0, 2.0], \"positions_per_event\": 30.47, \"diff_vs_template\": -98.44, \"boot95\": [-100.0, -96.35]}\n"}
24
+ ,{"stream_name":"stdout","time":102.909738537,"data":"{\"route\": \"text_abbrev\", \"set\": \"sealed_4_new\", \"n\": 192, \"acc\": 0.0, \"wilson95\": [0.0, 2.0], \"positions_per_event\": 33.37, \"diff_vs_template\": -94.27, \"boot95\": [-97.4, -90.62]}\n"}
25
+ ,{"stream_name":"stdout","time":105.034201776,"data":"{\"route\": \"text_csv\", \"set\": \"sealed_id\", \"n\": 200, \"acc\": 0.03, \"wilson95\": [1.4, 6.4], \"positions_per_event\": 30.67, \"diff_vs_template\": -96.5, \"boot95\": [-99.0, -93.5]}\n"}
26
+ ,{"stream_name":"stdout","time":109.355365425,"data":"{\"route\": \"text_csv\", \"set\": \"sealed_ood\", \"n\": 400, \"acc\": 0.005, \"wilson95\": [0.1, 1.8], \"positions_per_event\": 30.63, \"diff_vs_template\": -98.75, \"boot95\": [-99.75, -97.5]}\n"}
27
+ ,{"stream_name":"stdout","time":111.556024189,"data":"{\"route\": \"text_csv\", \"set\": \"sealed_1_new\", \"n\": 192, \"acc\": 0.005208333333333333, \"wilson95\": [0.1, 2.9], \"positions_per_event\": 32.1, \"diff_vs_template\": -95.83, \"boot95\": [-98.44, -92.71]}\n"}
28
+ ,{"stream_name":"stdout","time":113.84502117,"data":"{\"route\": \"text_csv\", \"set\": \"sealed_2_new\", \"n\": 192, \"acc\": 0.005208333333333333, \"wilson95\": [0.1, 2.9], \"positions_per_event\": 33.47, \"diff_vs_template\": -97.92, \"boot95\": [-99.48, -95.83]}\n"}
29
+ ,{"stream_name":"stdout","time":116.331066973,"data":"{\"route\": \"text_csv\", \"set\": \"sealed_4_new\", \"n\": 192, \"acc\": 0.0, \"wilson95\": [0.0, 2.0], \"positions_per_event\": 36.37, \"diff_vs_template\": -94.27, \"boot95\": [-97.4, -90.62]}\n"}
30
+ ,{"stream_name":"stdout","time":117.764029661,"data":"{\"route\": \"text_values_csv\", \"set\": \"sealed_id\", \"n\": 200, \"acc\": 0.03, \"wilson95\": [1.4, 6.4], \"positions_per_event\": 22.0, \"diff_vs_template\": -96.5, \"boot95\": [-99.0, -93.5]}\n"}
31
+ ,{"stream_name":"stdout","time":120.623199264,"data":"{\"route\": \"text_values_csv\", \"set\": \"sealed_ood\", \"n\": 400, \"acc\": 0.0025, \"wilson95\": [0.0, 1.4], \"positions_per_event\": 22.0, \"diff_vs_template\": -99.0, \"boot95\": [-99.75, -98.0]}\n"}
32
+ ,{"stream_name":"stdout","time":122.201757359,"data":"{\"route\": \"text_values_csv\", \"set\": \"sealed_1_new\", \"n\": 192, \"acc\": 0.0, \"wilson95\": [0.0, 2.0], \"positions_per_event\": 22.46, \"diff_vs_template\": -96.35, \"boot95\": [-98.96, -93.23]}\n"}
33
+ ,{"stream_name":"stdout","time":123.764041327,"data":"{\"route\": \"text_values_csv\", \"set\": \"sealed_2_new\", \"n\": 192, \"acc\": 0.0, \"wilson95\": [0.0, 2.0], \"positions_per_event\": 22.76, \"diff_vs_template\": -98.44, \"boot95\": [-100.0, -96.35]}\n"}
34
+ ,{"stream_name":"stdout","time":125.356988219,"data":"{\"route\": \"text_values_csv\", \"set\": \"sealed_4_new\", \"n\": 192, \"acc\": 0.0, \"wilson95\": [0.0, 2.0], \"positions_per_event\": 23.56, \"diff_vs_template\": -94.27, \"boot95\": [-97.4, -90.62]}\n"}
35
+ ,{"stream_name":"stdout","time":126.613655256,"data":"{\"route\": \"text_values\", \"set\": \"sealed_id\", \"n\": 200, \"acc\": 0.005, \"wilson95\": [0.1, 2.8], \"positions_per_event\": 19.0, \"diff_vs_template\": -99.0, \"boot95\": [-100.0, -97.5]}\n"}
36
+ ,{"stream_name":"stdout","time":129.126717252,"data":"{\"route\": \"text_values\", \"set\": \"sealed_ood\", \"n\": 400, \"acc\": 0.0025, \"wilson95\": [0.0, 1.4], \"positions_per_event\": 19.0, \"diff_vs_template\": -99.0, \"boot95\": [-99.75, -98.0]}\n"}
37
+ ,{"stream_name":"stdout","time":130.498097875,"data":"{\"route\": \"text_values\", \"set\": \"sealed_1_new\", \"n\": 192, \"acc\": 0.0, \"wilson95\": [0.0, 2.0], \"positions_per_event\": 19.46, \"diff_vs_template\": -96.35, \"boot95\": [-98.96, -93.23]}\n"}
38
+ ,{"stream_name":"stdout","time":131.90200467,"data":"{\"route\": \"text_values\", \"set\": \"sealed_2_new\", \"n\": 192, \"acc\": 0.0, \"wilson95\": [0.0, 2.0], \"positions_per_event\": 19.76, \"diff_vs_template\": -98.44, \"boot95\": [-100.0, -96.35]}\n"}
39
+ ,{"stream_name":"stdout","time":133.250639158,"data":"{\"route\": \"text_values\", \"set\": \"sealed_4_new\", \"n\": 192, \"acc\": 0.0, \"wilson95\": [0.0, 2.0], \"positions_per_event\": 20.56, \"diff_vs_template\": -94.27, \"boot95\": [-97.4, -90.62]}\n"}
40
+ ,{"stream_name":"stdout","time":134.808868904,"data":"{\"route\": \"k2_s17\", \"set\": \"sealed_id\", \"n\": 200, \"acc\": 0.985, \"wilson95\": [95.7, 99.5], \"positions_per_event\": 17, \"diff_vs_template\": -1.0, \"boot95\": [-3.0, 1.0]}\n"}
41
+ ,{"stream_name":"stdout","time":137.377613413,"data":"{\"route\": \"k2_s17\", \"set\": \"sealed_ood\", \"n\": 400, \"acc\": 0.99, \"wilson95\": [97.5, 99.6], \"positions_per_event\": 17, \"diff_vs_template\": -0.25, \"boot95\": [-1.5, 1.0]}\n"}
42
+ ,{"stream_name":"stdout","time":138.581706175,"data":"{\"route\": \"k2_s17\", \"set\": \"sealed_1_new\", \"n\": 192, \"acc\": 0.9583333333333334, \"wilson95\": [92.0, 97.9], \"positions_per_event\": 17, \"diff_vs_template\": -0.52, \"boot95\": [-4.17, 3.12]}\n"}
43
+ ,{"stream_name":"stdout","time":139.806274618,"data":"{\"route\": \"k2_s17\", \"set\": \"sealed_2_new\", \"n\": 192, \"acc\": 0.90625, \"wilson95\": [85.7, 94.0], \"positions_per_event\": 17, \"diff_vs_template\": -7.81, \"boot95\": [-12.5, -3.12]}\n"}
44
+ ,{"stream_name":"stdout","time":141.049152533,"data":"{\"route\": \"k2_s17\", \"set\": \"sealed_4_new\", \"n\": 192, \"acc\": 0.828125, \"wilson95\": [76.8, 87.5], \"positions_per_event\": 17, \"diff_vs_template\": -11.46, \"boot95\": [-17.19, -5.73]}\n"}
45
+ ,{"stream_name":"stdout","time":142.511667421,"data":"{\"route\": \"k2_s29\", \"set\": \"sealed_id\", \"n\": 200, \"acc\": 1.0, \"wilson95\": [98.1, 100.0], \"positions_per_event\": 17, \"diff_vs_template\": 0.5, \"boot95\": [0.0, 1.5]}\n"}
46
+ ,{"stream_name":"stdout","time":145.062978489,"data":"{\"route\": \"k2_s29\", \"set\": \"sealed_ood\", \"n\": 400, \"acc\": 0.9975, \"wilson95\": [98.6, 100.0], \"positions_per_event\": 17, \"diff_vs_template\": 0.5, \"boot95\": [-0.5, 1.5]}\n"}
47
+ ,{"stream_name":"stdout","time":146.304692095,"data":"{\"route\": \"k2_s29\", \"set\": \"sealed_1_new\", \"n\": 192, \"acc\": 0.984375, \"wilson95\": [95.5, 99.5], \"positions_per_event\": 17, \"diff_vs_template\": 2.08, \"boot95\": [-1.04, 5.21]}\n"}
48
+ ,{"stream_name":"stdout","time":147.545631656,"data":"{\"route\": \"k2_s29\", \"set\": \"sealed_2_new\", \"n\": 192, \"acc\": 0.9791666666666666, \"wilson95\": [94.8, 99.2], \"positions_per_event\": 17, \"diff_vs_template\": -0.52, \"boot95\": [-3.12, 2.08]}\n"}
49
+ ,{"stream_name":"stdout","time":148.784008133,"data":"{\"route\": \"k2_s29\", \"set\": \"sealed_4_new\", \"n\": 192, \"acc\": 0.875, \"wilson95\": [82.1, 91.5], \"positions_per_event\": 17, \"diff_vs_template\": -6.77, \"boot95\": [-11.98, -2.08]}\n"}
50
+ ,{"stream_name":"stdout","time":150.214376809,"data":"{\"route\": \"k2_s43\", \"set\": \"sealed_id\", \"n\": 200, \"acc\": 0.98, \"wilson95\": [95.0, 99.2], \"positions_per_event\": 17, \"diff_vs_template\": -1.5, \"boot95\": [-4.0, 0.5]}\n"}
51
+ ,{"stream_name":"stdout","time":152.814730981,"data":"{\"route\": \"k2_s43\", \"set\": \"sealed_ood\", \"n\": 400, \"acc\": 0.9825, \"wilson95\": [96.4, 99.1], \"positions_per_event\": 17, \"diff_vs_template\": -1.0, \"boot95\": [-2.5, 0.5]}\n"}
52
+ ,{"stream_name":"stdout","time":154.044099749,"data":"{\"route\": \"k2_s43\", \"set\": \"sealed_1_new\", \"n\": 192, \"acc\": 0.859375, \"wilson95\": [80.3, 90.2], \"positions_per_event\": 17, \"diff_vs_template\": -10.42, \"boot95\": [-15.62, -5.21]}\n"}
53
+ ,{"stream_name":"stdout","time":155.266531092,"data":"{\"route\": \"k2_s43\", \"set\": \"sealed_2_new\", \"n\": 192, \"acc\": 0.8697916666666666, \"wilson95\": [81.5, 91.0], \"positions_per_event\": 17, \"diff_vs_template\": -11.46, \"boot95\": [-16.67, -6.77]}\n"}
54
+ ,{"stream_name":"stdout","time":156.500330949,"data":"{\"route\": \"k2_s43\", \"set\": \"sealed_4_new\", \"n\": 192, \"acc\": 0.8229166666666666, \"wilson95\": [76.3, 87.0], \"positions_per_event\": 17, \"diff_vs_template\": -11.98, \"boot95\": [-18.23, -6.25]}\n"}
55
+ ,{"stream_name":"stdout","time":156.539517256,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_id\", \"acc\": 0.9883333333333333, \"vs\": \"text_template\", \"diff\": -0.67, \"boot95\": [-2.0, 0.83]}\n"}
56
+ ,{"stream_name":"stdout","time":156.579250879,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_id\", \"acc\": 0.9883333333333333, \"vs\": \"text_kv\", \"diff\": 92.33, \"boot95\": [88.67, 95.67]}\n"}
57
+ ,{"stream_name":"stdout","time":156.616884816,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_id\", \"acc\": 0.9883333333333333, \"vs\": \"text_abbrev\", \"diff\": 95.33, \"boot95\": [92.5, 97.83]}\n"}
58
+ ,{"stream_name":"stdout","time":156.65477194,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_id\", \"acc\": 0.9883333333333333, \"vs\": \"text_csv\", \"diff\": 95.83, \"boot95\": [93.17, 98.17]}\n"}
59
+ ,{"stream_name":"stdout","time":156.693577489,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_id\", \"acc\": 0.9883333333333333, \"vs\": \"text_values_csv\", \"diff\": 95.83, \"boot95\": [93.17, 98.17]}\n"}
60
+ ,{"stream_name":"stdout","time":156.732202027,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_id\", \"acc\": 0.9883333333333333, \"vs\": \"text_values\", \"diff\": 98.33, \"boot95\": [96.83, 99.5]}\n"}
61
+ ,{"stream_name":"stdout","time":156.826159533,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_ood\", \"acc\": 0.99, \"vs\": \"text_template\", \"diff\": -0.25, \"boot95\": [-1.25, 0.92]}\n"}
62
+ ,{"stream_name":"stdout","time":156.919513567,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_ood\", \"acc\": 0.99, \"vs\": \"text_kv\", \"diff\": 96.5, \"boot95\": [94.75, 98.0]}\n"}
63
+ ,{"stream_name":"stdout","time":157.011904627,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_ood\", \"acc\": 0.99, \"vs\": \"text_abbrev\", \"diff\": 99.0, \"boot95\": [98.33, 99.58]}\n"}
64
+ ,{"stream_name":"stdout","time":157.103279626,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_ood\", \"acc\": 0.99, \"vs\": \"text_csv\", \"diff\": 98.5, \"boot95\": [97.5, 99.33]}\n"}
65
+ ,{"stream_name":"stdout","time":157.195875214,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_ood\", \"acc\": 0.99, \"vs\": \"text_values_csv\", \"diff\": 98.75, \"boot95\": [97.83, 99.5]}\n"}
66
+ ,{"stream_name":"stdout","time":157.288146417,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_ood\", \"acc\": 0.99, \"vs\": \"text_values\", \"diff\": 98.75, \"boot95\": [97.83, 99.5]}\n"}
67
+ ,{"stream_name":"stdout","time":157.329154262,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_1_new\", \"acc\": 0.9340277777777778, \"vs\": \"text_template\", \"diff\": -2.95, \"boot95\": [-6.08, 0.35]}\n"}
68
+ ,{"stream_name":"stdout","time":157.365357174,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_1_new\", \"acc\": 0.9340277777777778, \"vs\": \"text_kv\", \"diff\": 92.36, \"boot95\": [89.41, 94.97]}\n"}
69
+ ,{"stream_name":"stdout","time":157.402069447,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_1_new\", \"acc\": 0.9340277777777778, \"vs\": \"text_abbrev\", \"diff\": 93.4, \"boot95\": [90.97, 95.49]}\n"}
70
+ ,{"stream_name":"stdout","time":157.438547212,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_1_new\", \"acc\": 0.9340277777777778, \"vs\": \"text_csv\", \"diff\": 92.88, \"boot95\": [90.28, 95.14]}\n"}
71
+ ,{"stream_name":"stdout","time":157.476708162,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_1_new\", \"acc\": 0.9340277777777778, \"vs\": \"text_values_csv\", \"diff\": 93.4, \"boot95\": [90.97, 95.49]}\n"}
72
+ ,{"stream_name":"stdout","time":157.513927911,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_1_new\", \"acc\": 0.9340277777777778, \"vs\": \"text_values\", \"diff\": 93.4, \"boot95\": [90.97, 95.49]}\n"}
73
+ ,{"stream_name":"stdout","time":157.551306283,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_2_new\", \"acc\": 0.9184027777777778, \"vs\": \"text_template\", \"diff\": -6.6, \"boot95\": [-9.72, -3.47]}\n"}
74
+ ,{"stream_name":"stdout","time":157.587669908,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_2_new\", \"acc\": 0.9184027777777778, \"vs\": \"text_kv\", \"diff\": 86.63, \"boot95\": [82.47, 90.62]}\n"}
75
+ ,{"stream_name":"stdout","time":157.625047843,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_2_new\", \"acc\": 0.9184027777777778, \"vs\": \"text_abbrev\", \"diff\": 91.84, \"boot95\": [89.06, 94.44]}\n"}
76
+ ,{"stream_name":"stdout","time":157.66255768,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_2_new\", \"acc\": 0.9184027777777778, \"vs\": \"text_csv\", \"diff\": 91.32, \"boot95\": [88.37, 94.1]}\n"}
77
+ ,{"stream_name":"stdout","time":157.700952112,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_2_new\", \"acc\": 0.9184027777777778, \"vs\": \"text_values_csv\", \"diff\": 91.84, \"boot95\": [89.06, 94.44]}\n"}
78
+ ,{"stream_name":"stdout","time":157.739386102,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_2_new\", \"acc\": 0.9184027777777778, \"vs\": \"text_values\", \"diff\": 91.84, \"boot95\": [89.06, 94.44]}\n"}
79
+ ,{"stream_name":"stdout","time":157.776225545,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_4_new\", \"acc\": 0.8420138888888888, \"vs\": \"text_template\", \"diff\": -10.07, \"boot95\": [-15.1, -5.38]}\n"}
80
+ ,{"stream_name":"stdout","time":157.813558255,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_4_new\", \"acc\": 0.8420138888888888, \"vs\": \"text_kv\", \"diff\": 75.87, \"boot95\": [70.14, 81.25]}\n"}
81
+ ,{"stream_name":"stdout","time":157.850692098,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_4_new\", \"acc\": 0.8420138888888888, \"vs\": \"text_abbrev\", \"diff\": 84.2, \"boot95\": [79.51, 88.37]}\n"}
82
+ ,{"stream_name":"stdout","time":157.890013899,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_4_new\", \"acc\": 0.8420138888888888, \"vs\": \"text_csv\", \"diff\": 84.2, \"boot95\": [79.51, 88.37]}\n"}
83
+ ,{"stream_name":"stdout","time":157.926849507,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_4_new\", \"acc\": 0.8420138888888888, \"vs\": \"text_values_csv\", \"diff\": 84.2, \"boot95\": [79.51, 88.37]}\n"}
84
+ ,{"stream_name":"stdout","time":157.965024562,"data":"{\"route\": \"k2_mean3\", \"set\": \"sealed_4_new\", \"acc\": 0.8420138888888888, \"vs\": \"text_values\", \"diff\": 84.2, \"boot95\": [79.51, 88.37]}\n"}
85
+ ,{"stream_name":"stdout","time":223.255143598,"data":"{\"route\": \"probes_k2_s17\", \"ridge_aug:sealed_ood\": 1.0, \"mlp_aug:sealed_ood\": 1.0, \"mlp_ctrl_shuffled:sealed_ood\": 0.0, \"mlp_cls_fit:sealed_ood\": 0.995, \"knn1:sealed_ood\": 0.0, \"knn5:sealed_ood\": 0.1325, \"ridge_aug:sealed_4_new\": 0.5833333333333334, \"mlp_aug:sealed_4_new\": 0.546875, \"mlp_ctrl_shuffled:sealed_4_new\": 0.0, \"mlp_cls_fit:sealed_4_new\": 0.0, \"knn1:sealed_4_new\": 0.0, \"knn5:sealed_4_new\": 0.0, \"ridge_aug:sealed_1_new\": 0.9322916666666666, \"mlp_aug:sealed_1_new\": 1.0, \"mlp_ctrl_shuffled:sealed_1_new\": 0.0, \"mlp_cls_fit:sealed_1_new\": 0.0, \"knn1:sealed_1_new\": 0.0, \"knn5:sealed_1_new\": 0.0}\n"}
86
+ ,{"stream_name":"stdout","time":276.767804206,"data":"{\"route\": \"probes_k2_s29\", \"ridge_aug:sealed_ood\": 1.0, \"mlp_aug:sealed_ood\": 1.0, \"mlp_ctrl_shuffled:sealed_ood\": 0.0, \"mlp_cls_fit:sealed_ood\": 0.9975, \"knn1:sealed_ood\": 0.0, \"knn5:sealed_ood\": 0.14, \"ridge_aug:sealed_4_new\": 0.6822916666666666, \"mlp_aug:sealed_4_new\": 0.6458333333333334, \"mlp_ctrl_shuffled:sealed_4_new\": 0.0, \"mlp_cls_fit:sealed_4_new\": 0.0, \"knn1:sealed_4_new\": 0.0, \"knn5:sealed_4_new\": 0.0, \"ridge_aug:sealed_1_new\": 0.9270833333333334, \"mlp_aug:sealed_1_new\": 1.0, \"mlp_ctrl_shuffled:sealed_1_new\": 0.0, \"mlp_cls_fit:sealed_1_new\": 0.0, \"knn1:sealed_1_new\": 0.0, \"knn5:sealed_1_new\": 0.0}\n"}
87
+ ,{"stream_name":"stdout","time":328.503478024,"data":"{\"route\": \"probes_k2_s43\", \"ridge_aug:sealed_ood\": 1.0, \"mlp_aug:sealed_ood\": 1.0, \"mlp_ctrl_shuffled:sealed_ood\": 0.0, \"mlp_cls_fit:sealed_ood\": 1.0, \"knn1:sealed_ood\": 0.0, \"knn5:sealed_ood\": 0.175, \"ridge_aug:sealed_4_new\": 0.671875, \"mlp_aug:sealed_4_new\": 0.5, \"mlp_ctrl_shuffled:sealed_4_new\": 0.0, \"mlp_cls_fit:sealed_4_new\": 0.0, \"knn1:sealed_4_new\": 0.0, \"knn5:sealed_4_new\": 0.0, \"ridge_aug:sealed_1_new\": 0.96875, \"mlp_aug:sealed_1_new\": 1.0, \"mlp_ctrl_shuffled:sealed_1_new\": 0.0, \"mlp_cls_fit:sealed_1_new\": 0.0, \"knn1:sealed_1_new\": 0.0, \"knn5:sealed_1_new\": 0.0}\n"}
88
+ ,{"stream_name":"stdout","time":328.621673763,"data":"{\"probe_mean3\": \"ridge_aug\", \"set\": \"sealed_ood\", \"acc\": 1.0, \"llm_k2_acc\": 0.99, \"diff_vs_llm\": 1.0, \"boot95\": [0.42, 1.67]}\n"}
89
+ ,{"stream_name":"stdout","time":328.726904848,"data":"{\"probe_mean3\": \"mlp_aug\", \"set\": \"sealed_ood\", \"acc\": 1.0, \"llm_k2_acc\": 0.99, \"diff_vs_llm\": 1.0, \"boot95\": [0.42, 1.67]}\n"}
90
+ ,{"stream_name":"stdout","time":328.82347804,"data":"{\"probe_mean3\": \"mlp_ctrl_shuffled\", \"set\": \"sealed_ood\", \"acc\": 0.0, \"llm_k2_acc\": 0.99, \"diff_vs_llm\": -99.0, \"boot95\": [-99.58, -98.33]}\n"}
91
+ ,{"stream_name":"stdout","time":328.914063195,"data":"{\"probe_mean3\": \"mlp_cls_fit\", \"set\": \"sealed_ood\", \"acc\": 0.9975, \"llm_k2_acc\": 0.99, \"diff_vs_llm\": 0.75, \"boot95\": [0.08, 1.5]}\n"}
92
+ ,{"stream_name":"stdout","time":329.008640798,"data":"{\"probe_mean3\": \"knn1\", \"set\": \"sealed_ood\", \"acc\": 0.0, \"llm_k2_acc\": 0.99, \"diff_vs_llm\": -99.0, \"boot95\": [-99.58, -98.33]}\n"}
93
+ ,{"stream_name":"stdout","time":329.105495539,"data":"{\"probe_mean3\": \"knn5\", \"set\": \"sealed_ood\", \"acc\": 0.14916666666666667, \"llm_k2_acc\": 0.99, \"diff_vs_llm\": -84.08, \"boot95\": [-87.0, -81.08]}\n"}
94
+ ,{"stream_name":"stdout","time":329.145507064,"data":"{\"probe_mean3\": \"ridge_aug\", \"set\": \"sealed_4_new\", \"acc\": 0.6458333333333334, \"llm_k2_acc\": 0.8420138888888888, \"diff_vs_llm\": -19.62, \"boot95\": [-26.74, -12.5]}\n"}
95
+ ,{"stream_name":"stdout","time":329.183671603,"data":"{\"probe_mean3\": \"mlp_aug\", \"set\": \"sealed_4_new\", \"acc\": 0.564236111111111, \"llm_k2_acc\": 0.8420138888888888, \"diff_vs_llm\": -27.78, \"boot95\": [-35.24, -20.14]}\n"}
96
+ ,{"stream_name":"stdout","time":329.222185721,"data":"{\"probe_mean3\": \"mlp_ctrl_shuffled\", \"set\": \"sealed_4_new\", \"acc\": 0.0, \"llm_k2_acc\": 0.8420138888888888, \"diff_vs_llm\": -84.2, \"boot95\": [-88.54, -79.69]}\n"}
97
+ ,{"stream_name":"stdout","time":329.259354621,"data":"{\"probe_mean3\": \"mlp_cls_fit\", \"set\": \"sealed_4_new\", \"acc\": 0.0, \"llm_k2_acc\": 0.8420138888888888, \"diff_vs_llm\": -84.2, \"boot95\": [-88.54, -79.69]}\n"}
98
+ ,{"stream_name":"stdout","time":329.296126324,"data":"{\"probe_mean3\": \"knn1\", \"set\": \"sealed_4_new\", \"acc\": 0.0, \"llm_k2_acc\": 0.8420138888888888, \"diff_vs_llm\": -84.2, \"boot95\": [-88.54, -79.69]}\n"}
99
+ ,{"stream_name":"stdout","time":329.334319863,"data":"{\"probe_mean3\": \"knn5\", \"set\": \"sealed_4_new\", \"acc\": 0.0, \"llm_k2_acc\": 0.8420138888888888, \"diff_vs_llm\": -84.2, \"boot95\": [-88.54, -79.69]}\n"}
100
+ ,{"stream_name":"stdout","time":329.37332952,"data":"{\"probe_mean3\": \"ridge_aug\", \"set\": \"sealed_1_new\", \"acc\": 0.9427083333333334, \"llm_k2_acc\": 0.9340277777777778, \"diff_vs_llm\": 0.87, \"boot95\": [-2.95, 4.51]}\n"}
101
+ ,{"stream_name":"stdout","time":329.411147075,"data":"{\"probe_mean3\": \"mlp_aug\", \"set\": \"sealed_1_new\", \"acc\": 1.0, \"llm_k2_acc\": 0.9340277777777778, \"diff_vs_llm\": 6.6, \"boot95\": [4.51, 8.85]}\n"}
102
+ ,{"stream_name":"stdout","time":329.44810494,"data":"{\"probe_mean3\": \"mlp_ctrl_shuffled\", \"set\": \"sealed_1_new\", \"acc\": 0.0, \"llm_k2_acc\": 0.9340277777777778, \"diff_vs_llm\": -93.4, \"boot95\": [-95.49, -91.15]}\n"}
103
+ ,{"stream_name":"stdout","time":329.485734684,"data":"{\"probe_mean3\": \"mlp_cls_fit\", \"set\": \"sealed_1_new\", \"acc\": 0.0, \"llm_k2_acc\": 0.9340277777777778, \"diff_vs_llm\": -93.4, \"boot95\": [-95.49, -91.15]}\n"}
104
+ ,{"stream_name":"stdout","time":329.523359846,"data":"{\"probe_mean3\": \"knn1\", \"set\": \"sealed_1_new\", \"acc\": 0.0, \"llm_k2_acc\": 0.9340277777777778, \"diff_vs_llm\": -93.4, \"boot95\": [-95.49, -91.15]}\n"}
105
+ ,{"stream_name":"stdout","time":329.561016123,"data":"{\"probe_mean3\": \"knn5\", \"set\": \"sealed_1_new\", \"acc\": 0.0, \"llm_k2_acc\": 0.9340277777777778, \"diff_vs_llm\": -93.4, \"boot95\": [-95.49, -91.15]}\n"}
106
+ ,{"stream_name":"stdout","time":342.977856481,"data":"{\"control\": \"random_receiver\", \"set\": \"sealed_ood\", \"text_template\": 0.0, \"k2_s17\": 0.0, \"k2_s29\": 0.0, \"k2_s43\": 0.0}\n"}
107
+ ,{"stream_name":"stdout","time":349.383162888,"data":"{\"control\": \"random_receiver\", \"set\": \"sealed_4_new\", \"text_template\": 0.0, \"k2_s17\": 0.0, \"k2_s29\": 0.0, \"k2_s43\": 0.0}\n"}
108
+ ,{"stream_name":"stderr","time":357.249388448,"data":"/usr/local/lib/python3.13/dist-packages/mistune.py:435: SyntaxWarning: invalid escape sequence '\\|'\n"}
109
+ ,{"stream_name":"stderr","time":357.249427362,"data":" cells[i][c] = re.sub('\\\\\\\\\\|', '|', cell)\n"}
110
+ ,{"stream_name":"stderr","time":357.476570044,"data":"/usr/local/lib/python3.13/dist-packages/nbconvert/filters/filter_links.py:36: SyntaxWarning: invalid escape sequence '\\_'\n"}
111
+ ,{"stream_name":"stderr","time":357.47660111,"data":" text = re.sub(r'_', '\\_', text) # Escape underscores in display text\n"}
112
+ ,{"stream_name":"stderr","time":358.244424653,"data":"[NbConvertApp] Converting notebook __script__.ipynb to html\n"}
113
+ ,{"stream_name":"stderr","time":358.758158185,"data":"[NbConvertApp] Writing 278746 bytes to __results__.html\n"}
114
+ ]
020/eval020.py ADDED
@@ -0,0 +1,127 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # CUERDAS-020: sealed test (evaluated once), compact text baselines, stronger no-LLM decoders.
2
+ import os, sys, json, hashlib, random, math, platform
3
+ import torch, torch.nn as nn, torch.nn.functional as F, transformers
4
+ sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
5
+ import run018 as R
6
+ from transformers import AutoModelForCausalLM, AutoConfig
7
+ INP = os.environ['INP']; OD = os.environ['OD']; NBOOT = 10000; SEEDS = [17, 29, 43]
8
+ raw = open(f'{INP}/sealed020.json').read()
9
+ H = hashlib.sha256(raw.encode()).hexdigest(); assert H == '4c7737f66364a44786b4d9acb4686a6f31829d78cb5914f19eca1df74fe73393', H
10
+ S = json.loads(raw); NEW = S['new_words']
11
+ SUMM = open(f'{OD}/results020.jsonl', 'w'); PRED = open(f'{OD}/predictions020.jsonl', 'w')
12
+ def log(d): SUMM.write(json.dumps(d) + '\n'); SUMM.flush(); print(json.dumps(d), flush=True)
13
+ log({'env': {'python': platform.python_version(), 'torch': torch.__version__, 'transformers': transformers.__version__, 'device': torch.cuda.get_device_name() if R.DEV == 'cuda' else 'cpu'}, 'sealed_sha256': H})
14
+ for j in range(4):
15
+ for w in NEW[j]: assert len(R.t1(w)) == 1, w
16
+ CAND = [torch.tensor(sorted({R.t1(w)[0] for w in R.V[j] + NEW[j]})) for j in range(4)]
17
+ assert all(len(CAND[j]) == len(R.V[j]) + 8 for j in range(4))
18
+ def gidx(j, w): return int((CAND[j] == R.t1(w)[0]).nonzero()[0])
19
+ SETS = ['sealed_id', 'sealed_ood', 'sealed_1_new', 'sealed_2_new', 'sealed_4_new']
20
+ FMT = {'template': R.TPL, 'kv': "action: {action}; target: {target}; status: {status}; window: {window}.",
21
+ 'abbrev': "act={action} tgt={target} st={status} win={window}", 'csv': "action,target,status,window\n{action},{target},{status},{window}",
22
+ 'values_csv': "{action}, {target}, {status}, {window}", 'values': "{action} {target} {status} {window}"}
23
+ def text_logits(msgs, fmt):
24
+ embs, mt = [], []
25
+ for i, m in enumerate(msgs):
26
+ for j in range(4):
27
+ embs.append(R.Emb[R.tok.encode(fmt.format(**dict(zip(R.FIELDS, m))) + R.SUF_MAIN.format(field=R.FIELDS[j]), add_special_tokens=False)]); mt.append((i, j))
28
+ return R.run_rows(embs, None), mt
29
+ @torch.no_grad()
30
+ def per_ex(msgs, fmt=None, cod=None, bs=24):
31
+ out = []
32
+ for b in range(0, len(msgs), bs):
33
+ mb = msgs[b:b+bs]; lg, mt = text_logits(mb, fmt) if cod is None else R.codec_logits(cod, mb, R.SUF_MAIN)
34
+ good = [True] * len(mb)
35
+ for r, (i, j) in enumerate(mt): good[i] &= int(lg[r][CAND[j]].argmax()) == gidx(j, mb[i][j])
36
+ out += good
37
+ return [float(g) for g in out]
38
+ def wilson(k, n, z=1.96):
39
+ p = k / n; d = 1 + z*z/n; c = (p + z*z/(2*n)) / d; h = z * math.sqrt(p*(1-p)/n + z*z/(4*n*n)) / d
40
+ return [round(100*(c-h), 1), round(100*(c+h), 1)]
41
+ def boot(a, b, seed=0):
42
+ g = torch.Generator().manual_seed(seed); a = torch.tensor(a, dtype=torch.float64, device='cpu'); b = torch.tensor(b, dtype=torch.float64, device='cpu')
43
+ idx = torch.randint(0, len(a), (NBOOT, len(a)), generator=g, device='cpu'); d = (a[idx].mean(1) - b[idx].mean(1)) * 100
44
+ return round(float((a.mean()-b.mean())*100), 2), [round(float(d.quantile(.025)), 2), round(float(d.quantile(.975)), 2)]
45
+ def positions(msgs, fmt): return sum(len(R.tok.encode(fmt.format(**dict(zip(R.FIELDS, m))), add_special_tokens=False)) for m in msgs) / len(msgs) + 15
46
+ # ---- text formats ----
47
+ TXT = {}
48
+ for fn, fmt in FMT.items():
49
+ for s in SETS:
50
+ r = per_ex(S[s], fmt); TXT[(fn, s)] = r; k = int(sum(r))
51
+ for i, c in enumerate(r): PRED.write(json.dumps({'route': 'text_' + fn, 'set': s, 'i': i, 'correct': bool(c)}) + '\n')
52
+ e = {'route': 'text_' + fn, 'set': s, 'n': len(r), 'acc': k/len(r), 'wilson95': wilson(k, len(r)), 'positions_per_event': round(positions(S[s], fmt), 2)}
53
+ if fn != 'template': e['diff_vs_template'], e['boot95'] = boot(r, TXT[('template', s)], seed=7)
54
+ log(e)
55
+ # ---- K=2, sealed, evaluated once ----
56
+ cods = {}; K2 = {}
57
+ for sd in SEEDS:
58
+ cod = R.PackK(2, sd); cod.load_state_dict(torch.load(f'{INP}/k2long_s{sd}.pt', map_location=R.DEV)); cod.eval().requires_grad_(False); cods[sd] = cod
59
+ for s in SETS:
60
+ r = per_ex(S[s], cod=cod); K2[(sd, s)] = r; k = int(sum(r))
61
+ for i, c in enumerate(r): PRED.write(json.dumps({'route': f'k2_s{sd}', 'set': s, 'i': i, 'correct': bool(c)}) + '\n')
62
+ d, ci = boot(r, TXT[('template', s)], seed=sd)
63
+ log({'route': f'k2_s{sd}', 'set': s, 'n': len(r), 'acc': k/len(r), 'wilson95': wilson(k, len(r)), 'positions_per_event': 17, 'diff_vs_template': d, 'boot95': ci})
64
+ for s in SETS:
65
+ a = [sum(K2[(sd, s)][i] for sd in SEEDS)/3 for i in range(len(S[s]))]
66
+ for fn in FMT:
67
+ d, ci = boot(a, TXT[(fn, s)], seed=1); log({'route': 'k2_mean3', 'set': s, 'acc': sum(a)/len(a), 'vs': 'text_' + fn, 'diff': d, 'boot95': ci})
68
+ # ---- no-LLM decoders on K=2 slots ----
69
+ def slots(cod, ms):
70
+ with torch.no_grad(): return torch.cat([cod(torch.stack([torch.stack([R.lex(w) for w in m]) for m in ms[i:i+512]])).flatten(1) for i in range(0, len(ms), 512)])
71
+ g = random.Random(2020); AUG = []
72
+ for _ in range(20000):
73
+ m = list(g.choice(R.FIT))
74
+ for j in range(4):
75
+ if g.random() < .4: m[j] = g.choice(R.POOL)
76
+ AUG.append(m)
77
+ YA = torch.stack([torch.stack([R.lex(w) for w in m]) for m in AUG]) # [N,4,D]
78
+ CE = [F.normalize(R.Emb[CAND[j]], dim=-1) for j in range(4)]
79
+ def dec_emb(P, ms): # P [n,4,D] predicted value embeddings -> exact-match
80
+ ok = torch.ones(len(ms), dtype=torch.bool, device='cpu')
81
+ for j in range(4): ok &= ((F.normalize(P[:, j], dim=-1) @ CE[j].T).argmax(1).cpu() == torch.tensor([gidx(j, m[j]) for m in ms], device='cpu'))
82
+ return ok.float().tolist()
83
+ def train_mlp(X, Y, shuffle=False, steps=4000, seed=0):
84
+ torch.manual_seed(seed); Y = Y[torch.randperm(len(Y))] if shuffle else Y
85
+ mu, sdv = X.mean(0), X.std(0) + 1e-6
86
+ net = nn.Sequential(nn.Linear(X.shape[1], 1024), nn.GELU(), nn.Linear(1024, 1024), nn.GELU(), nn.Linear(1024, 4*R.D)); opt = torch.optim.Adam(net.parameters(), 1e-3)
87
+ for _ in range(steps):
88
+ b = torch.randint(0, len(X), (256,)); p = net((X[b]-mu)/sdv).view(-1, 4, R.D)
89
+ loss = (1 - F.cosine_similarity(p, Y[b], -1)).mean(); opt.zero_grad(); loss.backward(); opt.step()
90
+ net.eval(); return lambda Z: net((Z-mu)/sdv).view(-1, 4, R.D).detach()
91
+ PROBE = {}
92
+ for sd in SEEDS:
93
+ cod = cods[sd]; XA = slots(cod, AUG); XF = slots(cod, R.FIT); res = {'route': f'probes_k2_s{sd}'}
94
+ A = torch.linalg.solve(XA.T@XA + 10*torch.eye(XA.shape[1]), XA.T@YA.flatten(1)); ridge = lambda Z: (Z@A).view(-1, 4, R.D)
95
+ mlp = train_mlp(XA, YA, seed=sd); ctrl = train_mlp(XA, YA, shuffle=True, seed=sd)
96
+ YF = torch.tensor([[R.V[j].index(m[j]) for j in range(4)] for m in R.FIT])
97
+ torch.manual_seed(sd); cls = nn.Sequential(nn.Linear(XF.shape[1], 1024), nn.GELU(), nn.Linear(1024, sum(len(v) for v in R.V))); opt = torch.optim.Adam(cls.parameters(), 1e-3)
98
+ off = [0, 12, 24, 40, 50]; muF, sdF = XF.mean(0), XF.std(0)+1e-6
99
+ for _ in range(2000):
100
+ b = torch.randint(0, len(XF), (128,)); o = cls((XF[b]-muF)/sdF); loss = sum(F.cross_entropy(o[:, off[j]:off[j+1]], YF[b, j]) for j in range(4)); opt.zero_grad(); loss.backward(); opt.step()
101
+ XFn = F.normalize(XF, dim=-1)
102
+ for s in ('sealed_ood', 'sealed_4_new', 'sealed_1_new'):
103
+ ms = S[s]; Z = slots(cod, ms)
104
+ for name, fn in (('ridge_aug', ridge), ('mlp_aug', mlp), ('mlp_ctrl_shuffled', ctrl)):
105
+ r = dec_emb(fn(Z), ms); PROBE[(sd, name, s)] = r; res[f'{name}:{s}'] = sum(r)/len(r)
106
+ with torch.no_grad(): o = cls((Z-muF)/sdF)
107
+ okc = []; okk1 = []; okk5 = []; sim = F.normalize(Z, dim=-1) @ XFn.T; top = sim.topk(5, dim=1).indices.cpu()
108
+ for i, m in enumerate(ms):
109
+ okc.append(float(all(m[j] in R.V[j] and int(o[i, off[j]:off[j+1]].argmax()) == R.V[j].index(m[j]) for j in range(4))))
110
+ n1 = R.FIT[int(top[i, 0])]; okk1.append(float(all(n1[j] == m[j] for j in range(4))))
111
+ vote = [max(set(R.FIT[int(t)][j] for t in top[i]), key=lambda v: sum(R.FIT[int(t)][j] == v for t in top[i])) for j in range(4)]
112
+ okk5.append(float(all(vote[j] == m[j] for j in range(4))))
113
+ for name, r in (('mlp_cls_fit', okc), ('knn1', okk1), ('knn5', okk5)): PROBE[(sd, name, s)] = r; res[f'{name}:{s}'] = sum(r)/len(r)
114
+ log(res)
115
+ for s in ('sealed_ood', 'sealed_4_new', 'sealed_1_new'):
116
+ for name in ('ridge_aug', 'mlp_aug', 'mlp_ctrl_shuffled', 'mlp_cls_fit', 'knn1', 'knn5'):
117
+ a = [sum(PROBE[(sd, name, s)][i] for sd in SEEDS)/3 for i in range(len(S[s]))]; l = [sum(K2[(sd, s)][i] for sd in SEEDS)/3 for i in range(len(S[s]))]
118
+ d, ci = boot(a, l, seed=3); log({'probe_mean3': name, 'set': s, 'acc': sum(a)/len(a), 'llm_k2_acc': sum(l)/len(l), 'diff_vs_llm': d, 'boot95': ci})
119
+ # ---- random-weight receiver ----
120
+ torch.manual_seed(0)
121
+ rnd = AutoModelForCausalLM.from_config(AutoConfig.from_pretrained(R.MP)).float().eval().requires_grad_(False).to(R.DEV)
122
+ rnd.model.embed_tokens.weight.data.copy_(R.Emb.data); R.lm = rnd
123
+ for s in ('sealed_ood', 'sealed_4_new'):
124
+ e = {'control': 'random_receiver', 'set': s, 'text_template': sum(per_ex(S[s], R.TPL))/len(S[s])}
125
+ for sd in SEEDS: e[f'k2_s{sd}'] = sum(per_ex(S[s], cod=cods[sd]))/len(S[s])
126
+ log(e)
127
+ SUMM.close(); PRED.close()
020/kaggle_kernel.py ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os,subprocess,sys,glob,shutil
2
+ subprocess.run([sys.executable,'-m','pip','install','-q','transformers==4.46.3'],check=True)
3
+ inp=os.path.dirname(glob.glob('/kaggle/input/**/k2long_s17.pt',recursive=True)[0])
4
+ S='/kaggle/working/src'; os.makedirs(S+'/data',exist_ok=True); shutil.copy(f'{inp}/expanded.json',f'{S}/data/')
5
+ for f in ('run018.py','eval020.py'): shutil.copy(f'{inp}/{f}','/kaggle/working/')
6
+ shutil.copy(f'{inp}/sealed020.json','/kaggle/working/')
7
+ from transformers import AutoTokenizer,AutoModelForCausalLM
8
+ mp='/kaggle/working/m135'; R='12fd25f77366fa6b3b4b768ec3050bf629380bac'
9
+ AutoTokenizer.from_pretrained('HuggingFaceTB/SmolLM2-135M-Instruct',revision=R).save_pretrained(mp); AutoModelForCausalLM.from_pretrained('HuggingFaceTB/SmolLM2-135M-Instruct',revision=R).save_pretrained(mp)
10
+ subprocess.run([sys.executable,'/kaggle/working/eval020.py'],env=dict(os.environ,SRC=S,MP=mp,OD='/kaggle/working',INP=inp),check=True)
11
+ shutil.rmtree(mp); shutil.rmtree(S); os.remove('/kaggle/working/run018.py')
020/make_sealed020.py ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # CUERDAS-020: sealed test generator. Fresh seed, never seen by any earlier iteration.
2
+ import json, random, itertools, re, hashlib, sys
3
+ from transformers import AutoTokenizer
4
+ tok = AutoTokenizer.from_pretrained(sys.argv[1]); SEED = 20261016
5
+ meta = json.load(open(sys.argv[2])); V = meta['vocabulary']; rows = meta['rows']; SPL = meta['splits']
6
+ def t1(w): return tok.encode(' ' + w, add_special_tokens=False)
7
+ used = {tuple(r['values']) for r in rows}
8
+ fit = [tuple(rows[i]['values']) for i in SPL['fit']]; seen_at = {v[:2] for v in fit}; seen_sw = {v[2:] for v in fit}
9
+ # replicate the 015-019 fresh test (seed 15015) so it is excluded
10
+ rng = random.Random(15015); uni = [v for v in itertools.product(*[range(len(x)) for x in V]) if v not in used]; rng.shuffle(uni)
11
+ isid = lambda v: v[:2] in seen_at and v[2:] in seen_sw
12
+ fresh = set([v for v in uni if isid(v)][:200] + [v for v in uni if not isid(v)][:400])
13
+ S013 = (7, 7, 8, 5)
14
+ elig = [v for v in itertools.product(*[range(len(x)) for x in V]) if v not in used and v not in fresh and not all(v[j] < S013[j] for j in range(4))]
15
+ r = random.Random(SEED); r.shuffle(elig)
16
+ sid = [v for v in elig if isid(v)][:200]; sood = [v for v in elig if not isid(v)][:400]
17
+ W = lambda v: [V[j][v[j]] for j in range(4)]
18
+ # unseen words: single-token, >=10 lowercase letters (the 018 training pool only has 3-9 letters, so these can never have been trained on)
19
+ OLD_NEW = 'inspect repair cancel approve archive scan print send sensor battery printer server truck door screen cable broken stable late empty full warm missing urgent night evening weekend noon spring summer winter midnight'.split()
20
+ cand = sorted(t[1:] for t in tok.get_vocab() if t.startswith('Ġ') and re.fullmatch(r'[a-z]{10,14}', t[1:]))
21
+ r.shuffle(cand); usedtok = {t1(w)[0] for vs in V for w in vs}; NEW = [[] for _ in range(4)]
22
+ for w in cand:
23
+ j = min(range(4), key=lambda k: len(NEW[k]))
24
+ if len(NEW[j]) >= 8: break
25
+ if len(t1(w)) == 1 and t1(w)[0] not in usedtok and w not in OLD_NEW: NEW[j].append(w); usedtok.add(t1(w)[0])
26
+ def novel(n_new, n=192):
27
+ out = []
28
+ for k in range(n):
29
+ b = W(sood[k % 400]); js = r.sample(range(4), n_new)
30
+ for j in js: b[j] = r.choice(NEW[j])
31
+ out.append(b)
32
+ return out
33
+ SEALED = {'seed': SEED, 'n_eligible_universe': len(elig) + len(sid) * 0, 'new_words': NEW,
34
+ 'sealed_id': [W(v) for v in sid], 'sealed_ood': [W(v) for v in sood],
35
+ 'sealed_1_new': novel(1), 'sealed_2_new': novel(2), 'sealed_4_new': novel(4)}
36
+ s = json.dumps(SEALED, sort_keys=True); open(sys.argv[3], 'w').write(s)
37
+ print('sha256', hashlib.sha256(s.encode()).hexdigest(), 'eligible', len(elig), {k: len(v) for k, v in SEALED.items() if isinstance(v, list)})
020/predictions020.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
020/preregistro020.md ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # CUERDAS-020 — Pre-registro (escrito ANTES de cualquier evaluación)
2
+
3
+ Fecha: 2026-10-16 (America/Mexico_City). Autor: Erik Sanchez (con Grok Bot).
4
+
5
+ ## Conjunto sellado
6
+ - Generado por `make_sealed020.py` con semilla nueva **20261016** (nunca usada antes).
7
+ - Universo elegible: combinaciones del vocabulario 014 que NO están en filas 014, NO están en el test fresco 015–019 (semilla 15015) y NO están en el universo 013 → 17,494 combinaciones.
8
+ - `sealed_id` (200), `sealed_ood` (400, con al menos un par retenido), `sealed_1_new` / `sealed_2_new` / `sealed_4_new` (192 c/u; palabras nuevas de 1 token con ≥10 letras, imposibles en el pool de entrenamiento 018 que solo tiene 3–9 letras).
9
+ - Archivo `kd/sealed020.json`, **SHA-256 4c7737f66364a44786b4d9acb4686a6f31829d78cb5914f19eca1df74fe73393**. El kernel verifica el hash antes de evaluar.
10
+ - Opciones por campo (cerradas): vocabulario 014 del campo + las 8 palabras nuevas del campo.
11
+ - Se evalúa **una sola vez**, con los checkpoints 018 `k2long_s{17,29,43}.pt` (elegidos solo con validación en 018). Sin reentrenar, sin elegir nada después.
12
+
13
+ ## Métricas
14
+ - Exactitud exact-match (los 4 campos), lectura cerrada de primer token, consulta por defecto. IC Wilson 95%.
15
+ - Diferencia K=2 − plantilla de texto (pp), bootstrap pareado 10,000 remuestreos, por semilla y media de 3 semillas.
16
+ - Posiciones por evento = tokens del mensaje + 15 (7 compartidos + 4×2 colas); K=2 = 17.
17
+
18
+ ## Experimento 2: baselines de texto compactos (receptor congelado, zero-shot)
19
+ `template` (20 tokens de 014), `kv` = "action: a; target: t; status: s; window: w.", `abbrev` = "act=a tgt=t st=s win=w", `csv` = "action,target,status,window\na,t,s,w", `values_csv` = "a, t, s, w", `values` = "a t s w". Mismas consultas, opciones y conjuntos sellados.
20
+
21
+ ## Experimento 3: decodificadores sin LLM sobre las ranuras K=2 (hiperparámetros fijados aquí, sin ajuste)
22
+ - Entrenamiento: 20,000 mensajes de FIT con aumento de palabras aleatorias del pool de entrenamiento 018 (p=0.4 por campo, igual que el codificador). Objetivo: embedding léxico del valor de cada campo. Decodificación: coseno contra las opciones.
23
+ - `ridge_aug` (λ=10); `mlp_aug` (1152→1024→GELU→1024→GELU→4×576, Adam lr 1e-3, 4,000 pasos, lote 256, pérdida coseno); `mlp_ctrl` = igual pero con objetivos barajados (control de Hewitt & Liang); `knn1`/`knn5` = vecinos más cercanos (coseno) sobre las ranuras de FIT (solo pueden emitir valores conocidos); `mlp_cls_fit` = clasificador MLP sobre FIT sin aumento (solo valores conocidos).
24
+ - Receptor con pesos aleatorios (semilla 0, mismo léxico y cabeza): texto y K=2 en `sealed_ood` y `sealed_4_new`.
25
+
26
+ ## Reglas de decisión (fijadas antes)
27
+ 1. **Test sellado**: la ventaja "K=2 ≈ texto" sobrevive si la media de 3 semillas en `sealed_ood` está a ≥ −2 pp de la plantilla Y el límite inferior del bootstrap de la media es > −3 pp.
28
+ 2. **Baselines compactos**: la ventaja en posiciones "sobrevive" frente al mejor formato de texto que esté a ≥ −2 pp de la plantilla en `sealed_ood`; se reporta el ahorro K=2 vs ese formato (posiciones y %). Si ese formato también gana en palabras nuevas, se dice.
29
+ 3. **Lookup sin LLM**: si algún decodificador sin LLM (excepto el control) alcanza ≥ la exactitud del LLM con K=2 en `sealed_4_new` (media de semillas), se concluye que la información es recuperable sin el LLM y se retira la afirmación de "no-lookup". Si queda muy por debajo, solo se dice que *estos* decodificadores fallan.
30
+ Todo resultado se reporta, sea bueno o malo.
020/resultados020.md ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # CUERDAS-020 — Test sellado, baselines de texto compactos y decodificadores más fuertes
2
+
3
+ Fecha: 16 oct 2026 (hora CDMX). Kernel privado de Kaggle `cuerdas020-sealed` (Tesla T4, PyTorch 2.11.0+cu128, transformers 4.46.3, Python 3.13.15), ≈6 min de GPU. Sin entrenamiento: se usaron los checkpoints 018 `k2long_s{17,29,43}.pt`, elegidos en 018 solo con validación.
4
+
5
+ Pre-registro: `preregistro020.md` (escrito antes de correr; SHA-256 5bc64fb6…). Conjunto sellado `sealed020.json`, SHA-256 `4c7737f6…3393`, verificado por el kernel antes de evaluar. **Se evaluó una sola vez.** Datos crudos: `gpu/results020.jsonl`, `gpu/predictions020.jsonl`.
6
+
7
+ ## 1. Test sellado (semilla nueva 20261016, nunca vista)
8
+
9
+ Exactitud exact-match (%), lectura cerrada de primer token, consulta por defecto. IC Wilson 95% entre corchetes. Opciones por campo = vocabulario 014 + 8 palabras nuevas.
10
+
11
+ | Conjunto (n) | Texto plantilla (35 pos.) | K=2 s17 | K=2 s29 | K=2 s43 | K=2 media | Δ media vs texto [bootstrap 95%] |
12
+ |---|---|---|---|---|---|---|
13
+ | sealed_id (200) | 99.5 [97.2, 99.9] | 98.5 [95.7, 99.5] | 100 [98.1, 100] | 98.0 [95.0, 99.2] | 98.83 | −0.67 [−2.00, +0.83] |
14
+ | **sealed_ood (400)** | **99.25 [97.8, 99.7]** | 99.0 [97.5, 99.6] | 99.75 [98.6, 100] | 98.25 [96.4, 99.1] | **99.0** | **−0.25 [−1.25, +0.92]** |
15
+ | sealed_1_new (192) | 96.4 [92.7, 98.2] | 95.8 | 98.4 | 85.9 | 93.4 | −2.95 [−6.08, +0.35] |
16
+ | sealed_2_new (192) | 98.4 [95.5, 99.5] | 90.6 | 97.9 | 87.0 | 91.8 | −6.60 [−9.72, −3.47] |
17
+ | sealed_4_new (192) | 94.3 [90.0, 96.8] | 82.8 [76.8, 87.5] | 87.5 [82.1, 91.5] | 82.3 [76.3, 87.0] | 84.2 | −10.07 [−15.10, −5.38] |
18
+
19
+ Por semilla en sealed_ood: s17 −0.25 [−1.5, +1.0]; s29 +0.5 [−0.5, +1.5]; s43 −1.0 [−2.5, +0.5] pp.
20
+
21
+ **Regla 1 (pre-registrada): se cumple.** La media en sealed_ood está a −0.25 pp del texto (≥ −2 pp) y el límite inferior del bootstrap (−1.25) es > −3 pp. La paridad con la plantilla en combinaciones nuevas de palabras conocidas **sobrevive** en un test que nadie había visto. Sigue sin ser una prueba de equivalencia formal (modelos fijos, 3 semillas).
22
+
23
+ Palabras nuevas: la brecha sigue siendo real (−3 a −10 pp), aunque menor que en desarrollo (4 nuevas: 84% sellado vs 55–73% en dev). Las palabras nuevas selladas son largas (≥10 letras) y de otro tipo que las de desarrollo, así que no se debe leer como mejora.
24
+
25
+ ## 2. Baselines de texto compactos (receptor congelado, zero-shot)
26
+
27
+ | Formato | Posiciones/evento (sealed_ood) | sealed_ood | sealed_id | sealed_4_new |
28
+ |---|---:|---:|---:|---:|
29
+ | plantilla 014 ("The action is …") | 35.0 | **99.25** | 99.5 | 94.3 |
30
+ | kv ("action: …; target: …") | 31.0 | 2.5 | 6.5 | 8.3 |
31
+ | csv con encabezado | 30.6 | 0.5 | 3.0 | 0.0 |
32
+ | abbrev ("act=… tgt=…") | 27.6 | 0.0 | 3.5 | 0.0 |
33
+ | values_csv ("a, t, s, w") | 22.0 | 0.25 | 3.0 | 0.0 |
34
+ | values ("a t s w") | 19.0 | 0.25 | 0.5 | 0.0 |
35
+ | **K=2 (media 3 semillas)** | **17** | **99.0** | 98.8 | 84.2 |
36
+
37
+ **Regla 2: ningún formato compacto queda a ≤2 pp de la plantilla**, así que el ahorro de posiciones de K=2 se mide contra la plantilla (35 → 17, −51%). Pero hay que decirlo claro: los formatos compactos fallan porque este receptor de 135M, zero-shot y con la consulta "According to the record, the {field} is", no sabe leerlos; no porque la información falte. La comparación **no es justa en entrenamiento**: el codificador K=2 se entrenó, el texto no. Un formato compacto con ejemplos (few-shot), con una consulta que use la misma etiqueta ("action:") o con un prompt aprendido podría funcionar con 19–23 posiciones. Eso no se probó. Conclusión honesta: K=2 le gana a texto compacto *zero-shot*, pero falta el baseline compacto *adaptado*.
38
+
39
+ ## 3. Decodificadores sin LLM sobre las ranuras K=2
40
+
41
+ Entrenados en 20,000 mensajes de FIT con palabras aleatorias del pool de entrenamiento (p=0.4 por campo), objetivo = embedding del valor, decodificación por coseno contra las opciones. Media de 3 semillas (por semilla entre paréntesis).
42
+
43
+ | Decodificador | sealed_ood (400) | sealed_1_new (192) | sealed_4_new (192) |
44
+ |---|---|---|---|
45
+ | LLM congelado (K=2) | 99.0 | 93.4 | **84.2** |
46
+ | ridge_aug (λ=10) | 100 (100/100/100) | 94.3 (93.2/92.7/96.9) | **64.6** (58.3/68.2/67.2) |
47
+ | mlp_aug (2 capas ocultas) | 100 | **100** (100/100/100) | 56.4 (54.7/64.6/50.0) |
48
+ | mlp con objetivos barajados (control) | 0 | 0 | 0 |
49
+ | mlp clasificador solo FIT | 99.75 | 0 | 0 |
50
+ | kNN-1 / kNN-5 sobre ranuras de FIT | 0 / 14.9 | 0 / 0 | 0 / 0 |
51
+ | Receptor con pesos aleatorios: texto / K=2 | 0 / 0 | — | 0 / 0 |
52
+
53
+ Δ vs LLM en sealed_4_new: ridge_aug −19.6 [−26.7, −12.5] pp; mlp_aug −27.8 [−35.2, −20.1] pp. En sealed_1_new: mlp_aug **+6.6 [+4.5, +8.9] pp** (mejor que el LLM).
54
+
55
+ (kNN no puede emitir palabras nuevas ni combinaciones fuera de FIT, por construcción; su 0% no dice nada sobre lookup.)
56
+
57
+ **Regla 3: ningún decodificador sin LLM alcanza al LLM en sealed_4_new**, así que formalmente no se retira la afirmación. Pero el resultado cambia la historia: el "0–2%" de la ridge/coseno de 018 era un artefacto de sondas que nunca vieron palabras nuevas. Con aumento de vocabulario, una ridge lineal recupera 58–68% de los mensajes con 4 palabras nuevas, y un MLP lee el 100% de los mensajes con 1 palabra nueva. **Las ranuras son en gran parte decodificables sin el LLM**: la información de las palabras viaja casi lineal en las ranuras. El LLM agrega ≈20 pp solo en el caso más difícil (4 nuevas). La afirmación "no es un lookup" debe bajarse a: "no es un lookup *simple*; un decodificador lineal entrenado con aumento recupera la mayoría".
58
+
59
+ ## 4. Veredicto honesto
60
+
61
+ - **Sobrevive:** la paridad K=2 ≈ plantilla de texto en combinaciones nuevas de palabras conocidas, ahora en un test sellado y pre-registrado (99.0% vs 99.25%, Δ −0.25 pp [−1.25, +0.92]) con 17 vs 35 posiciones.
62
+ - **No sobrevive tal cual:** la idea de que solo el LLM puede leer las ranuras. Una sonda lineal con aumento lee 58–68% con 4 palabras nuevas y ~93–100% con 1.
63
+ - **Sin resolver:** el ahorro de posiciones frente a un texto compacto *bien adaptado*. Zero-shot, todos los formatos compactos colapsan (0–8%), pero eso mide al receptor, no al formato.
64
+ - **Sigue débil:** transferencia a palabras nuevas (−3 a −10 pp vs texto).
65
+
66
+ Siguiente paso sugerido: baseline compacto few-shot o con prompt aprendido (mismo presupuesto de entrenamiento que K=2), generación abierta, y 10 semillas.
020/results020.jsonl ADDED
@@ -0,0 +1,99 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"env": {"python": "3.13.15", "torch": "2.11.0+cu128", "transformers": "4.46.3", "device": "Tesla T4"}, "sealed_sha256": "4c7737f66364a44786b4d9acb4686a6f31829d78cb5914f19eca1df74fe73393"}
2
+ {"route": "text_template", "set": "sealed_id", "n": 200, "acc": 0.995, "wilson95": [97.2, 99.9], "positions_per_event": 35.0}
3
+ {"route": "text_template", "set": "sealed_ood", "n": 400, "acc": 0.9925, "wilson95": [97.8, 99.7], "positions_per_event": 35.0}
4
+ {"route": "text_template", "set": "sealed_1_new", "n": 192, "acc": 0.9635416666666666, "wilson95": [92.7, 98.2], "positions_per_event": 35.0}
5
+ {"route": "text_template", "set": "sealed_2_new", "n": 192, "acc": 0.984375, "wilson95": [95.5, 99.5], "positions_per_event": 35.0}
6
+ {"route": "text_template", "set": "sealed_4_new", "n": 192, "acc": 0.9427083333333334, "wilson95": [90.0, 96.8], "positions_per_event": 35.0}
7
+ {"route": "text_kv", "set": "sealed_id", "n": 200, "acc": 0.065, "wilson95": [3.8, 10.8], "positions_per_event": 31.0, "diff_vs_template": -93.0, "boot95": [-96.5, -89.0]}
8
+ {"route": "text_kv", "set": "sealed_ood", "n": 400, "acc": 0.025, "wilson95": [1.4, 4.5], "positions_per_event": 31.0, "diff_vs_template": -96.75, "boot95": [-98.25, -95.0]}
9
+ {"route": "text_kv", "set": "sealed_1_new", "n": 192, "acc": 0.010416666666666666, "wilson95": [0.3, 3.7], "positions_per_event": 31.0, "diff_vs_template": -95.31, "boot95": [-98.44, -91.67]}
10
+ {"route": "text_kv", "set": "sealed_2_new", "n": 192, "acc": 0.052083333333333336, "wilson95": [2.9, 9.3], "positions_per_event": 31.0, "diff_vs_template": -93.23, "boot95": [-96.35, -89.58]}
11
+ {"route": "text_kv", "set": "sealed_4_new", "n": 192, "acc": 0.08333333333333333, "wilson95": [5.2, 13.1], "positions_per_event": 31.0, "diff_vs_template": -85.94, "boot95": [-90.62, -80.73]}
12
+ {"route": "text_abbrev", "set": "sealed_id", "n": 200, "acc": 0.035, "wilson95": [1.7, 7.0], "positions_per_event": 27.67, "diff_vs_template": -96.0, "boot95": [-98.5, -93.0]}
13
+ {"route": "text_abbrev", "set": "sealed_ood", "n": 400, "acc": 0.0, "wilson95": [0.0, 1.0], "positions_per_event": 27.63, "diff_vs_template": -99.25, "boot95": [-100.0, -98.25]}
14
+ {"route": "text_abbrev", "set": "sealed_1_new", "n": 192, "acc": 0.0, "wilson95": [0.0, 2.0], "positions_per_event": 29.1, "diff_vs_template": -96.35, "boot95": [-98.96, -93.23]}
15
+ {"route": "text_abbrev", "set": "sealed_2_new", "n": 192, "acc": 0.0, "wilson95": [0.0, 2.0], "positions_per_event": 30.47, "diff_vs_template": -98.44, "boot95": [-100.0, -96.35]}
16
+ {"route": "text_abbrev", "set": "sealed_4_new", "n": 192, "acc": 0.0, "wilson95": [0.0, 2.0], "positions_per_event": 33.37, "diff_vs_template": -94.27, "boot95": [-97.4, -90.62]}
17
+ {"route": "text_csv", "set": "sealed_id", "n": 200, "acc": 0.03, "wilson95": [1.4, 6.4], "positions_per_event": 30.67, "diff_vs_template": -96.5, "boot95": [-99.0, -93.5]}
18
+ {"route": "text_csv", "set": "sealed_ood", "n": 400, "acc": 0.005, "wilson95": [0.1, 1.8], "positions_per_event": 30.63, "diff_vs_template": -98.75, "boot95": [-99.75, -97.5]}
19
+ {"route": "text_csv", "set": "sealed_1_new", "n": 192, "acc": 0.005208333333333333, "wilson95": [0.1, 2.9], "positions_per_event": 32.1, "diff_vs_template": -95.83, "boot95": [-98.44, -92.71]}
20
+ {"route": "text_csv", "set": "sealed_2_new", "n": 192, "acc": 0.005208333333333333, "wilson95": [0.1, 2.9], "positions_per_event": 33.47, "diff_vs_template": -97.92, "boot95": [-99.48, -95.83]}
21
+ {"route": "text_csv", "set": "sealed_4_new", "n": 192, "acc": 0.0, "wilson95": [0.0, 2.0], "positions_per_event": 36.37, "diff_vs_template": -94.27, "boot95": [-97.4, -90.62]}
22
+ {"route": "text_values_csv", "set": "sealed_id", "n": 200, "acc": 0.03, "wilson95": [1.4, 6.4], "positions_per_event": 22.0, "diff_vs_template": -96.5, "boot95": [-99.0, -93.5]}
23
+ {"route": "text_values_csv", "set": "sealed_ood", "n": 400, "acc": 0.0025, "wilson95": [0.0, 1.4], "positions_per_event": 22.0, "diff_vs_template": -99.0, "boot95": [-99.75, -98.0]}
24
+ {"route": "text_values_csv", "set": "sealed_1_new", "n": 192, "acc": 0.0, "wilson95": [0.0, 2.0], "positions_per_event": 22.46, "diff_vs_template": -96.35, "boot95": [-98.96, -93.23]}
25
+ {"route": "text_values_csv", "set": "sealed_2_new", "n": 192, "acc": 0.0, "wilson95": [0.0, 2.0], "positions_per_event": 22.76, "diff_vs_template": -98.44, "boot95": [-100.0, -96.35]}
26
+ {"route": "text_values_csv", "set": "sealed_4_new", "n": 192, "acc": 0.0, "wilson95": [0.0, 2.0], "positions_per_event": 23.56, "diff_vs_template": -94.27, "boot95": [-97.4, -90.62]}
27
+ {"route": "text_values", "set": "sealed_id", "n": 200, "acc": 0.005, "wilson95": [0.1, 2.8], "positions_per_event": 19.0, "diff_vs_template": -99.0, "boot95": [-100.0, -97.5]}
28
+ {"route": "text_values", "set": "sealed_ood", "n": 400, "acc": 0.0025, "wilson95": [0.0, 1.4], "positions_per_event": 19.0, "diff_vs_template": -99.0, "boot95": [-99.75, -98.0]}
29
+ {"route": "text_values", "set": "sealed_1_new", "n": 192, "acc": 0.0, "wilson95": [0.0, 2.0], "positions_per_event": 19.46, "diff_vs_template": -96.35, "boot95": [-98.96, -93.23]}
30
+ {"route": "text_values", "set": "sealed_2_new", "n": 192, "acc": 0.0, "wilson95": [0.0, 2.0], "positions_per_event": 19.76, "diff_vs_template": -98.44, "boot95": [-100.0, -96.35]}
31
+ {"route": "text_values", "set": "sealed_4_new", "n": 192, "acc": 0.0, "wilson95": [0.0, 2.0], "positions_per_event": 20.56, "diff_vs_template": -94.27, "boot95": [-97.4, -90.62]}
32
+ {"route": "k2_s17", "set": "sealed_id", "n": 200, "acc": 0.985, "wilson95": [95.7, 99.5], "positions_per_event": 17, "diff_vs_template": -1.0, "boot95": [-3.0, 1.0]}
33
+ {"route": "k2_s17", "set": "sealed_ood", "n": 400, "acc": 0.99, "wilson95": [97.5, 99.6], "positions_per_event": 17, "diff_vs_template": -0.25, "boot95": [-1.5, 1.0]}
34
+ {"route": "k2_s17", "set": "sealed_1_new", "n": 192, "acc": 0.9583333333333334, "wilson95": [92.0, 97.9], "positions_per_event": 17, "diff_vs_template": -0.52, "boot95": [-4.17, 3.12]}
35
+ {"route": "k2_s17", "set": "sealed_2_new", "n": 192, "acc": 0.90625, "wilson95": [85.7, 94.0], "positions_per_event": 17, "diff_vs_template": -7.81, "boot95": [-12.5, -3.12]}
36
+ {"route": "k2_s17", "set": "sealed_4_new", "n": 192, "acc": 0.828125, "wilson95": [76.8, 87.5], "positions_per_event": 17, "diff_vs_template": -11.46, "boot95": [-17.19, -5.73]}
37
+ {"route": "k2_s29", "set": "sealed_id", "n": 200, "acc": 1.0, "wilson95": [98.1, 100.0], "positions_per_event": 17, "diff_vs_template": 0.5, "boot95": [0.0, 1.5]}
38
+ {"route": "k2_s29", "set": "sealed_ood", "n": 400, "acc": 0.9975, "wilson95": [98.6, 100.0], "positions_per_event": 17, "diff_vs_template": 0.5, "boot95": [-0.5, 1.5]}
39
+ {"route": "k2_s29", "set": "sealed_1_new", "n": 192, "acc": 0.984375, "wilson95": [95.5, 99.5], "positions_per_event": 17, "diff_vs_template": 2.08, "boot95": [-1.04, 5.21]}
40
+ {"route": "k2_s29", "set": "sealed_2_new", "n": 192, "acc": 0.9791666666666666, "wilson95": [94.8, 99.2], "positions_per_event": 17, "diff_vs_template": -0.52, "boot95": [-3.12, 2.08]}
41
+ {"route": "k2_s29", "set": "sealed_4_new", "n": 192, "acc": 0.875, "wilson95": [82.1, 91.5], "positions_per_event": 17, "diff_vs_template": -6.77, "boot95": [-11.98, -2.08]}
42
+ {"route": "k2_s43", "set": "sealed_id", "n": 200, "acc": 0.98, "wilson95": [95.0, 99.2], "positions_per_event": 17, "diff_vs_template": -1.5, "boot95": [-4.0, 0.5]}
43
+ {"route": "k2_s43", "set": "sealed_ood", "n": 400, "acc": 0.9825, "wilson95": [96.4, 99.1], "positions_per_event": 17, "diff_vs_template": -1.0, "boot95": [-2.5, 0.5]}
44
+ {"route": "k2_s43", "set": "sealed_1_new", "n": 192, "acc": 0.859375, "wilson95": [80.3, 90.2], "positions_per_event": 17, "diff_vs_template": -10.42, "boot95": [-15.62, -5.21]}
45
+ {"route": "k2_s43", "set": "sealed_2_new", "n": 192, "acc": 0.8697916666666666, "wilson95": [81.5, 91.0], "positions_per_event": 17, "diff_vs_template": -11.46, "boot95": [-16.67, -6.77]}
46
+ {"route": "k2_s43", "set": "sealed_4_new", "n": 192, "acc": 0.8229166666666666, "wilson95": [76.3, 87.0], "positions_per_event": 17, "diff_vs_template": -11.98, "boot95": [-18.23, -6.25]}
47
+ {"route": "k2_mean3", "set": "sealed_id", "acc": 0.9883333333333333, "vs": "text_template", "diff": -0.67, "boot95": [-2.0, 0.83]}
48
+ {"route": "k2_mean3", "set": "sealed_id", "acc": 0.9883333333333333, "vs": "text_kv", "diff": 92.33, "boot95": [88.67, 95.67]}
49
+ {"route": "k2_mean3", "set": "sealed_id", "acc": 0.9883333333333333, "vs": "text_abbrev", "diff": 95.33, "boot95": [92.5, 97.83]}
50
+ {"route": "k2_mean3", "set": "sealed_id", "acc": 0.9883333333333333, "vs": "text_csv", "diff": 95.83, "boot95": [93.17, 98.17]}
51
+ {"route": "k2_mean3", "set": "sealed_id", "acc": 0.9883333333333333, "vs": "text_values_csv", "diff": 95.83, "boot95": [93.17, 98.17]}
52
+ {"route": "k2_mean3", "set": "sealed_id", "acc": 0.9883333333333333, "vs": "text_values", "diff": 98.33, "boot95": [96.83, 99.5]}
53
+ {"route": "k2_mean3", "set": "sealed_ood", "acc": 0.99, "vs": "text_template", "diff": -0.25, "boot95": [-1.25, 0.92]}
54
+ {"route": "k2_mean3", "set": "sealed_ood", "acc": 0.99, "vs": "text_kv", "diff": 96.5, "boot95": [94.75, 98.0]}
55
+ {"route": "k2_mean3", "set": "sealed_ood", "acc": 0.99, "vs": "text_abbrev", "diff": 99.0, "boot95": [98.33, 99.58]}
56
+ {"route": "k2_mean3", "set": "sealed_ood", "acc": 0.99, "vs": "text_csv", "diff": 98.5, "boot95": [97.5, 99.33]}
57
+ {"route": "k2_mean3", "set": "sealed_ood", "acc": 0.99, "vs": "text_values_csv", "diff": 98.75, "boot95": [97.83, 99.5]}
58
+ {"route": "k2_mean3", "set": "sealed_ood", "acc": 0.99, "vs": "text_values", "diff": 98.75, "boot95": [97.83, 99.5]}
59
+ {"route": "k2_mean3", "set": "sealed_1_new", "acc": 0.9340277777777778, "vs": "text_template", "diff": -2.95, "boot95": [-6.08, 0.35]}
60
+ {"route": "k2_mean3", "set": "sealed_1_new", "acc": 0.9340277777777778, "vs": "text_kv", "diff": 92.36, "boot95": [89.41, 94.97]}
61
+ {"route": "k2_mean3", "set": "sealed_1_new", "acc": 0.9340277777777778, "vs": "text_abbrev", "diff": 93.4, "boot95": [90.97, 95.49]}
62
+ {"route": "k2_mean3", "set": "sealed_1_new", "acc": 0.9340277777777778, "vs": "text_csv", "diff": 92.88, "boot95": [90.28, 95.14]}
63
+ {"route": "k2_mean3", "set": "sealed_1_new", "acc": 0.9340277777777778, "vs": "text_values_csv", "diff": 93.4, "boot95": [90.97, 95.49]}
64
+ {"route": "k2_mean3", "set": "sealed_1_new", "acc": 0.9340277777777778, "vs": "text_values", "diff": 93.4, "boot95": [90.97, 95.49]}
65
+ {"route": "k2_mean3", "set": "sealed_2_new", "acc": 0.9184027777777778, "vs": "text_template", "diff": -6.6, "boot95": [-9.72, -3.47]}
66
+ {"route": "k2_mean3", "set": "sealed_2_new", "acc": 0.9184027777777778, "vs": "text_kv", "diff": 86.63, "boot95": [82.47, 90.62]}
67
+ {"route": "k2_mean3", "set": "sealed_2_new", "acc": 0.9184027777777778, "vs": "text_abbrev", "diff": 91.84, "boot95": [89.06, 94.44]}
68
+ {"route": "k2_mean3", "set": "sealed_2_new", "acc": 0.9184027777777778, "vs": "text_csv", "diff": 91.32, "boot95": [88.37, 94.1]}
69
+ {"route": "k2_mean3", "set": "sealed_2_new", "acc": 0.9184027777777778, "vs": "text_values_csv", "diff": 91.84, "boot95": [89.06, 94.44]}
70
+ {"route": "k2_mean3", "set": "sealed_2_new", "acc": 0.9184027777777778, "vs": "text_values", "diff": 91.84, "boot95": [89.06, 94.44]}
71
+ {"route": "k2_mean3", "set": "sealed_4_new", "acc": 0.8420138888888888, "vs": "text_template", "diff": -10.07, "boot95": [-15.1, -5.38]}
72
+ {"route": "k2_mean3", "set": "sealed_4_new", "acc": 0.8420138888888888, "vs": "text_kv", "diff": 75.87, "boot95": [70.14, 81.25]}
73
+ {"route": "k2_mean3", "set": "sealed_4_new", "acc": 0.8420138888888888, "vs": "text_abbrev", "diff": 84.2, "boot95": [79.51, 88.37]}
74
+ {"route": "k2_mean3", "set": "sealed_4_new", "acc": 0.8420138888888888, "vs": "text_csv", "diff": 84.2, "boot95": [79.51, 88.37]}
75
+ {"route": "k2_mean3", "set": "sealed_4_new", "acc": 0.8420138888888888, "vs": "text_values_csv", "diff": 84.2, "boot95": [79.51, 88.37]}
76
+ {"route": "k2_mean3", "set": "sealed_4_new", "acc": 0.8420138888888888, "vs": "text_values", "diff": 84.2, "boot95": [79.51, 88.37]}
77
+ {"route": "probes_k2_s17", "ridge_aug:sealed_ood": 1.0, "mlp_aug:sealed_ood": 1.0, "mlp_ctrl_shuffled:sealed_ood": 0.0, "mlp_cls_fit:sealed_ood": 0.995, "knn1:sealed_ood": 0.0, "knn5:sealed_ood": 0.1325, "ridge_aug:sealed_4_new": 0.5833333333333334, "mlp_aug:sealed_4_new": 0.546875, "mlp_ctrl_shuffled:sealed_4_new": 0.0, "mlp_cls_fit:sealed_4_new": 0.0, "knn1:sealed_4_new": 0.0, "knn5:sealed_4_new": 0.0, "ridge_aug:sealed_1_new": 0.9322916666666666, "mlp_aug:sealed_1_new": 1.0, "mlp_ctrl_shuffled:sealed_1_new": 0.0, "mlp_cls_fit:sealed_1_new": 0.0, "knn1:sealed_1_new": 0.0, "knn5:sealed_1_new": 0.0}
78
+ {"route": "probes_k2_s29", "ridge_aug:sealed_ood": 1.0, "mlp_aug:sealed_ood": 1.0, "mlp_ctrl_shuffled:sealed_ood": 0.0, "mlp_cls_fit:sealed_ood": 0.9975, "knn1:sealed_ood": 0.0, "knn5:sealed_ood": 0.14, "ridge_aug:sealed_4_new": 0.6822916666666666, "mlp_aug:sealed_4_new": 0.6458333333333334, "mlp_ctrl_shuffled:sealed_4_new": 0.0, "mlp_cls_fit:sealed_4_new": 0.0, "knn1:sealed_4_new": 0.0, "knn5:sealed_4_new": 0.0, "ridge_aug:sealed_1_new": 0.9270833333333334, "mlp_aug:sealed_1_new": 1.0, "mlp_ctrl_shuffled:sealed_1_new": 0.0, "mlp_cls_fit:sealed_1_new": 0.0, "knn1:sealed_1_new": 0.0, "knn5:sealed_1_new": 0.0}
79
+ {"route": "probes_k2_s43", "ridge_aug:sealed_ood": 1.0, "mlp_aug:sealed_ood": 1.0, "mlp_ctrl_shuffled:sealed_ood": 0.0, "mlp_cls_fit:sealed_ood": 1.0, "knn1:sealed_ood": 0.0, "knn5:sealed_ood": 0.175, "ridge_aug:sealed_4_new": 0.671875, "mlp_aug:sealed_4_new": 0.5, "mlp_ctrl_shuffled:sealed_4_new": 0.0, "mlp_cls_fit:sealed_4_new": 0.0, "knn1:sealed_4_new": 0.0, "knn5:sealed_4_new": 0.0, "ridge_aug:sealed_1_new": 0.96875, "mlp_aug:sealed_1_new": 1.0, "mlp_ctrl_shuffled:sealed_1_new": 0.0, "mlp_cls_fit:sealed_1_new": 0.0, "knn1:sealed_1_new": 0.0, "knn5:sealed_1_new": 0.0}
80
+ {"probe_mean3": "ridge_aug", "set": "sealed_ood", "acc": 1.0, "llm_k2_acc": 0.99, "diff_vs_llm": 1.0, "boot95": [0.42, 1.67]}
81
+ {"probe_mean3": "mlp_aug", "set": "sealed_ood", "acc": 1.0, "llm_k2_acc": 0.99, "diff_vs_llm": 1.0, "boot95": [0.42, 1.67]}
82
+ {"probe_mean3": "mlp_ctrl_shuffled", "set": "sealed_ood", "acc": 0.0, "llm_k2_acc": 0.99, "diff_vs_llm": -99.0, "boot95": [-99.58, -98.33]}
83
+ {"probe_mean3": "mlp_cls_fit", "set": "sealed_ood", "acc": 0.9975, "llm_k2_acc": 0.99, "diff_vs_llm": 0.75, "boot95": [0.08, 1.5]}
84
+ {"probe_mean3": "knn1", "set": "sealed_ood", "acc": 0.0, "llm_k2_acc": 0.99, "diff_vs_llm": -99.0, "boot95": [-99.58, -98.33]}
85
+ {"probe_mean3": "knn5", "set": "sealed_ood", "acc": 0.14916666666666667, "llm_k2_acc": 0.99, "diff_vs_llm": -84.08, "boot95": [-87.0, -81.08]}
86
+ {"probe_mean3": "ridge_aug", "set": "sealed_4_new", "acc": 0.6458333333333334, "llm_k2_acc": 0.8420138888888888, "diff_vs_llm": -19.62, "boot95": [-26.74, -12.5]}
87
+ {"probe_mean3": "mlp_aug", "set": "sealed_4_new", "acc": 0.564236111111111, "llm_k2_acc": 0.8420138888888888, "diff_vs_llm": -27.78, "boot95": [-35.24, -20.14]}
88
+ {"probe_mean3": "mlp_ctrl_shuffled", "set": "sealed_4_new", "acc": 0.0, "llm_k2_acc": 0.8420138888888888, "diff_vs_llm": -84.2, "boot95": [-88.54, -79.69]}
89
+ {"probe_mean3": "mlp_cls_fit", "set": "sealed_4_new", "acc": 0.0, "llm_k2_acc": 0.8420138888888888, "diff_vs_llm": -84.2, "boot95": [-88.54, -79.69]}
90
+ {"probe_mean3": "knn1", "set": "sealed_4_new", "acc": 0.0, "llm_k2_acc": 0.8420138888888888, "diff_vs_llm": -84.2, "boot95": [-88.54, -79.69]}
91
+ {"probe_mean3": "knn5", "set": "sealed_4_new", "acc": 0.0, "llm_k2_acc": 0.8420138888888888, "diff_vs_llm": -84.2, "boot95": [-88.54, -79.69]}
92
+ {"probe_mean3": "ridge_aug", "set": "sealed_1_new", "acc": 0.9427083333333334, "llm_k2_acc": 0.9340277777777778, "diff_vs_llm": 0.87, "boot95": [-2.95, 4.51]}
93
+ {"probe_mean3": "mlp_aug", "set": "sealed_1_new", "acc": 1.0, "llm_k2_acc": 0.9340277777777778, "diff_vs_llm": 6.6, "boot95": [4.51, 8.85]}
94
+ {"probe_mean3": "mlp_ctrl_shuffled", "set": "sealed_1_new", "acc": 0.0, "llm_k2_acc": 0.9340277777777778, "diff_vs_llm": -93.4, "boot95": [-95.49, -91.15]}
95
+ {"probe_mean3": "mlp_cls_fit", "set": "sealed_1_new", "acc": 0.0, "llm_k2_acc": 0.9340277777777778, "diff_vs_llm": -93.4, "boot95": [-95.49, -91.15]}
96
+ {"probe_mean3": "knn1", "set": "sealed_1_new", "acc": 0.0, "llm_k2_acc": 0.9340277777777778, "diff_vs_llm": -93.4, "boot95": [-95.49, -91.15]}
97
+ {"probe_mean3": "knn5", "set": "sealed_1_new", "acc": 0.0, "llm_k2_acc": 0.9340277777777778, "diff_vs_llm": -93.4, "boot95": [-95.49, -91.15]}
98
+ {"control": "random_receiver", "set": "sealed_ood", "text_template": 0.0, "k2_s17": 0.0, "k2_s29": 0.0, "k2_s43": 0.0}
99
+ {"control": "random_receiver", "set": "sealed_4_new", "text_template": 0.0, "k2_s17": 0.0, "k2_s29": 0.0, "k2_s43": 0.0}
020/sealed020.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"n_eligible_universe": 17494, "new_words": [["diagnostics", "manipulative", "favourable", "intelligence", "phonological", "persecuted", "contingency", "quantifying"], ["unaffected", "prerequisites", "dehydration", "intonation", "discerning", "kilometres", "watersheds", "foreseeable"], ["companions", "pediatrics", "epithelium", "complained", "multiplied", "sunglasses", "redistribute", "disclosure"], ["similarities", "applicability", "abnormality", "encountering", "predominantly", "provincial", "modalities", "universality"]], "sealed_1_new": [["start", "discerning", "open", "none"], ["update", "job", "sunglasses", "hour"], ["stop", "prerequisites", "active", "year"], ["start", "meeting", "complained", "month"], ["stop", "unaffected", "active", "minute"], ["read", "pump", "paid", "applicability"], ["execute", "order", "sunglasses", "day"], ["delete", "machine", "disclosure", "minute"], ["read", "inventory", "epithelium", "day"], ["favourable", "invoice", "failed", "month"], ["stop", "unaffected", "failed", "minute"], ["write", "pump", "multiplied", "today"], ["manipulative", "inventory", "high", "today"], ["quantifying", "meeting", "high", "hour"], ["query", "file", "companions", "week"], ["create", "pump", "pending", "abnormality"], ["quantifying", "customer", "idle", "year"], ["contingency", "motor", "pending", "tomorrow"], ["stop", "watersheds", "active", "now"], ["report", "unaffected", "running", "day"], ["write", "unaffected", "ready", "now"], ["persecuted", "pump", "paid", "none"], ["update", "watersheds", "blocked", "week"], ["notify", "prerequisites", "running", "hour"], ["create", "valve", "complete", "predominantly"], ["quantifying", "file", "running", "hour"], ["query", "file", "normal", "abnormality"], ["manipulative", "valve", "blocked", "tomorrow"], ["phonological", "valve", "low", "now"], ["contingency", "order", "ready", "month"], ["update", "order", "multiplied", "tomorrow"], ["persecuted", "invoice", "none", "today"], ["write", "pump", "none", "universality"], ["persecuted", "order", "high", "today"], ["notify", "kilometres", "ready", "tomorrow"], ["write", "intonation", "low", "day"], ["check", "customer", "multiplied", "day"], ["start", "meeting", "pediatrics", "hour"], ["report", "motor", "complained", "minute"], ["check", "machine", "complained", "today"], ["check", "file", "idle", "universality"], ["check", "file", "multiplied", "year"], ["write", "watersheds", "pending", "week"], ["start", "kilometres", "low", "today"], ["query", "machine", "companions", "today"], ["write", "valve", "sunglasses", "month"], ["query", "motor", "ready", "abnormality"], ["write", "valve", "pediatrics", "minute"], ["diagnostics", "inventory", "pending", "year"], ["start", "invoice", "paid", "universality"], ["write", "prerequisites", "closed", "day"], ["execute", "valve", "disclosure", "month"], ["read", "file", "multiplied", "tomorrow"], ["write", "dehydration", "low", "day"], ["persecuted", "valve", "idle", "tomorrow"], ["query", "file", "paid", "modalities"], ["stop", "email", "epithelium", "none"], ["check", "foreseeable", "ready", "hour"], ["write", "watersheds", "high", "hour"], ["report", "kilometres", "normal", "minute"], ["quantifying", "order", "running", "tomorrow"], ["notify", "motor", "none", "universality"], ["read", "machine", "high", "abnormality"], ["notify", "order", "companions", "week"], ["report", "discerning", "failed", "none"], ["intelligence", "order", "none", "year"], ["contingency", "customer", "complete", "year"], ["diagnostics", "order", "high", "hour"], ["update", "watersheds", "running", "day"], ["check", "motor", "open", "modalities"], ["execute", "customer", "epithelium", "now"], ["contingency", "machine", "active", "minute"], ["execute", "meeting", "closed", "provincial"], ["read", "kilometres", "complete", "year"], ["diagnostics", "job", "running", "now"], ["start", "meeting", "running", "modalities"], ["quantifying", "file", "overload", "year"], ["stop", "machine", "epithelium", "tomorrow"], ["delete", "machine", "idle", "similarities"], ["contingency", "meeting", "active", "week"], ["update", "order", "high", "encountering"], ["favourable", "machine", "overload", "month"], ["check", "file", "low", "similarities"], ["check", "pump", "epithelium", "day"], ["intelligence", "inventory", "idle", "year"], ["check", "motor", "open", "modalities"], ["check", "file", "pediatrics", "week"], ["read", "kilometres", "idle", "today"], ["write", "discerning", "overload", "today"], ["delete", "machine", "running", "predominantly"], ["persecuted", "email", "pending", "month"], ["favourable", "invoice", "blocked", "none"], ["update", "watersheds", "paid", "month"], ["notify", "email", "blocked", "universality"], ["phonological", "meeting", "normal", "year"], ["write", "intonation", "normal", "day"], ["delete", "customer", "complained", "month"], ["report", "pump", "pediatrics", "hour"], ["manipulative", "file", "high", "hour"], ["report", "customer", "high", "similarities"], ["write", "pump", "complained", "month"], ["intelligence", "email", "open", "tomorrow"], ["diagnostics", "inventory", "running", "month"], ["notify", "customer", "running", "universality"], ["notify", "valve", "redistribute", "hour"], ["execute", "dehydration", "active", "year"], ["query", "motor", "redistribute", "none"], ["quantifying", "pump", "closed", "day"], ["write", "meeting", "idle", "encountering"], ["query", "foreseeable", "normal", "none"], ["write", "motor", "complained", "none"], ["execute", "watersheds", "active", "year"], ["manipulative", "order", "pending", "minute"], ["report", "motor", "open", "predominantly"], ["intelligence", "pump", "open", "tomorrow"], ["update", "customer", "failed", "applicability"], ["persecuted", "meeting", "complete", "day"], ["check", "valve", "failed", "encountering"], ["notify", "customer", "complete", "predominantly"], ["delete", "pump", "sunglasses", "minute"], ["phonological", "invoice", "failed", "hour"], ["check", "unaffected", "normal", "hour"], ["report", "invoice", "high", "provincial"], ["persecuted", "motor", "high", "minute"], ["execute", "intonation", "running", "minute"], ["notify", "dehydration", "idle", "tomorrow"], ["persecuted", "order", "none", "now"], ["execute", "kilometres", "closed", "year"], ["manipulative", "motor", "normal", "year"], ["execute", "valve", "redistribute", "none"], ["favourable", "motor", "closed", "minute"], ["notify", "job", "ready", "encountering"], ["persecuted", "customer", "failed", "minute"], ["report", "intonation", "overload", "month"], ["write", "valve", "closed", "similarities"], ["query", "machine", "complained", "minute"], ["execute", "inventory", "epithelium", "hour"], ["execute", "pump", "low", "provincial"], ["read", "job", "epithelium", "year"], ["read", "customer", "redistribute", "minute"], ["read", "foreseeable", "complete", "day"], ["stop", "inventory", "complained", "now"], ["query", "pump", "failed", "provincial"], ["report", "customer", "pending", "provincial"], ["query", "meeting", "sunglasses", "year"], ["update", "foreseeable", "normal", "minute"], ["check", "valve", "failed", "provincial"], ["notify", "inventory", "pediatrics", "day"], ["update", "inventory", "none", "modalities"], ["persecuted", "order", "ready", "hour"], ["read", "unaffected", "ready", "month"], ["query", "unaffected", "idle", "day"], ["stop", "inventory", "redistribute", "week"], ["start", "email", "sunglasses", "day"], ["diagnostics", "job", "ready", "month"], ["favourable", "customer", "closed", "none"], ["check", "intonation", "running", "minute"], ["create", "dehydration", "closed", "day"], ["start", "watersheds", "pending", "week"], ["update", "order", "idle", "predominantly"], ["delete", "machine", "redistribute", "tomorrow"], ["report", "customer", "disclosure", "today"], ["read", "foreseeable", "closed", "day"], ["check", "intonation", "running", "week"], ["create", "job", "sunglasses", "month"], ["intelligence", "machine", "blocked", "minute"], ["start", "customer", "paid", "encountering"], ["notify", "kilometres", "ready", "tomorrow"], ["contingency", "meeting", "blocked", "tomorrow"], ["query", "invoice", "failed", "abnormality"], ["persecuted", "job", "ready", "none"], ["manipulative", "invoice", "paid", "minute"], ["read", "prerequisites", "paid", "month"], ["contingency", "customer", "overload", "week"], ["persecuted", "invoice", "blocked", "minute"], ["execute", "valve", "pediatrics", "month"], ["delete", "pump", "none", "applicability"], ["report", "invoice", "open", "predominantly"], ["contingency", "invoice", "active", "minute"], ["delete", "machine", "sunglasses", "today"], ["persecuted", "file", "open", "hour"], ["start", "invoice", "redistribute", "week"], ["persecuted", "order", "open", "tomorrow"], ["check", "file", "multiplied", "day"], ["stop", "intonation", "low", "day"], ["manipulative", "invoice", "idle", "minute"], ["start", "order", "disclosure", "month"], ["report", "customer", "low", "universality"], ["stop", "dehydration", "idle", "tomorrow"], ["execute", "discerning", "idle", "tomorrow"], ["execute", "unaffected", "running", "year"], ["quantifying", "valve", "high", "today"]], "sealed_2_new": [["start", "discerning", "open", "universality"], ["update", "kilometres", "failed", "applicability"], ["phonological", "customer", "pediatrics", "year"], ["intelligence", "meeting", "low", "modalities"], ["stop", "intonation", "active", "encountering"], ["read", "foreseeable", "paid", "predominantly"], ["execute", "order", "sunglasses", "predominantly"], ["delete", "machine", "complained", "abnormality"], ["read", "inventory", "companions", "encountering"], ["read", "dehydration", "failed", "abnormality"], ["stop", "intonation", "sunglasses", "minute"], ["write", "pump", "multiplied", "provincial"], ["phonological", "discerning", "high", "today"], ["write", "meeting", "redistribute", "encountering"], ["query", "file", "disclosure", "encountering"], ["manipulative", "unaffected", "pending", "week"], ["persecuted", "customer", "idle", "modalities"], ["quantifying", "motor", "complained", "tomorrow"], ["intelligence", "inventory", "companions", "now"], ["report", "customer", "complained", "encountering"], ["contingency", "email", "ready", "applicability"], ["quantifying", "intonation", "paid", "none"], ["update", "watersheds", "epithelium", "week"], ["notify", "job", "epithelium", "similarities"], ["create", "valve", "pediatrics", "modalities"], ["diagnostics", "kilometres", "running", "hour"], ["phonological", "file", "companions", "today"], ["write", "valve", "epithelium", "predominantly"], ["diagnostics", "valve", "low", "predominantly"], ["phonological", "order", "ready", "predominantly"], ["update", "order", "disclosure", "modalities"], ["start", "unaffected", "companions", "today"], ["write", "prerequisites", "none", "universality"], ["persecuted", "order", "high", "abnormality"], ["favourable", "job", "pediatrics", "tomorrow"], ["write", "meeting", "multiplied", "abnormality"], ["check", "customer", "redistribute", "universality"], ["quantifying", "intonation", "blocked", "hour"], ["intelligence", "motor", "overload", "predominantly"], ["check", "machine", "disclosure", "applicability"], ["diagnostics", "file", "idle", "predominantly"], ["phonological", "kilometres", "high", "year"], ["write", "watersheds", "pending", "provincial"], ["contingency", "unaffected", "low", "today"], ["persecuted", "machine", "closed", "universality"], ["persecuted", "valve", "paid", "applicability"], ["persecuted", "motor", "ready", "universality"], ["diagnostics", "valve", "idle", "applicability"], ["stop", "inventory", "redistribute", "provincial"], ["persecuted", "foreseeable", "paid", "none"], ["quantifying", "email", "closed", "modalities"], ["execute", "prerequisites", "running", "similarities"], ["quantifying", "file", "complained", "tomorrow"], ["favourable", "valve", "redistribute", "day"], ["contingency", "foreseeable", "idle", "tomorrow"], ["quantifying", "file", "epithelium", "year"], ["diagnostics", "dehydration", "open", "none"], ["check", "dehydration", "ready", "modalities"], ["manipulative", "foreseeable", "high", "hour"], ["quantifying", "motor", "multiplied", "minute"], ["execute", "intonation", "complained", "tomorrow"], ["quantifying", "dehydration", "none", "month"], ["read", "machine", "sunglasses", "predominantly"], ["favourable", "order", "redistribute", "week"], ["report", "kilometres", "sunglasses", "none"], ["persecuted", "order", "none", "similarities"], ["execute", "prerequisites", "complete", "modalities"], ["intelligence", "order", "complained", "hour"], ["favourable", "pump", "disclosure", "day"], ["check", "foreseeable", "epithelium", "week"], ["execute", "intonation", "disclosure", "now"], ["create", "prerequisites", "active", "provincial"], ["execute", "meeting", "companions", "applicability"], ["favourable", "intonation", "complete", "year"], ["diagnostics", "job", "companions", "now"], ["persecuted", "meeting", "multiplied", "month"], ["intelligence", "file", "companions", "year"], ["diagnostics", "intonation", "open", "tomorrow"], ["delete", "intonation", "idle", "predominantly"], ["create", "meeting", "multiplied", "similarities"], ["update", "unaffected", "high", "applicability"], ["manipulative", "machine", "overload", "similarities"], ["manipulative", "file", "low", "abnormality"], ["intelligence", "pump", "pediatrics", "day"], ["update", "intonation", "companions", "year"], ["check", "kilometres", "pediatrics", "day"], ["check", "foreseeable", "companions", "week"], ["read", "pump", "disclosure", "modalities"], ["write", "email", "complained", "predominantly"], ["manipulative", "discerning", "running", "now"], ["persecuted", "email", "pending", "modalities"], ["phonological", "invoice", "blocked", "encountering"], ["update", "intonation", "redistribute", "month"], ["notify", "kilometres", "blocked", "predominantly"], ["persecuted", "discerning", "normal", "year"], ["persecuted", "kilometres", "normal", "day"], ["delete", "dehydration", "ready", "provincial"], ["manipulative", "watersheds", "none", "hour"], ["create", "watersheds", "high", "modalities"], ["report", "watersheds", "pediatrics", "day"], ["write", "prerequisites", "paid", "encountering"], ["contingency", "email", "open", "applicability"], ["update", "dehydration", "companions", "month"], ["manipulative", "discerning", "running", "day"], ["diagnostics", "valve", "epithelium", "hour"], ["execute", "unaffected", "companions", "year"], ["favourable", "motor", "open", "similarities"], ["start", "pump", "pediatrics", "applicability"], ["write", "discerning", "idle", "provincial"], ["query", "kilometres", "sunglasses", "none"], ["write", "intonation", "blocked", "modalities"], ["execute", "foreseeable", "active", "encountering"], ["update", "order", "multiplied", "applicability"], ["report", "foreseeable", "open", "similarities"], ["quantifying", "pump", "epithelium", "tomorrow"], ["update", "customer", "complained", "predominantly"], ["stop", "unaffected", "complete", "predominantly"], ["check", "kilometres", "sunglasses", "today"], ["quantifying", "customer", "redistribute", "day"], ["diagnostics", "pump", "multiplied", "minute"], ["delete", "watersheds", "pediatrics", "hour"], ["quantifying", "file", "normal", "predominantly"], ["report", "watersheds", "high", "modalities"], ["check", "foreseeable", "high", "provincial"], ["quantifying", "valve", "running", "universality"], ["quantifying", "unaffected", "idle", "tomorrow"], ["persecuted", "prerequisites", "none", "now"], ["execute", "order", "companions", "provincial"], ["check", "motor", "redistribute", "abnormality"], ["intelligence", "valve", "active", "encountering"], ["report", "watersheds", "complained", "minute"], ["favourable", "job", "pediatrics", "month"], ["intelligence", "dehydration", "failed", "minute"], ["report", "watersheds", "overload", "encountering"], ["diagnostics", "valve", "closed", "abnormality"], ["contingency", "prerequisites", "idle", "minute"], ["phonological", "inventory", "paid", "universality"], ["execute", "pump", "pediatrics", "encountering"], ["contingency", "job", "companions", "year"], ["favourable", "watersheds", "active", "minute"], ["diagnostics", "prerequisites", "complete", "day"], ["stop", "discerning", "overload", "abnormality"], ["query", "unaffected", "pediatrics", "hour"], ["contingency", "customer", "pending", "modalities"], ["manipulative", "meeting", "complete", "predominantly"], ["update", "foreseeable", "normal", "predominantly"], ["check", "intonation", "failed", "applicability"], ["notify", "discerning", "multiplied", "day"], ["manipulative", "kilometres", "none", "hour"], ["read", "dehydration", "ready", "modalities"], ["phonological", "file", "ready", "applicability"], ["query", "file", "disclosure", "applicability"], ["stop", "inventory", "multiplied", "universality"], ["favourable", "email", "pediatrics", "day"], ["manipulative", "job", "ready", "provincial"], ["delete", "intonation", "closed", "encountering"], ["check", "unaffected", "running", "applicability"], ["persecuted", "email", "epithelium", "day"], ["start", "intonation", "pending", "encountering"], ["favourable", "order", "pediatrics", "year"], ["phonological", "dehydration", "high", "tomorrow"], ["intelligence", "customer", "paid", "abnormality"], ["read", "prerequisites", "disclosure", "day"], ["check", "kilometres", "multiplied", "week"], ["intelligence", "job", "ready", "applicability"], ["contingency", "machine", "epithelium", "minute"], ["start", "kilometres", "epithelium", "hour"], ["notify", "unaffected", "ready", "provincial"], ["delete", "kilometres", "blocked", "encountering"], ["persecuted", "invoice", "failed", "universality"], ["phonological", "job", "ready", "universality"], ["quantifying", "watersheds", "paid", "minute"], ["read", "email", "epithelium", "provincial"], ["report", "customer", "companions", "provincial"], ["quantifying", "prerequisites", "blocked", "minute"], ["diagnostics", "valve", "disclosure", "month"], ["intelligence", "pump", "none", "modalities"], ["report", "unaffected", "pediatrics", "tomorrow"], ["quantifying", "foreseeable", "active", "minute"], ["phonological", "machine", "multiplied", "today"], ["diagnostics", "file", "open", "applicability"], ["start", "foreseeable", "sunglasses", "week"], ["write", "intonation", "open", "predominantly"], ["persecuted", "file", "normal", "encountering"], ["stop", "invoice", "complained", "predominantly"], ["quantifying", "invoice", "idle", "abnormality"], ["phonological", "prerequisites", "paid", "month"], ["favourable", "customer", "disclosure", "year"], ["stop", "prerequisites", "idle", "provincial"], ["execute", "pump", "disclosure", "similarities"], ["persecuted", "foreseeable", "running", "year"], ["manipulative", "valve", "high", "similarities"]], "sealed_4_new": [["diagnostics", "watersheds", "redistribute", "predominantly"], ["phonological", "unaffected", "companions", "applicability"], ["manipulative", "kilometres", "redistribute", "similarities"], ["quantifying", "foreseeable", "multiplied", "modalities"], ["favourable", "discerning", "sunglasses", "encountering"], ["intelligence", "prerequisites", "complained", "applicability"], ["favourable", "intonation", "epithelium", "predominantly"], ["quantifying", "kilometres", "complained", "abnormality"], ["intelligence", "discerning", "companions", "modalities"], ["persecuted", "watersheds", "complained", "abnormality"], ["manipulative", "watersheds", "pediatrics", "universality"], ["phonological", "unaffected", "sunglasses", "encountering"], ["phonological", "kilometres", "multiplied", "provincial"], ["diagnostics", "intonation", "complained", "similarities"], ["persecuted", "foreseeable", "epithelium", "applicability"], ["phonological", "discerning", "sunglasses", "universality"], ["quantifying", "foreseeable", "pediatrics", "universality"], ["manipulative", "prerequisites", "redistribute", "modalities"], ["contingency", "prerequisites", "pediatrics", "universality"], ["contingency", "dehydration", "redistribute", "universality"], ["persecuted", "discerning", "companions", "predominantly"], ["diagnostics", "foreseeable", "epithelium", "similarities"], ["persecuted", "watersheds", "redistribute", "abnormality"], ["persecuted", "intonation", "multiplied", "provincial"], ["phonological", "prerequisites", "redistribute", "predominantly"], ["favourable", "dehydration", "pediatrics", "provincial"], ["intelligence", "watersheds", "pediatrics", "abnormality"], ["diagnostics", "discerning", "complained", "similarities"], ["manipulative", "kilometres", "sunglasses", "applicability"], ["manipulative", "foreseeable", "disclosure", "predominantly"], ["quantifying", "dehydration", "pediatrics", "universality"], ["persecuted", "watersheds", "sunglasses", "provincial"], ["contingency", "watersheds", "pediatrics", "applicability"], ["phonological", "watersheds", "sunglasses", "predominantly"], ["favourable", "prerequisites", "epithelium", "abnormality"], ["manipulative", "dehydration", "multiplied", "provincial"], ["favourable", "prerequisites", "sunglasses", "provincial"], ["persecuted", "discerning", "multiplied", "predominantly"], ["phonological", "kilometres", "sunglasses", "similarities"], ["phonological", "dehydration", "multiplied", "encountering"], ["persecuted", "intonation", "pediatrics", "predominantly"], ["manipulative", "foreseeable", "multiplied", "applicability"], ["manipulative", "unaffected", "companions", "similarities"], ["phonological", "prerequisites", "complained", "encountering"], ["favourable", "dehydration", "complained", "encountering"], ["diagnostics", "foreseeable", "multiplied", "abnormality"], ["quantifying", "unaffected", "disclosure", "applicability"], ["diagnostics", "foreseeable", "sunglasses", "encountering"], ["manipulative", "dehydration", "disclosure", "modalities"], ["contingency", "prerequisites", "epithelium", "encountering"], ["intelligence", "watersheds", "pediatrics", "provincial"], ["persecuted", "dehydration", "complained", "universality"], ["contingency", "foreseeable", "epithelium", "similarities"], ["quantifying", "watersheds", "complained", "universality"], ["quantifying", "foreseeable", "multiplied", "predominantly"], ["persecuted", "foreseeable", "epithelium", "abnormality"], ["persecuted", "watersheds", "multiplied", "abnormality"], ["manipulative", "intonation", "redistribute", "abnormality"], ["persecuted", "prerequisites", "sunglasses", "predominantly"], ["phonological", "intonation", "sunglasses", "similarities"], ["diagnostics", "unaffected", "sunglasses", "universality"], ["persecuted", "unaffected", "epithelium", "applicability"], ["contingency", "kilometres", "redistribute", "modalities"], ["favourable", "prerequisites", "epithelium", "modalities"], ["persecuted", "prerequisites", "companions", "abnormality"], ["contingency", "dehydration", "companions", "abnormality"], ["manipulative", "discerning", "disclosure", "abnormality"], ["quantifying", "prerequisites", "redistribute", "applicability"], ["manipulative", "discerning", "redistribute", "predominantly"], ["persecuted", "discerning", "multiplied", "universality"], ["favourable", "dehydration", "complained", "encountering"], ["quantifying", "prerequisites", "disclosure", "universality"], ["diagnostics", "dehydration", "epithelium", "predominantly"], ["contingency", "unaffected", "multiplied", "abnormality"], ["persecuted", "unaffected", "sunglasses", "universality"], ["favourable", "kilometres", "sunglasses", "abnormality"], ["intelligence", "discerning", "multiplied", "abnormality"], ["quantifying", "unaffected", "pediatrics", "universality"], ["phonological", "foreseeable", "complained", "predominantly"], ["phonological", "unaffected", "multiplied", "applicability"], ["diagnostics", "watersheds", "companions", "predominantly"], ["persecuted", "discerning", "sunglasses", "encountering"], ["quantifying", "watersheds", "companions", "encountering"], ["quantifying", "watersheds", "epithelium", "predominantly"], ["manipulative", "discerning", "complained", "modalities"], ["intelligence", "foreseeable", "complained", "encountering"], ["contingency", "prerequisites", "redistribute", "predominantly"], ["intelligence", "kilometres", "multiplied", "applicability"], ["manipulative", "intonation", "companions", "abnormality"], ["persecuted", "prerequisites", "epithelium", "provincial"], ["phonological", "discerning", "complained", "provincial"], ["manipulative", "foreseeable", "multiplied", "abnormality"], ["persecuted", "kilometres", "complained", "abnormality"], ["contingency", "watersheds", "companions", "provincial"], ["quantifying", "prerequisites", "complained", "modalities"], ["intelligence", "watersheds", "complained", "modalities"], ["persecuted", "foreseeable", "multiplied", "applicability"], ["favourable", "foreseeable", "disclosure", "predominantly"], ["quantifying", "intonation", "sunglasses", "predominantly"], ["persecuted", "discerning", "epithelium", "encountering"], ["persecuted", "intonation", "multiplied", "modalities"], ["favourable", "unaffected", "companions", "predominantly"], ["diagnostics", "watersheds", "multiplied", "universality"], ["phonological", "kilometres", "multiplied", "similarities"], ["favourable", "dehydration", "pediatrics", "provincial"], ["diagnostics", "discerning", "redistribute", "applicability"], ["phonological", "dehydration", "redistribute", "applicability"], ["favourable", "prerequisites", "companions", "similarities"], ["quantifying", "intonation", "redistribute", "similarities"], ["intelligence", "intonation", "multiplied", "abnormality"], ["persecuted", "intonation", "epithelium", "predominantly"], ["contingency", "discerning", "multiplied", "provincial"], ["persecuted", "unaffected", "redistribute", "modalities"], ["persecuted", "discerning", "sunglasses", "abnormality"], ["phonological", "unaffected", "redistribute", "applicability"], ["contingency", "intonation", "sunglasses", "encountering"], ["quantifying", "unaffected", "redistribute", "predominantly"], ["persecuted", "intonation", "pediatrics", "modalities"], ["persecuted", "watersheds", "complained", "abnormality"], ["intelligence", "prerequisites", "multiplied", "encountering"], ["diagnostics", "dehydration", "companions", "universality"], ["favourable", "dehydration", "complained", "encountering"], ["persecuted", "intonation", "complained", "provincial"], ["quantifying", "intonation", "epithelium", "similarities"], ["manipulative", "dehydration", "sunglasses", "applicability"], ["intelligence", "dehydration", "redistribute", "abnormality"], ["diagnostics", "intonation", "multiplied", "encountering"], ["manipulative", "prerequisites", "multiplied", "modalities"], ["manipulative", "foreseeable", "pediatrics", "universality"], ["manipulative", "dehydration", "pediatrics", "modalities"], ["contingency", "intonation", "redistribute", "modalities"], ["manipulative", "prerequisites", "disclosure", "universality"], ["manipulative", "prerequisites", "sunglasses", "provincial"], ["favourable", "intonation", "sunglasses", "similarities"], ["manipulative", "unaffected", "companions", "universality"], ["quantifying", "discerning", "disclosure", "similarities"], ["contingency", "dehydration", "pediatrics", "similarities"], ["contingency", "intonation", "multiplied", "modalities"], ["manipulative", "discerning", "redistribute", "universality"], ["phonological", "kilometres", "redistribute", "encountering"], ["favourable", "kilometres", "companions", "modalities"], ["phonological", "dehydration", "pediatrics", "predominantly"], ["phonological", "unaffected", "disclosure", "applicability"], ["persecuted", "watersheds", "sunglasses", "abnormality"], ["diagnostics", "prerequisites", "complained", "provincial"], ["quantifying", "dehydration", "redistribute", "universality"], ["phonological", "intonation", "companions", "abnormality"], ["manipulative", "watersheds", "complained", "universality"], ["favourable", "intonation", "disclosure", "modalities"], ["contingency", "dehydration", "redistribute", "similarities"], ["phonological", "prerequisites", "multiplied", "applicability"], ["favourable", "discerning", "redistribute", "abnormality"], ["quantifying", "unaffected", "sunglasses", "applicability"], ["favourable", "unaffected", "companions", "predominantly"], ["intelligence", "unaffected", "disclosure", "universality"], ["manipulative", "watersheds", "pediatrics", "provincial"], ["persecuted", "foreseeable", "redistribute", "modalities"], ["favourable", "foreseeable", "complained", "universality"], ["persecuted", "prerequisites", "pediatrics", "applicability"], ["persecuted", "watersheds", "multiplied", "modalities"], ["diagnostics", "dehydration", "disclosure", "provincial"], ["phonological", "foreseeable", "pediatrics", "modalities"], ["manipulative", "intonation", "multiplied", "predominantly"], ["quantifying", "intonation", "pediatrics", "similarities"], ["quantifying", "unaffected", "epithelium", "encountering"], ["quantifying", "kilometres", "complained", "abnormality"], ["contingency", "prerequisites", "sunglasses", "modalities"], ["phonological", "kilometres", "companions", "modalities"], ["quantifying", "foreseeable", "epithelium", "similarities"], ["intelligence", "intonation", "companions", "applicability"], ["persecuted", "discerning", "epithelium", "similarities"], ["phonological", "dehydration", "epithelium", "abnormality"], ["diagnostics", "unaffected", "sunglasses", "applicability"], ["quantifying", "intonation", "epithelium", "modalities"], ["intelligence", "intonation", "redistribute", "similarities"], ["contingency", "unaffected", "sunglasses", "similarities"], ["diagnostics", "discerning", "disclosure", "applicability"], ["contingency", "discerning", "sunglasses", "provincial"], ["diagnostics", "discerning", "redistribute", "modalities"], ["persecuted", "unaffected", "pediatrics", "universality"], ["contingency", "discerning", "sunglasses", "predominantly"], ["manipulative", "prerequisites", "redistribute", "applicability"], ["contingency", "prerequisites", "sunglasses", "encountering"], ["quantifying", "kilometres", "pediatrics", "encountering"], ["persecuted", "prerequisites", "companions", "applicability"], ["persecuted", "dehydration", "disclosure", "predominantly"], ["persecuted", "discerning", "multiplied", "predominantly"], ["persecuted", "unaffected", "multiplied", "provincial"], ["manipulative", "dehydration", "epithelium", "modalities"], ["diagnostics", "foreseeable", "companions", "universality"], ["favourable", "foreseeable", "multiplied", "universality"], ["quantifying", "prerequisites", "redistribute", "similarities"]], "sealed_id": [["check", "machine", "pending", "today"], ["write", "machine", "pending", "now"], ["delete", "pump", "ready", "day"], ["update", "meeting", "blocked", "hour"], ["stop", "customer", "open", "now"], ["read", "order", "overload", "minute"], ["report", "file", "open", "hour"], ["write", "invoice", "failed", "year"], ["start", "motor", "paid", "week"], ["start", "job", "blocked", "day"], ["delete", "job", "low", "hour"], ["notify", "motor", "blocked", "month"], ["report", "email", "open", "hour"], ["delete", "pump", "active", "month"], ["check", "machine", "blocked", "hour"], ["query", "meeting", "blocked", "month"], ["execute", "job", "active", "today"], ["delete", "inventory", "active", "now"], ["notify", "meeting", "pending", "hour"], ["start", "valve", "normal", "month"], ["write", "order", "ready", "week"], ["execute", "meeting", "normal", "today"], ["notify", "valve", "normal", "now"], ["check", "email", "none", "week"], ["read", "file", "low", "today"], ["notify", "machine", "running", "today"], ["notify", "machine", "complete", "hour"], ["read", "inventory", "pending", "year"], ["notify", "order", "running", "month"], ["check", "email", "low", "hour"], ["create", "invoice", "active", "tomorrow"], ["execute", "machine", "closed", "hour"], ["write", "invoice", "open", "year"], ["create", "order", "closed", "none"], ["delete", "email", "idle", "today"], ["create", "motor", "failed", "year"], ["stop", "machine", "none", "day"], ["update", "valve", "high", "tomorrow"], ["check", "job", "running", "hour"], ["read", "machine", "open", "week"], ["stop", "pump", "active", "month"], ["check", "machine", "ready", "minute"], ["check", "valve", "blocked", "hour"], ["check", "order", "low", "week"], ["check", "meeting", "running", "none"], ["read", "valve", "pending", "today"], ["create", "machine", "ready", "today"], ["check", "machine", "closed", "today"], ["report", "machine", "overload", "hour"], ["delete", "meeting", "blocked", "now"], ["delete", "valve", "blocked", "year"], ["write", "customer", "ready", "tomorrow"], ["execute", "machine", "normal", "minute"], ["delete", "job", "complete", "today"], ["write", "meeting", "failed", "week"], ["update", "meeting", "idle", "minute"], ["write", "pump", "idle", "none"], ["delete", "order", "active", "tomorrow"], ["create", "pump", "paid", "year"], ["report", "order", "normal", "now"], ["notify", "motor", "ready", "year"], ["check", "inventory", "active", "week"], ["execute", "email", "blocked", "now"], ["stop", "machine", "blocked", "year"], ["start", "customer", "running", "now"], ["write", "machine", "pending", "tomorrow"], ["start", "machine", "open", "month"], ["create", "file", "active", "day"], ["execute", "job", "normal", "minute"], ["read", "valve", "running", "tomorrow"], ["execute", "motor", "low", "hour"], ["query", "job", "ready", "week"], ["delete", "file", "open", "minute"], ["delete", "order", "low", "week"], ["execute", "email", "idle", "year"], ["read", "machine", "active", "now"], ["update", "meeting", "pending", "today"], ["report", "job", "low", "minute"], ["stop", "customer", "failed", "minute"], ["create", "order", "high", "tomorrow"], ["stop", "customer", "closed", "now"], ["query", "inventory", "overload", "today"], ["check", "job", "blocked", "hour"], ["read", "machine", "complete", "today"], ["read", "meeting", "blocked", "now"], ["notify", "machine", "blocked", "today"], ["execute", "customer", "low", "hour"], ["update", "customer", "overload", "year"], ["delete", "pump", "failed", "now"], ["start", "order", "complete", "minute"], ["check", "order", "overload", "minute"], ["write", "machine", "blocked", "hour"], ["stop", "pump", "pending", "minute"], ["create", "valve", "idle", "day"], ["write", "motor", "failed", "day"], ["check", "meeting", "overload", "today"], ["create", "machine", "active", "now"], ["update", "pump", "ready", "none"], ["stop", "machine", "high", "now"], ["stop", "motor", "pending", "hour"], ["query", "order", "ready", "minute"], ["notify", "valve", "pending", "minute"], ["execute", "meeting", "idle", "now"], ["read", "order", "high", "tomorrow"], ["read", "customer", "normal", "hour"], ["check", "email", "complete", "minute"], ["report", "invoice", "idle", "hour"], ["execute", "customer", "blocked", "now"], ["check", "job", "complete", "none"], ["query", "order", "paid", "minute"], ["start", "motor", "running", "now"], ["check", "pump", "low", "week"], ["query", "inventory", "overload", "minute"], ["report", "job", "blocked", "day"], ["check", "order", "ready", "year"], ["notify", "order", "open", "hour"], ["stop", "motor", "normal", "today"], ["execute", "invoice", "failed", "tomorrow"], ["notify", "machine", "open", "minute"], ["write", "pump", "high", "month"], ["execute", "email", "none", "tomorrow"], ["query", "order", "normal", "none"], ["delete", "order", "ready", "minute"], ["query", "motor", "active", "hour"], ["check", "invoice", "ready", "week"], ["write", "job", "high", "none"], ["report", "email", "normal", "none"], ["start", "file", "running", "hour"], ["notify", "inventory", "failed", "year"], ["update", "meeting", "high", "year"], ["read", "order", "none", "today"], ["create", "email", "pending", "today"], ["notify", "inventory", "running", "hour"], ["check", "meeting", "closed", "today"], ["write", "motor", "high", "month"], ["read", "machine", "normal", "now"], ["write", "pump", "overload", "tomorrow"], ["start", "inventory", "idle", "hour"], ["execute", "meeting", "normal", "now"], ["notify", "meeting", "blocked", "minute"], ["query", "motor", "overload", "tomorrow"], ["check", "email", "none", "minute"], ["update", "pump", "high", "year"], ["stop", "pump", "high", "tomorrow"], ["delete", "meeting", "active", "week"], ["report", "meeting", "normal", "minute"], ["start", "pump", "low", "hour"], ["start", "job", "ready", "now"], ["execute", "email", "normal", "month"], ["delete", "inventory", "normal", "hour"], ["notify", "meeting", "ready", "today"], ["query", "invoice", "low", "today"], ["notify", "invoice", "paid", "tomorrow"], ["update", "machine", "active", "week"], ["stop", "machine", "overload", "day"], ["report", "file", "closed", "month"], ["write", "invoice", "pending", "today"], ["notify", "motor", "idle", "now"], ["create", "motor", "active", "tomorrow"], ["write", "inventory", "open", "today"], ["update", "email", "normal", "today"], ["stop", "pump", "paid", "none"], ["report", "pump", "ready", "day"], ["stop", "job", "overload", "year"], ["stop", "valve", "blocked", "month"], ["delete", "motor", "normal", "week"], ["delete", "email", "high", "minute"], ["check", "machine", "blocked", "week"], ["check", "customer", "paid", "none"], ["query", "meeting", "high", "tomorrow"], ["stop", "pump", "overload", "week"], ["delete", "email", "paid", "today"], ["start", "job", "low", "minute"], ["stop", "invoice", "complete", "hour"], ["report", "file", "normal", "tomorrow"], ["notify", "motor", "high", "day"], ["start", "inventory", "ready", "day"], ["stop", "machine", "closed", "year"], ["update", "invoice", "complete", "hour"], ["create", "pump", "closed", "tomorrow"], ["create", "inventory", "active", "none"], ["execute", "pump", "none", "now"], ["execute", "invoice", "ready", "day"], ["read", "file", "running", "minute"], ["write", "job", "complete", "month"], ["start", "valve", "paid", "week"], ["update", "machine", "ready", "now"], ["check", "customer", "idle", "none"], ["stop", "motor", "idle", "now"], ["create", "email", "closed", "none"], ["notify", "meeting", "pending", "tomorrow"], ["notify", "customer", "complete", "month"], ["create", "email", "blocked", "week"], ["stop", "machine", "blocked", "hour"], ["delete", "email", "normal", "none"], ["create", "valve", "overload", "none"], ["delete", "inventory", "blocked", "week"], ["execute", "meeting", "overload", "hour"], ["stop", "customer", "high", "tomorrow"], ["check", "machine", "overload", "tomorrow"]], "sealed_ood": [["start", "invoice", "open", "none"], ["update", "job", "failed", "hour"], ["stop", "customer", "active", "year"], ["start", "meeting", "low", "month"], ["stop", "inventory", "active", "minute"], ["read", "pump", "paid", "tomorrow"], ["execute", "order", "paid", "day"], ["delete", "machine", "overload", "minute"], ["read", "inventory", "normal", "day"], ["read", "invoice", "failed", "month"], ["stop", "inventory", "failed", "minute"], ["write", "pump", "high", "today"], ["execute", "inventory", "high", "today"], ["write", "meeting", "high", "hour"], ["query", "file", "closed", "week"], ["create", "pump", "pending", "week"], ["delete", "customer", "idle", "year"], ["check", "motor", "pending", "tomorrow"], ["stop", "inventory", "active", "now"], ["report", "customer", "running", "day"], ["write", "email", "ready", "now"], ["read", "pump", "paid", "none"], ["update", "inventory", "blocked", "week"], ["notify", "job", "running", "hour"], ["create", "valve", "complete", "year"], ["check", "file", "running", "hour"], ["query", "file", "normal", "today"], ["write", "valve", "blocked", "tomorrow"], ["write", "valve", "low", "now"], ["query", "order", "ready", "month"], ["update", "order", "paid", "tomorrow"], ["start", "invoice", "none", "today"], ["write", "pump", "none", "month"], ["execute", "order", "high", "today"], ["notify", "job", "ready", "tomorrow"], ["write", "meeting", "low", "day"], ["check", "customer", "normal", "day"], ["start", "meeting", "blocked", "hour"], ["report", "motor", "overload", "minute"], ["check", "machine", "high", "today"], ["check", "file", "idle", "tomorrow"], ["check", "file", "high", "year"], ["write", "inventory", "pending", "week"], ["start", "meeting", "low", "today"], ["query", "machine", "closed", "today"], ["write", "valve", "paid", "month"], ["query", "motor", "ready", "month"], ["write", "valve", "idle", "minute"], ["stop", "inventory", "pending", "year"], ["start", "invoice", "paid", "none"], ["write", "email", "closed", "day"], ["execute", "valve", "running", "month"], ["read", "file", "blocked", "tomorrow"], ["write", "valve", "low", "day"], ["delete", "valve", "idle", "tomorrow"], ["query", "file", "paid", "year"], ["stop", "email", "open", "none"], ["check", "customer", "ready", "hour"], ["write", "customer", "high", "hour"], ["report", "motor", "normal", "minute"], ["execute", "order", "running", "tomorrow"], ["notify", "motor", "none", "month"], ["read", "machine", "high", "hour"], ["notify", "order", "closed", "week"], ["report", "customer", "failed", "none"], ["update", "order", "none", "year"], ["execute", "customer", "complete", "year"], ["stop", "order", "high", "hour"], ["update", "pump", "running", "day"], ["check", "motor", "open", "week"], ["execute", "customer", "overload", "now"], ["create", "machine", "active", "minute"], ["execute", "meeting", "closed", "day"], ["read", "email", "complete", "year"], ["create", "job", "running", "now"], ["start", "meeting", "running", "month"], ["query", "file", "overload", "year"], ["stop", "machine", "open", "tomorrow"], ["delete", "machine", "idle", "now"], ["create", "meeting", "active", "week"], ["update", "order", "high", "hour"], ["start", "machine", "overload", "month"], ["check", "file", "low", "year"], ["check", "pump", "complete", "day"], ["update", "inventory", "idle", "year"], ["check", "motor", "open", "day"], ["check", "file", "idle", "week"], ["read", "pump", "idle", "today"], ["write", "email", "overload", "today"], ["delete", "machine", "running", "now"], ["stop", "email", "pending", "month"], ["notify", "invoice", "blocked", "none"], ["update", "valve", "paid", "month"], ["notify", "email", "blocked", "tomorrow"], ["update", "meeting", "normal", "year"], ["write", "inventory", "normal", "day"], ["delete", "customer", "ready", "month"], ["report", "pump", "none", "hour"], ["create", "file", "high", "hour"], ["report", "customer", "high", "day"], ["write", "pump", "paid", "month"], ["start", "email", "open", "tomorrow"], ["update", "inventory", "running", "month"], ["notify", "customer", "running", "day"], ["notify", "valve", "high", "hour"], ["execute", "invoice", "active", "year"], ["query", "motor", "open", "none"], ["start", "pump", "closed", "day"], ["write", "meeting", "idle", "tomorrow"], ["query", "file", "normal", "none"], ["write", "motor", "blocked", "none"], ["execute", "email", "active", "year"], ["update", "order", "pending", "minute"], ["report", "motor", "open", "month"], ["query", "pump", "open", "tomorrow"], ["update", "customer", "failed", "today"], ["stop", "meeting", "complete", "day"], ["check", "valve", "failed", "today"], ["notify", "customer", "complete", "day"], ["delete", "pump", "active", "minute"], ["delete", "invoice", "failed", "hour"], ["check", "file", "normal", "hour"], ["report", "invoice", "high", "hour"], ["check", "motor", "high", "minute"], ["execute", "valve", "running", "minute"], ["notify", "motor", "idle", "tomorrow"], ["execute", "order", "none", "now"], ["execute", "order", "closed", "year"], ["check", "motor", "normal", "year"], ["execute", "valve", "active", "none"], ["report", "motor", "closed", "minute"], ["notify", "job", "ready", "month"], ["report", "customer", "failed", "minute"], ["report", "order", "overload", "month"], ["write", "valve", "closed", "tomorrow"], ["query", "machine", "idle", "minute"], ["execute", "inventory", "paid", "hour"], ["execute", "pump", "low", "year"], ["read", "job", "running", "year"], ["read", "customer", "active", "minute"], ["read", "job", "complete", "day"], ["stop", "inventory", "overload", "now"], ["query", "pump", "failed", "hour"], ["report", "customer", "pending", "now"], ["query", "meeting", "complete", "year"], ["update", "inventory", "normal", "minute"], ["check", "valve", "failed", "hour"], ["notify", "inventory", "closed", "day"], ["update", "inventory", "none", "hour"], ["read", "order", "ready", "hour"], ["read", "file", "ready", "month"], ["query", "file", "idle", "day"], ["stop", "inventory", "paid", "week"], ["start", "email", "low", "day"], ["report", "job", "ready", "month"], ["delete", "customer", "closed", "none"], ["check", "file", "running", "minute"], ["create", "email", "closed", "day"], ["start", "pump", "pending", "week"], ["update", "order", "idle", "year"], ["delete", "machine", "high", "tomorrow"], ["report", "customer", "paid", "today"], ["read", "email", "closed", "day"], ["check", "motor", "running", "week"], ["create", "job", "ready", "month"], ["delete", "machine", "blocked", "minute"], ["start", "customer", "paid", "hour"], ["notify", "pump", "ready", "tomorrow"], ["delete", "meeting", "blocked", "tomorrow"], ["query", "invoice", "failed", "today"], ["create", "job", "ready", "none"], ["read", "invoice", "paid", "minute"], ["read", "email", "paid", "month"], ["report", "customer", "overload", "week"], ["start", "invoice", "blocked", "minute"], ["execute", "valve", "paid", "month"], ["delete", "pump", "none", "month"], ["report", "invoice", "open", "tomorrow"], ["read", "invoice", "active", "minute"], ["delete", "machine", "blocked", "today"], ["check", "file", "open", "hour"], ["start", "invoice", "overload", "week"], ["write", "order", "open", "tomorrow"], ["check", "file", "normal", "day"], ["stop", "invoice", "low", "day"], ["start", "invoice", "idle", "minute"], ["start", "order", "paid", "month"], ["report", "customer", "low", "year"], ["stop", "motor", "idle", "tomorrow"], ["execute", "pump", "idle", "tomorrow"], ["execute", "order", "running", "year"], ["read", "valve", "high", "today"], ["write", "email", "high", "day"], ["notify", "pump", "pending", "week"], ["create", "motor", "normal", "year"], ["check", "file", "normal", "week"], ["update", "job", "low", "year"], ["read", "job", "high", "today"], ["write", "valve", "pending", "month"], ["write", "email", "paid", "month"], ["start", "meeting", "none", "year"], ["notify", "job", "none", "hour"], ["update", "valve", "blocked", "none"], ["execute", "order", "closed", "hour"], ["read", "pump", "blocked", "year"], ["check", "meeting", "blocked", "none"], ["read", "invoice", "open", "day"], ["delete", "machine", "overload", "hour"], ["start", "invoice", "low", "today"], ["query", "pump", "open", "none"], ["create", "job", "normal", "tomorrow"], ["delete", "inventory", "high", "today"], ["start", "invoice", "open", "hour"], ["notify", "pump", "low", "week"], ["stop", "inventory", "normal", "month"], ["stop", "inventory", "active", "none"], ["delete", "machine", "ready", "tomorrow"], ["execute", "customer", "low", "year"], ["delete", "machine", "active", "hour"], ["query", "machine", "normal", "day"], ["stop", "email", "running", "tomorrow"], ["create", "job", "paid", "tomorrow"], ["check", "meeting", "idle", "tomorrow"], ["notify", "pump", "running", "today"], ["write", "valve", "paid", "week"], ["report", "customer", "failed", "today"], ["report", "inventory", "normal", "year"], ["check", "valve", "none", "hour"], ["stop", "inventory", "pending", "today"], ["report", "motor", "active", "year"], ["stop", "email", "failed", "week"], ["stop", "inventory", "closed", "now"], ["write", "motor", "active", "year"], ["check", "job", "open", "tomorrow"], ["start", "valve", "high", "hour"], ["write", "email", "closed", "year"], ["create", "invoice", "failed", "today"], ["report", "customer", "complete", "year"], ["report", "machine", "low", "year"], ["write", "email", "normal", "year"], ["notify", "job", "none", "today"], ["check", "file", "active", "year"], ["create", "meeting", "closed", "year"], ["update", "inventory", "ready", "week"], ["delete", "machine", "complete", "tomorrow"], ["write", "file", "high", "today"], ["check", "file", "normal", "year"], ["start", "pump", "blocked", "none"], ["delete", "job", "normal", "year"], ["execute", "valve", "overload", "today"], ["execute", "order", "paid", "none"], ["read", "inventory", "blocked", "tomorrow"], ["execute", "valve", "normal", "hour"], ["execute", "customer", "closed", "day"], ["stop", "inventory", "closed", "month"], ["create", "order", "normal", "day"], ["check", "file", "overload", "month"], ["query", "file", "low", "hour"], ["delete", "machine", "paid", "year"], ["write", "meeting", "ready", "month"], ["stop", "inventory", "overload", "hour"], ["read", "valve", "overload", "now"], ["stop", "inventory", "normal", "minute"], ["report", "customer", "ready", "week"], ["write", "valve", "none", "month"], ["query", "machine", "ready", "now"], ["read", "file", "running", "day"], ["create", "meeting", "ready", "day"], ["delete", "machine", "idle", "day"], ["notify", "motor", "open", "tomorrow"], ["check", "order", "blocked", "tomorrow"], ["delete", "invoice", "none", "hour"], ["notify", "job", "pending", "tomorrow"], ["check", "motor", "blocked", "today"], ["update", "job", "idle", "tomorrow"], ["create", "valve", "failed", "hour"], ["update", "pump", "open", "none"], ["write", "customer", "closed", "day"], ["update", "meeting", "active", "year"], ["query", "pump", "low", "day"], ["execute", "valve", "failed", "week"], ["query", "machine", "running", "today"], ["start", "customer", "active", "year"], ["delete", "order", "overload", "month"], ["notify", "job", "none", "year"], ["notify", "order", "pending", "week"], ["stop", "email", "complete", "day"], ["read", "inventory", "overload", "now"], ["read", "invoice", "failed", "year"], ["delete", "customer", "pending", "today"], ["stop", "order", "overload", "month"], ["start", "motor", "overload", "now"], ["report", "job", "running", "year"], ["update", "inventory", "failed", "hour"], ["create", "meeting", "closed", "now"], ["update", "order", "running", "day"], ["write", "file", "open", "none"], ["execute", "pump", "complete", "day"], ["read", "invoice", "paid", "none"], ["read", "valve", "open", "none"], ["report", "customer", "low", "today"], ["check", "pump", "paid", "month"], ["write", "email", "none", "today"], ["update", "file", "ready", "month"], ["write", "valve", "complete", "none"], ["read", "pump", "running", "minute"], ["read", "invoice", "paid", "year"], ["read", "job", "running", "day"], ["update", "email", "failed", "hour"], ["write", "order", "ready", "hour"], ["read", "order", "pending", "week"], ["start", "pump", "active", "year"], ["check", "order", "normal", "day"], ["read", "motor", "ready", "month"], ["report", "order", "open", "none"], ["report", "customer", "closed", "hour"], ["stop", "email", "ready", "today"], ["delete", "job", "blocked", "tomorrow"], ["create", "valve", "low", "day"], ["delete", "machine", "ready", "minute"], ["write", "email", "paid", "hour"], ["start", "meeting", "overload", "today"], ["update", "job", "closed", "week"], ["create", "valve", "open", "none"], ["execute", "machine", "active", "minute"], ["update", "valve", "none", "month"], ["report", "motor", "active", "day"], ["delete", "machine", "open", "tomorrow"], ["stop", "inventory", "high", "none"], ["check", "inventory", "ready", "hour"], ["write", "meeting", "pending", "day"], ["check", "valve", "high", "today"], ["report", "motor", "paid", "hour"], ["report", "customer", "pending", "week"], ["query", "machine", "overload", "hour"], ["read", "email", "running", "year"], ["report", "motor", "active", "none"], ["query", "file", "normal", "week"], ["create", "meeting", "none", "tomorrow"], ["start", "email", "none", "hour"], ["check", "file", "closed", "now"], ["delete", "file", "idle", "tomorrow"], ["check", "machine", "idle", "tomorrow"], ["report", "customer", "high", "today"], ["notify", "pump", "overload", "tomorrow"], ["start", "pump", "idle", "tomorrow"], ["check", "file", "active", "tomorrow"], ["check", "meeting", "high", "hour"], ["execute", "email", "ready", "month"], ["check", "motor", "running", "today"], ["check", "valve", "blocked", "none"], ["query", "valve", "running", "year"], ["write", "email", "active", "month"], ["report", "pump", "overload", "month"], ["write", "valve", "high", "week"], ["check", "motor", "running", "month"], ["report", "motor", "high", "hour"], ["read", "pump", "overload", "year"], ["create", "motor", "low", "day"], ["start", "customer", "idle", "tomorrow"], ["read", "invoice", "paid", "hour"], ["delete", "customer", "failed", "week"], ["read", "valve", "none", "hour"], ["create", "file", "open", "tomorrow"], ["read", "invoice", "closed", "now"], ["notify", "invoice", "complete", "year"], ["delete", "customer", "paid", "week"], ["stop", "machine", "paid", "month"], ["check", "inventory", "failed", "today"], ["create", "job", "active", "month"], ["report", "motor", "idle", "today"], ["read", "customer", "complete", "day"], ["check", "file", "blocked", "now"], ["notify", "machine", "high", "today"], ["read", "invoice", "none", "minute"], ["stop", "email", "closed", "none"], ["notify", "job", "running", "tomorrow"], ["write", "valve", "failed", "none"], ["notify", "pump", "overload", "none"], ["read", "motor", "high", "today"], ["check", "invoice", "idle", "tomorrow"], ["update", "customer", "normal", "year"], ["query", "machine", "high", "now"], ["delete", "machine", "ready", "day"], ["create", "job", "pending", "today"], ["delete", "job", "failed", "hour"], ["create", "order", "paid", "hour"], ["write", "invoice", "overload", "now"], ["write", "valve", "active", "year"], ["report", "inventory", "low", "year"], ["stop", "motor", "none", "month"], ["stop", "email", "low", "minute"], ["check", "file", "pending", "day"], ["check", "file", "open", "none"], ["start", "file", "low", "year"], ["report", "motor", "failed", "today"], ["query", "motor", "failed", "hour"], ["start", "file", "open", "tomorrow"], ["read", "email", "blocked", "tomorrow"], ["check", "motor", "running", "none"]], "seed": 20261016}
021/c021.py ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # CUERDAS-021 common: sealed set, varied replacement pool, training loop, evaluation.
2
+ import os, sys, json, hashlib, random, re, time
3
+ import torch, torch.nn as nn, torch.nn.functional as F
4
+ sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
5
+ import run018 as R
6
+ INP = os.environ['INP']; OD = os.environ['OD']
7
+ raw = open(f'{INP}/sealed021.json').read(); H = hashlib.sha256(raw.encode()).hexdigest()
8
+ assert H == 'a86a6a5ee375ae67a3e79f969c9553b0e5548ed3b0e80e5315862e6552ae72bc', H
9
+ S = json.loads(raw); NEW = S['new_values']; EXCL = set(S['exclude_token_ids'])
10
+ SETS = ['sealed_id', 'sealed_ood', 'sealed_1_new', 'sealed_2_new', 'sealed_4_new']
11
+ SMOKE = bool(os.environ.get('SMOKE'))
12
+ if SMOKE: # dry run on dev messages only; sealed sets are never touched
13
+ for s_ in SETS: S[s_] = R.FRESH['ood'][:6]
14
+ STEPS = 4 if SMOKE else 3000; EVERY = 2 if SMOKE else 250
15
+ t1 = R.t1; tok = R.tok; Emb = R.Emb; D = R.D; FIELDS = R.FIELDS; V = R.V
16
+ CAND = [torch.tensor(sorted({t1(w)[0] for w in V[j] + NEW[j]})) for j in range(4)]
17
+ assert all(len(CAND[j]) == len(V[j]) + 8 for j in range(4))
18
+ def gidx(cand, j, w): return int((cand[j] == t1(w)[0]).nonzero()[0])
19
+ # ---- varied pool (training) and held-out validation pool; all disjoint from sealed tokens and from 014/dev/020 values ----
20
+ voc = tok.get_vocab(); BAN = EXCL | R.BANNED
21
+ def ok_tok(t): return voc[t] not in BAN
22
+ WORDS = sorted(t[1:] for t in voc if re.fullmatch(r'Ġ[a-z]{3,14}', t) and ok_tok(t))
23
+ NAMES = sorted(t[1:] for t in voc if re.fullmatch(r'Ġ[A-Z][a-z]{2,10}', t) and ok_tok(t))
24
+ SUBS = sorted(t for t in voc if re.fullmatch(r'[a-z]{2,5}', t))
25
+ g0 = random.Random(2101); g0.shuffle(WORDS); g0.shuffle(NAMES)
26
+ VW, VN = WORDS[-300:], NAMES[-300:]; WORDS, NAMES = WORDS[:-300], NAMES[:-300]
27
+ def sample_value(g, W_=None, N_=None):
28
+ W_ = W_ or WORDS; N_ = N_ or NAMES; k = g.randrange(5)
29
+ if k == 0: return g.choice(W_)
30
+ if k == 1: return g.choice(N_)
31
+ if k == 2: return str(g.randrange(10000))
32
+ if k == 3: return ' '.join(g.choice(W_) for _ in range(g.choice((2, 3))))
33
+ return g.choice(W_) + g.choice(SUBS)
34
+ def sample_msgs(g, bs=12):
35
+ out = []
36
+ for _ in range(bs):
37
+ m = list(g.choice(R.FIT))
38
+ for j in range(4):
39
+ if g.random() < .4: m[j] = sample_value(g)
40
+ out.append(m)
41
+ return out
42
+ # restricted transfer validation: 50 held-out values per field (no numbers: first token not unique), options = field vocab + those
43
+ gv = random.Random(2102); CANDV = []; VPOOL = []
44
+ for j in range(4):
45
+ used = {t1(w)[0] for w in V[j]}; ws = []
46
+ while len(ws) < 50:
47
+ k = gv.randrange(3); w = gv.choice(VW) if k == 0 else gv.choice(VN) if k == 1 else gv.choice(VW) + ' ' + gv.choice(VW)
48
+ if t1(w)[0] not in used: used.add(t1(w)[0]); ws.append(w)
49
+ VPOOL.append(ws); CANDV.append(torch.tensor(sorted(used)))
50
+ VAL_TR = [[gv.choice(VPOOL[j]) for j in range(4)] for _ in range(96)]
51
+ VALC = R.VAL[:96]
52
+ if SMOKE: VALC, VAL_TR = VALC[:6], VAL_TR[:6]
53
+ # ---- generic readout: rows = prefix (embeddings) + body tokens ----
54
+ def run_h(embs):
55
+ L = max(e.shape[0] for e in embs); B = len(embs); x = torch.zeros(B, L, D); m = torch.zeros(B, L, dtype=torch.long)
56
+ for i, e in enumerate(embs): x[i, :e.shape[0]] = e; m[i, :e.shape[0]] = 1
57
+ h = R.lm.model(inputs_embeds=x, attention_mask=m).last_hidden_state[torch.arange(B), m.sum(1) - 1]
58
+ return h, R.lm.lm_head(h)
59
+ def k2_rows(cod, msgs, tpl):
60
+ Sl = cod(torch.stack([torch.stack([R.lex(w) for w in m]) for m in msgs]))
61
+ return [torch.cat([Sl[i], Emb[R.sufids(tpl, j)]]) for i in range(len(msgs)) for j in range(4)]
62
+ def text_rows(msgs, tpl, fmt, prefix=None):
63
+ rows = []
64
+ for m in msgs:
65
+ for j in range(4):
66
+ e = Emb[tok.encode(fmt.format(**dict(zip(FIELDS, m))) + tpl.format(field=FIELDS[j]), add_special_tokens=False)]
67
+ rows.append(e if prefix is None else torch.cat([prefix, e]))
68
+ return rows
69
+ @torch.no_grad()
70
+ def per_ex(rowfn, msgs, cand, bs=24):
71
+ out = []
72
+ for b in range(0, len(msgs), bs):
73
+ mb = msgs[b:b+bs]; _, lg = run_h(rowfn(mb)); good = [True] * len(mb)
74
+ for r_ in range(len(mb) * 4):
75
+ i, j = divmod(r_, 4); good[i] &= int(lg[r_][cand[j]].argmax()) == gidx(cand, j, mb[i][j])
76
+ out += good
77
+ return [float(x) for x in out]
78
+ def val_score(rowfn_main):
79
+ return sum(per_ex(rowfn_main, VALC, R.CAND)) / len(VALC) + sum(per_ex(rowfn_main, VAL_TR, CANDV)) / len(VAL_TR)
80
+ def train(params, rowfn, teach, seed, tag, log, steps=None, lr=5e-4, every=None, set_mode=lambda t: None):
81
+ """Same loss/schedule/selection as 018 distill2: CE + KL(teacher) + 0.5(1-cos) ; OneCycle; select on validation only."""
82
+ steps = steps or STEPS; every = every or EVERY
83
+ g = random.Random(seed + 2100); opt = torch.optim.AdamW(params, lr=lr, weight_decay=1e-3)
84
+ sch = torch.optim.lr_scheduler.OneCycleLR(opt, max_lr=lr, total_steps=steps, pct_start=.05)
85
+ def snap(): return [p.detach().clone() for p in params]
86
+ set_mode(False); v = val_score(lambda mb: rowfn(mb, R.SUF_MAIN)); set_mode(True); best = (v, snap(), 0); log({'run': tag, 'step': 0, 'val': v})
87
+ t0 = time.time()
88
+ for s_ in range(1, steps + 1):
89
+ msgs = sample_msgs(g); tpl = g.choice(R.SUF_TRAIN)
90
+ with torch.no_grad(): ht, lt = R.readout_h(teach, msgs, tpl)
91
+ hs, ls = run_h(rowfn(msgs, tpl)); tg = torch.tensor([t1(m[j])[0] for m in msgs for j in range(4)])
92
+ loss = F.cross_entropy(ls, tg) + F.kl_div(F.log_softmax(ls, -1), F.log_softmax(lt, -1), log_target=True, reduction='batchmean') + .5 * (1 - F.cosine_similarity(hs, ht, -1).mean())
93
+ opt.zero_grad(); loss.backward(); torch.nn.utils.clip_grad_norm_(params, 1.0); opt.step(); sch.step()
94
+ if s_ % every == 0:
95
+ set_mode(False); v = val_score(lambda mb: rowfn(mb, R.SUF_MAIN)); set_mode(True)
96
+ log({'run': tag, 'step': s_, 'val': v, 'loss': float(loss), 's': round(time.time() - t0)})
97
+ if v > best[0]: best = (v, snap(), s_)
98
+ with torch.no_grad():
99
+ for p, b in zip(params, best[1]): p.copy_(b)
100
+ set_mode(False); return best[2], best[0]
101
+ def load_teacher(sd):
102
+ t = R.Binder(sd); t.load_state_dict(torch.load(f'{INP}/binder4_1000_s{sd}.pt', map_location=R.DEV)); return t.eval().requires_grad_(False)
103
+ def opener(name):
104
+ f = open(f'{OD}/{name}', 'w')
105
+ def log(d): f.write(json.dumps(d) + '\n'); f.flush(); print(json.dumps(d), flush=True)
106
+ return log
021/cuerdas021-a.log ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [{"stream_name":"stdout","time":6.186180912,"data":" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 44.1/44.1 kB 1.4 MB/s eta 0:00:00\n"}
2
+ ,{"stream_name":"stdout","time":7.168235528,"data":" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 10.0/10.0 MB 81.7 MB/s eta 0:00:00\n"}
3
+ ,{"stream_name":"stdout","time":7.200287347,"data":" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 566.4/566.4 kB 29.8 MB/s eta 0:00:00\n"}
4
+ ,{"stream_name":"stdout","time":7.269625736,"data":" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 3.0/3.0 MB 53.7 MB/s eta 0:00:00\n"}
5
+ ,{"stream_name":"stderr","time":21.344989509,"data":"ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.\n"}
6
+ ,{"stream_name":"stderr","time":21.345019411,"data":"litellm 1.85.7 requires tokenizers\u003c1.0,\u003e=0.21.0, but you have tokenizers 0.20.3 which is incompatible.\n"}
7
+ ,{"stream_name":"stderr","time":21.345025011,"data":"gradio 6.26.0 requires huggingface-hub\u003c2.0,\u003e=1.16.0, but you have huggingface-hub 0.36.2 which is incompatible.\n"}
8
+ ,{"stream_name":"stderr","time":21.345030991,"data":"diffusers 0.40.0 requires huggingface-hub\u003c2.0,\u003e=1.23.0, but you have huggingface-hub 0.36.2 which is incompatible.\n"}
9
+ ,{"stream_name":"stdout","time":75.159877172,"data":"{\"sealed_sha256\": \"a86a6a5ee375ae67a3e79f969c9553b0e5548ed3b0e80e5315862e6552ae72bc\", \"torch\": \"2.11.0+cu128\", \"pool_sizes\": {\"words\": 20654, \"names\": 8188, \"subs\": 7189}}\n"}
10
+ ,{"stream_name":"stdout","time":88.355573409,"data":"{\"sys\": \"text_template\", \"sealed_id\": 1.0, \"sealed_ood\": 0.985, \"sealed_1_new\": 0.96875, \"sealed_2_new\": 0.9739583333333334, \"sealed_4_new\": 0.984375}\n"}
11
+ ,{"stream_name":"stdout","time":95.592428269,"data":"{\"sys\": \"k2_018_s17\", \"sealed_id\": 1.0, \"sealed_ood\": 0.99, \"sealed_1_new\": 0.875, \"sealed_2_new\": 0.625, \"sealed_4_new\": 0.2864583333333333}\n"}
12
+ ,{"stream_name":"stdout","time":102.609933367,"data":"{\"sys\": \"k2_018_s29\", \"sealed_id\": 0.995, \"sealed_ood\": 0.99, \"sealed_1_new\": 0.875, \"sealed_2_new\": 0.578125, \"sealed_4_new\": 0.265625}\n"}
13
+ ,{"stream_name":"stdout","time":109.535114583,"data":"{\"sys\": \"k2_018_s43\", \"sealed_id\": 0.975, \"sealed_ood\": 0.9775, \"sealed_1_new\": 0.7552083333333334, \"sealed_2_new\": 0.4583333333333333, \"sealed_4_new\": 0.25}\n"}
14
+ ,{"stream_name":"stdout","time":110.769495618,"data":"{\"run\": \"k2_021_s17\", \"step\": 0, \"val\": 1.1145833333333333}\n"}
15
+ ,{"stream_name":"stderr","time":161.364732013,"data":"/usr/local/lib/python3.13/dist-packages/torch/utils/_device.py:116: UserWarning: Converting a tensor with requires_grad=True to a scalar may lead to unexpected behavior.\n"}
16
+ ,{"stream_name":"stderr","time":161.364784997,"data":"Consider using tensor.detach() first. (Triggered internally at /pytorch/torch/csrc/autograd/generated/python_variable_methods.cpp:836.)\n"}
17
+ ,{"stream_name":"stderr","time":161.364795195,"data":" return func(*args, **kwargs)\n"}
18
+ ,{"stream_name":"stdout","time":161.364818194,"data":"{\"run\": \"k2_021_s17\", \"step\": 250, \"val\": 1.0833333333333333, \"loss\": 4.719815254211426, \"s\": 51}\n"}
19
+ ,{"stream_name":"stdout","time":212.200652872,"data":"{\"run\": \"k2_021_s17\", \"step\": 500, \"val\": 1.0833333333333333, \"loss\": 5.05856990814209, \"s\": 101}\n"}
20
+ ,{"stream_name":"stdout","time":264.450103516,"data":"{\"run\": \"k2_021_s17\", \"step\": 750, \"val\": 1.0625, \"loss\": 4.387829780578613, \"s\": 154}\n"}
21
+ ,{"stream_name":"stdout","time":316.827434258,"data":"{\"run\": \"k2_021_s17\", \"step\": 1000, \"val\": 1.09375, \"loss\": 4.11569881439209, \"s\": 206}\n"}
22
+ ,{"stream_name":"stdout","time":369.502860696,"data":"{\"run\": \"k2_021_s17\", \"step\": 1250, \"val\": 1.1354166666666667, \"loss\": 4.992859363555908, \"s\": 259}\n"}
23
+ ,{"stream_name":"stdout","time":422.265248199,"data":"{\"run\": \"k2_021_s17\", \"step\": 1500, \"val\": 1.1458333333333335, \"loss\": 4.392324447631836, \"s\": 311}\n"}
24
+ ,{"stream_name":"stdout","time":474.904848973,"data":"{\"run\": \"k2_021_s17\", \"step\": 1750, \"val\": 1.1354166666666667, \"loss\": 4.7229413986206055, \"s\": 364}\n"}
25
+ ,{"stream_name":"stdout","time":527.486876182,"data":"{\"run\": \"k2_021_s17\", \"step\": 2000, \"val\": 1.1458333333333333, \"loss\": 4.124819278717041, \"s\": 417}\n"}
26
+ ,{"stream_name":"stdout","time":579.923370702,"data":"{\"run\": \"k2_021_s17\", \"step\": 2250, \"val\": 1.1666666666666667, \"loss\": 3.794137477874756, \"s\": 469}\n"}
27
+ ,{"stream_name":"stdout","time":631.86940922,"data":"{\"run\": \"k2_021_s17\", \"step\": 2500, \"val\": 1.1666666666666667, \"loss\": 2.527881622314453, \"s\": 521}\n"}
28
+ ,{"stream_name":"stdout","time":684.082385,"data":"{\"run\": \"k2_021_s17\", \"step\": 2750, \"val\": 1.1770833333333333, \"loss\": 3.1677093505859375, \"s\": 573}\n"}
29
+ ,{"stream_name":"stdout","time":736.534519603,"data":"{\"run\": \"k2_021_s17\", \"step\": 3000, \"val\": 1.1666666666666667, \"loss\": 1.943210482597351, \"s\": 626}\n"}
30
+ ,{"stream_name":"stdout","time":736.561427792,"data":"{\"selected\": \"k2_021_s17\", \"step\": 2750, \"val\": 1.1770833333333333}\n"}
31
+ ,{"stream_name":"stdout","time":744.219044901,"data":"{\"sys\": \"k2_021_s17\", \"sealed_id\": 1.0, \"sealed_ood\": 0.9975, \"sealed_1_new\": 0.90625, \"sealed_2_new\": 0.6302083333333334, \"sealed_4_new\": 0.3125}\n"}
32
+ ,{"stream_name":"stdout","time":745.734205364,"data":"{\"run\": \"k2_021_s29\", \"step\": 0, \"val\": 1.0729166666666667}\n"}
33
+ ,{"stream_name":"stdout","time":797.886349926,"data":"{\"run\": \"k2_021_s29\", \"step\": 250, \"val\": 1.0520833333333335, \"loss\": 4.769656181335449, \"s\": 52}\n"}
34
+ ,{"stream_name":"stdout","time":850.649595978,"data":"{\"run\": \"k2_021_s29\", \"step\": 500, \"val\": 1.0833333333333333, \"loss\": 4.504295825958252, \"s\": 105}\n"}
35
+ ,{"stream_name":"stdout","time":902.865580802,"data":"{\"run\": \"k2_021_s29\", \"step\": 750, \"val\": 1.15625, \"loss\": 3.0690648555755615, \"s\": 157}\n"}
36
+ ,{"stream_name":"stdout","time":955.496481741,"data":"{\"run\": \"k2_021_s29\", \"step\": 1000, \"val\": 1.125, \"loss\": 2.2323856353759766, \"s\": 210}\n"}
37
+ ,{"stream_name":"stdout","time":1008.341980537,"data":"{\"run\": \"k2_021_s29\", \"step\": 1250, \"val\": 1.1666666666666667, \"loss\": 4.31823205947876, \"s\": 263}\n"}
38
+ ,{"stream_name":"stdout","time":1060.535361573,"data":"{\"run\": \"k2_021_s29\", \"step\": 1500, \"val\": 1.1770833333333333, \"loss\": 4.3195061683654785, \"s\": 315}\n"}
39
+ ,{"stream_name":"stdout","time":1112.771534115,"data":"{\"run\": \"k2_021_s29\", \"step\": 1750, \"val\": 1.1354166666666667, \"loss\": 3.4602582454681396, \"s\": 367}\n"}
40
+ ,{"stream_name":"stdout","time":1164.997576792,"data":"{\"run\": \"k2_021_s29\", \"step\": 2000, \"val\": 1.1041666666666667, \"loss\": 3.7912955284118652, \"s\": 419}\n"}
41
+ ,{"stream_name":"stdout","time":1217.644480861,"data":"{\"run\": \"k2_021_s29\", \"step\": 2250, \"val\": 1.09375, \"loss\": 3.81158709526062, \"s\": 472}\n"}
42
+ ,{"stream_name":"stdout","time":1270.042174474,"data":"{\"run\": \"k2_021_s29\", \"step\": 2500, \"val\": 1.1354166666666667, \"loss\": 2.8178346157073975, \"s\": 524}\n"}
43
+ ,{"stream_name":"stdout","time":1322.556426989,"data":"{\"run\": \"k2_021_s29\", \"step\": 2750, \"val\": 1.1145833333333333, \"loss\": 3.228869676589966, \"s\": 577}\n"}
44
+ ,{"stream_name":"stdout","time":1374.978668944,"data":"{\"run\": \"k2_021_s29\", \"step\": 3000, \"val\": 1.1145833333333333, \"loss\": 3.688642978668213, \"s\": 629}\n"}
45
+ ,{"stream_name":"stdout","time":1374.998128133,"data":"{\"selected\": \"k2_021_s29\", \"step\": 1500, \"val\": 1.1770833333333333}\n"}
46
+ ,{"stream_name":"stdout","time":1382.613613675,"data":"{\"sys\": \"k2_021_s29\", \"sealed_id\": 0.99, \"sealed_ood\": 0.985, \"sealed_1_new\": 0.8854166666666666, \"sealed_2_new\": 0.5677083333333334, \"sealed_4_new\": 0.3020833333333333}\n"}
47
+ ,{"stream_name":"stdout","time":1384.182599023,"data":"{\"run\": \"k2_021_s43\", \"step\": 0, \"val\": 1.1041666666666667}\n"}
48
+ ,{"stream_name":"stdout","time":1436.533073168,"data":"{\"run\": \"k2_021_s43\", \"step\": 250, \"val\": 1.0625, \"loss\": 3.4185612201690674, \"s\": 52}\n"}
49
+ ,{"stream_name":"stdout","time":1488.680435255,"data":"{\"run\": \"k2_021_s43\", \"step\": 500, \"val\": 1.0, \"loss\": 4.755449295043945, \"s\": 104}\n"}
50
+ ,{"stream_name":"stdout","time":1541.653197742,"data":"{\"run\": \"k2_021_s43\", \"step\": 750, \"val\": 1.0520833333333335, \"loss\": 4.684804916381836, \"s\": 157}\n"}
51
+ ,{"stream_name":"stdout","time":1594.18542137,"data":"{\"run\": \"k2_021_s43\", \"step\": 1000, \"val\": 1.09375, \"loss\": 4.735645771026611, \"s\": 210}\n"}
52
+ ,{"stream_name":"stdout","time":1647.036378086,"data":"{\"run\": \"k2_021_s43\", \"step\": 1250, \"val\": 1.1458333333333333, \"loss\": 3.7620046138763428, \"s\": 263}\n"}
53
+ ,{"stream_name":"stdout","time":1699.512461238,"data":"{\"run\": \"k2_021_s43\", \"step\": 1500, \"val\": 1.1354166666666667, \"loss\": 4.758123397827148, \"s\": 315}\n"}
54
+ ,{"stream_name":"stdout","time":1751.688814083,"data":"{\"run\": \"k2_021_s43\", \"step\": 1750, \"val\": 1.1145833333333333, \"loss\": 4.265966892242432, \"s\": 368}\n"}
55
+ ,{"stream_name":"stdout","time":1803.454996524,"data":"{\"run\": \"k2_021_s43\", \"step\": 2000, \"val\": 1.0625, \"loss\": 3.849872589111328, \"s\": 419}\n"}
56
+ ,{"stream_name":"stdout","time":1855.058249654,"data":"{\"run\": \"k2_021_s43\", \"step\": 2250, \"val\": 1.1041666666666665, \"loss\": 4.972720146179199, \"s\": 471}\n"}
57
+ ,{"stream_name":"stdout","time":1907.043534368,"data":"{\"run\": \"k2_021_s43\", \"step\": 2500, \"val\": 1.125, \"loss\": 2.9099667072296143, \"s\": 523}\n"}
58
+ ,{"stream_name":"stdout","time":1958.882514238,"data":"{\"run\": \"k2_021_s43\", \"step\": 2750, \"val\": 1.125, \"loss\": 3.000356674194336, \"s\": 575}\n"}
59
+ ,{"stream_name":"stdout","time":2010.517405519,"data":"{\"run\": \"k2_021_s43\", \"step\": 3000, \"val\": 1.125, \"loss\": 2.915379762649536, \"s\": 626}\n"}
60
+ ,{"stream_name":"stdout","time":2010.535140228,"data":"{\"selected\": \"k2_021_s43\", \"step\": 1250, \"val\": 1.1458333333333333}\n"}
61
+ ,{"stream_name":"stdout","time":2018.04167291,"data":"{\"sys\": \"k2_021_s43\", \"sealed_id\": 0.995, \"sealed_ood\": 0.9775, \"sealed_1_new\": 0.6614583333333334, \"sealed_2_new\": 0.3802083333333333, \"sealed_4_new\": 0.265625}\n"}
62
+ ,{"stream_name":"stdout","time":2075.919319203,"data":"{\"probe\": \"ridge_aug\", \"seed\": 17, \"set\": \"sealed_ood\", \"acc\": 1.0}\n"}
63
+ ,{"stream_name":"stdout","time":2076.041626453,"data":"{\"probe\": \"ridge_aug\", \"seed\": 17, \"set\": \"sealed_1_new\", \"acc\": 0.890625}\n"}
64
+ ,{"stream_name":"stdout","time":2076.170779267,"data":"{\"probe\": \"ridge_aug\", \"seed\": 17, \"set\": \"sealed_2_new\", \"acc\": 0.7135416666666666}\n"}
65
+ ,{"stream_name":"stdout","time":2076.301331179,"data":"{\"probe\": \"ridge_aug\", \"seed\": 17, \"set\": \"sealed_4_new\", \"acc\": 0.4322916666666667}\n"}
66
+ ,{"stream_name":"stdout","time":2076.563257368,"data":"{\"probe\": \"mlp_aug\", \"seed\": 17, \"set\": \"sealed_ood\", \"acc\": 1.0}\n"}
67
+ ,{"stream_name":"stdout","time":2076.684802789,"data":"{\"probe\": \"mlp_aug\", \"seed\": 17, \"set\": \"sealed_1_new\", \"acc\": 1.0}\n"}
68
+ ,{"stream_name":"stdout","time":2076.808395106,"data":"{\"probe\": \"mlp_aug\", \"seed\": 17, \"set\": \"sealed_2_new\", \"acc\": 0.984375}\n"}
69
+ ,{"stream_name":"stdout","time":2076.931173743,"data":"{\"probe\": \"mlp_aug\", \"seed\": 17, \"set\": \"sealed_4_new\", \"acc\": 0.24479166666666666}\n"}
70
+ ,{"stream_name":"stdout","time":2077.182778765,"data":"{\"probe\": \"mlp_ctrl_shuffled\", \"seed\": 17, \"set\": \"sealed_ood\", \"acc\": 0.0}\n"}
71
+ ,{"stream_name":"stdout","time":2077.302429164,"data":"{\"probe\": \"mlp_ctrl_shuffled\", \"seed\": 17, \"set\": \"sealed_1_new\", \"acc\": 0.0}\n"}
72
+ ,{"stream_name":"stdout","time":2077.423131583,"data":"{\"probe\": \"mlp_ctrl_shuffled\", \"seed\": 17, \"set\": \"sealed_2_new\", \"acc\": 0.0}\n"}
73
+ ,{"stream_name":"stdout","time":2077.545791347,"data":"{\"probe\": \"mlp_ctrl_shuffled\", \"seed\": 17, \"set\": \"sealed_4_new\", \"acc\": 0.0}\n"}
74
+ ,{"stream_name":"stdout","time":2121.917288868,"data":"{\"probe\": \"ridge_aug\", \"seed\": 29, \"set\": \"sealed_ood\", \"acc\": 1.0}\n"}
75
+ ,{"stream_name":"stdout","time":2122.039396132,"data":"{\"probe\": \"ridge_aug\", \"seed\": 29, \"set\": \"sealed_1_new\", \"acc\": 0.9114583333333334}\n"}
76
+ ,{"stream_name":"stdout","time":2122.164308907,"data":"{\"probe\": \"ridge_aug\", \"seed\": 29, \"set\": \"sealed_2_new\", \"acc\": 0.7604166666666666}\n"}
77
+ ,{"stream_name":"stdout","time":2122.28846083,"data":"{\"probe\": \"ridge_aug\", \"seed\": 29, \"set\": \"sealed_4_new\", \"acc\": 0.421875}\n"}
78
+ ,{"stream_name":"stdout","time":2122.539580224,"data":"{\"probe\": \"mlp_aug\", \"seed\": 29, \"set\": \"sealed_ood\", \"acc\": 1.0}\n"}
79
+ ,{"stream_name":"stdout","time":2122.66336736,"data":"{\"probe\": \"mlp_aug\", \"seed\": 29, \"set\": \"sealed_1_new\", \"acc\": 1.0}\n"}
80
+ ,{"stream_name":"stdout","time":2122.788177567,"data":"{\"probe\": \"mlp_aug\", \"seed\": 29, \"set\": \"sealed_2_new\", \"acc\": 0.984375}\n"}
81
+ ,{"stream_name":"stdout","time":2122.913684039,"data":"{\"probe\": \"mlp_aug\", \"seed\": 29, \"set\": \"sealed_4_new\", \"acc\": 0.359375}\n"}
82
+ ,{"stream_name":"stdout","time":2123.169120479,"data":"{\"probe\": \"mlp_ctrl_shuffled\", \"seed\": 29, \"set\": \"sealed_ood\", \"acc\": 0.0}\n"}
83
+ ,{"stream_name":"stdout","time":2123.29559726,"data":"{\"probe\": \"mlp_ctrl_shuffled\", \"seed\": 29, \"set\": \"sealed_1_new\", \"acc\": 0.0}\n"}
84
+ ,{"stream_name":"stdout","time":2123.421828004,"data":"{\"probe\": \"mlp_ctrl_shuffled\", \"seed\": 29, \"set\": \"sealed_2_new\", \"acc\": 0.0}\n"}
85
+ ,{"stream_name":"stdout","time":2123.548216944,"data":"{\"probe\": \"mlp_ctrl_shuffled\", \"seed\": 29, \"set\": \"sealed_4_new\", \"acc\": 0.0}\n"}
86
+ ,{"stream_name":"stdout","time":2168.125527831,"data":"{\"probe\": \"ridge_aug\", \"seed\": 43, \"set\": \"sealed_ood\", \"acc\": 1.0}\n"}
87
+ ,{"stream_name":"stdout","time":2168.262863917,"data":"{\"probe\": \"ridge_aug\", \"seed\": 43, \"set\": \"sealed_1_new\", \"acc\": 0.8802083333333334}\n"}
88
+ ,{"stream_name":"stdout","time":2168.391112545,"data":"{\"probe\": \"ridge_aug\", \"seed\": 43, \"set\": \"sealed_2_new\", \"acc\": 0.7083333333333334}\n"}
89
+ ,{"stream_name":"stdout","time":2168.517144571,"data":"{\"probe\": \"ridge_aug\", \"seed\": 43, \"set\": \"sealed_4_new\", \"acc\": 0.4791666666666667}\n"}
90
+ ,{"stream_name":"stdout","time":2168.76732195,"data":"{\"probe\": \"mlp_aug\", \"seed\": 43, \"set\": \"sealed_ood\", \"acc\": 1.0}\n"}
91
+ ,{"stream_name":"stdout","time":2168.89285081,"data":"{\"probe\": \"mlp_aug\", \"seed\": 43, \"set\": \"sealed_1_new\", \"acc\": 1.0}\n"}
92
+ ,{"stream_name":"stdout","time":2169.022178723,"data":"{\"probe\": \"mlp_aug\", \"seed\": 43, \"set\": \"sealed_2_new\", \"acc\": 0.9635416666666666}\n"}
93
+ ,{"stream_name":"stdout","time":2169.154125255,"data":"{\"probe\": \"mlp_aug\", \"seed\": 43, \"set\": \"sealed_4_new\", \"acc\": 0.3229166666666667}\n"}
94
+ ,{"stream_name":"stdout","time":2169.415876585,"data":"{\"probe\": \"mlp_ctrl_shuffled\", \"seed\": 43, \"set\": \"sealed_ood\", \"acc\": 0.0}\n"}
95
+ ,{"stream_name":"stdout","time":2169.541812178,"data":"{\"probe\": \"mlp_ctrl_shuffled\", \"seed\": 43, \"set\": \"sealed_1_new\", \"acc\": 0.0}\n"}
96
+ ,{"stream_name":"stdout","time":2169.667708314,"data":"{\"probe\": \"mlp_ctrl_shuffled\", \"seed\": 43, \"set\": \"sealed_2_new\", \"acc\": 0.0}\n"}
97
+ ,{"stream_name":"stdout","time":2169.796014845,"data":"{\"probe\": \"mlp_ctrl_shuffled\", \"seed\": 43, \"set\": \"sealed_4_new\", \"acc\": 0.0}\n"}
98
+ ,{"stream_name":"stdout","time":2183.039922099,"data":"{\"control\": \"random_receiver\", \"set\": \"sealed_ood\", \"text_template\": 0.0, \"k2_021_s17\": 0.0, \"k2_021_s29\": 0.0, \"k2_021_s43\": 0.0}\n"}
99
+ ,{"stream_name":"stdout","time":2189.400291153,"data":"{\"control\": \"random_receiver\", \"set\": \"sealed_4_new\", \"text_template\": 0.0, \"k2_021_s17\": 0.0, \"k2_021_s29\": 0.0, \"k2_021_s43\": 0.0}\n"}
100
+ ,{"stream_name":"stderr","time":2198.480349048,"data":"/usr/local/lib/python3.13/dist-packages/mistune.py:435: SyntaxWarning: invalid escape sequence '\\|'\n"}
101
+ ,{"stream_name":"stderr","time":2198.480380816,"data":" cells[i][c] = re.sub('\\\\\\\\\\|', '|', cell)\n"}
102
+ ,{"stream_name":"stderr","time":2198.73720129,"data":"/usr/local/lib/python3.13/dist-packages/nbconvert/filters/filter_links.py:36: SyntaxWarning: invalid escape sequence '\\_'\n"}
103
+ ,{"stream_name":"stderr","time":2198.737254349,"data":" text = re.sub(r'_', '\\_', text) # Escape underscores in display text\n"}
104
+ ,{"stream_name":"stderr","time":2199.560446684,"data":"[NbConvertApp] Converting notebook __script__.ipynb to html\n"}
105
+ ,{"stream_name":"stderr","time":2200.058837332,"data":"[NbConvertApp] Writing 278885 bytes to __results__.html\n"}
106
+ ]
021/cuerdas021-b.log ADDED
@@ -0,0 +1,152 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [{"stream_name":"stdout","time":4.361350775,"data":" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 44.1/44.1 kB 1.4 MB/s eta 0:00:00\n"}
2
+ ,{"stream_name":"stdout","time":5.34820775,"data":" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 10.0/10.0 MB 87.4 MB/s eta 0:00:00\n"}
3
+ ,{"stream_name":"stdout","time":5.383756769,"data":" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 566.4/566.4 kB 24.8 MB/s eta 0:00:00\n"}
4
+ ,{"stream_name":"stdout","time":5.438858748,"data":" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 3.0/3.0 MB 71.6 MB/s eta 0:00:00\n"}
5
+ ,{"stream_name":"stderr","time":18.988138587,"data":"ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.\n"}
6
+ ,{"stream_name":"stderr","time":18.988160772,"data":"litellm 1.85.7 requires tokenizers\u003c1.0,\u003e=0.21.0, but you have tokenizers 0.20.3 which is incompatible.\n"}
7
+ ,{"stream_name":"stderr","time":18.988165608,"data":"gradio 6.26.0 requires huggingface-hub\u003c2.0,\u003e=1.16.0, but you have huggingface-hub 0.36.2 which is incompatible.\n"}
8
+ ,{"stream_name":"stderr","time":18.988169421,"data":"diffusers 0.40.0 requires huggingface-hub\u003c2.0,\u003e=1.23.0, but you have huggingface-hub 0.36.2 which is incompatible.\n"}
9
+ ,{"stream_name":"stdout","time":71.795176862,"data":"{\"positions_per_event\": \"values\", \"sealed_id\": 19.0, \"sealed_ood\": 19.0, \"sealed_1_new\": 19.4, \"sealed_2_new\": 19.95, \"sealed_4_new\": 20.85}\n"}
10
+ ,{"stream_name":"stdout","time":71.846933674,"data":"{\"positions_per_event\": \"values_csv\", \"sealed_id\": 22.0, \"sealed_ood\": 22.0, \"sealed_1_new\": 22.4, \"sealed_2_new\": 22.95, \"sealed_4_new\": 23.85}\n"}
11
+ ,{"stream_name":"stdout","time":71.929283084,"data":"{\"positions_per_event\": \"kv\", \"sealed_id\": 31.0, \"sealed_ood\": 31.0, \"sealed_1_new\": 31.21, \"sealed_2_new\": 31.41, \"sealed_4_new\": 32.01}\n"}
12
+ ,{"stream_name":"stdout","time":72.048858006,"data":"{\"positions_per_event\": \"template\", \"sealed_id\": 35.0, \"sealed_ood\": 35.0, \"sealed_1_new\": 35.21, \"sealed_2_new\": 35.41, \"sealed_4_new\": 36.01}\n"}
13
+ ,{"stream_name":"stdout","time":117.831483784,"data":"{\"sys\": \"fewshot_values\", \"shared_prefix_positions\": 99, \"sealed_id\": 0.345, \"sealed_ood\": 0.325, \"sealed_1_new\": 0.03125, \"sealed_2_new\": 0.0, \"sealed_4_new\": 0.0}\n"}
14
+ ,{"stream_name":"stdout","time":174.966069479,"data":"{\"sys\": \"fewshot_values_csv\", \"shared_prefix_positions\": 105, \"sealed_id\": 0.89, \"sealed_ood\": 0.9075, \"sealed_1_new\": 0.5, \"sealed_2_new\": 0.24479166666666666, \"sealed_4_new\": 0.046875}\n"}
15
+ ,{"stream_name":"stdout","time":245.00421725,"data":"{\"sys\": \"fewshot_kv\", \"shared_prefix_positions\": 123, \"sealed_id\": 0.945, \"sealed_ood\": 0.9175, \"sealed_1_new\": 0.84375, \"sealed_2_new\": 0.8645833333333334, \"sealed_4_new\": 0.8958333333333334}\n"}
16
+ ,{"stream_name":"stdout","time":321.724758965,"data":"{\"sys\": \"fewshot_template\", \"shared_prefix_positions\": 131, \"sealed_id\": 0.9, \"sealed_ood\": 0.865, \"sealed_1_new\": 0.6666666666666666, \"sealed_2_new\": 0.5833333333333334, \"sealed_4_new\": 0.3854166666666667}\n"}
17
+ ,{"stream_name":"stdout","time":323.808552115,"data":"{\"run\": \"sp_values_s17\", \"step\": 0, \"val\": 0.0}\n"}
18
+ ,{"stream_name":"stderr","time":411.215770849,"data":"/usr/local/lib/python3.13/dist-packages/torch/utils/_device.py:116: UserWarning: Converting a tensor with requires_grad=True to a scalar may lead to unexpected behavior.\n"}
19
+ ,{"stream_name":"stderr","time":411.215814739,"data":"Consider using tensor.detach() first. (Triggered internally at /pytorch/torch/csrc/autograd/generated/python_variable_methods.cpp:836.)\n"}
20
+ ,{"stream_name":"stderr","time":411.215821065,"data":" return func(*args, **kwargs)\n"}
21
+ ,{"stream_name":"stdout","time":411.21584086,"data":"{\"run\": \"sp_values_s17\", \"step\": 250, \"val\": 1.6145833333333333, \"loss\": 3.044680595397949, \"s\": 87}\n"}
22
+ ,{"stream_name":"stdout","time":497.745659182,"data":"{\"run\": \"sp_values_s17\", \"step\": 500, \"val\": 1.6666666666666667, \"loss\": 3.143272638320923, \"s\": 174}\n"}
23
+ ,{"stream_name":"stdout","time":584.606442804,"data":"{\"run\": \"sp_values_s17\", \"step\": 750, \"val\": 1.6770833333333335, \"loss\": 2.6209170818328857, \"s\": 261}\n"}
24
+ ,{"stream_name":"stdout","time":671.475163664,"data":"{\"run\": \"sp_values_s17\", \"step\": 1000, \"val\": 1.6041666666666665, \"loss\": 3.025043487548828, \"s\": 348}\n"}
25
+ ,{"stream_name":"stdout","time":758.830169521,"data":"{\"run\": \"sp_values_s17\", \"step\": 1250, \"val\": 1.7083333333333335, \"loss\": 2.137320041656494, \"s\": 435}\n"}
26
+ ,{"stream_name":"stdout","time":846.891577041,"data":"{\"run\": \"sp_values_s17\", \"step\": 1500, \"val\": 1.71875, \"loss\": 1.954566240310669, \"s\": 523}\n"}
27
+ ,{"stream_name":"stdout","time":934.268076228,"data":"{\"run\": \"sp_values_s17\", \"step\": 1750, \"val\": 1.6354166666666665, \"loss\": 3.326815128326416, \"s\": 610}\n"}
28
+ ,{"stream_name":"stdout","time":1022.820649516,"data":"{\"run\": \"sp_values_s17\", \"step\": 2000, \"val\": 1.7291666666666665, \"loss\": 2.481618642807007, \"s\": 699}\n"}
29
+ ,{"stream_name":"stdout","time":1110.19483707,"data":"{\"run\": \"sp_values_s17\", \"step\": 2250, \"val\": 1.7083333333333335, \"loss\": 2.0035409927368164, \"s\": 786}\n"}
30
+ ,{"stream_name":"stdout","time":1196.253999266,"data":"{\"run\": \"sp_values_s17\", \"step\": 2500, \"val\": 1.7395833333333335, \"loss\": 1.278429627418518, \"s\": 872}\n"}
31
+ ,{"stream_name":"stdout","time":1283.403222855,"data":"{\"run\": \"sp_values_s17\", \"step\": 2750, \"val\": 1.7708333333333335, \"loss\": 1.8047574758529663, \"s\": 960}\n"}
32
+ ,{"stream_name":"stdout","time":1370.785137672,"data":"{\"run\": \"sp_values_s17\", \"step\": 3000, \"val\": 1.7395833333333335, \"loss\": 1.2187349796295166, \"s\": 1047}\n"}
33
+ ,{"stream_name":"stdout","time":1382.603182089,"data":"{\"sys\": \"sp_values_s17\", \"params\": 3842368, \"selected_step\": 2750, \"val\": 1.7708333333333335, \"soft_prompt_positions\": 8, \"sealed_id\": 1.0, \"sealed_ood\": 1.0, \"sealed_1_new\": 1.0, \"sealed_2_new\": 0.984375, \"sealed_4_new\": 0.96875}\n"}
34
+ ,{"stream_name":"stdout","time":1384.615533132,"data":"{\"run\": \"sp_values_s29\", \"step\": 0, \"val\": 0.0}\n"}
35
+ ,{"stream_name":"stdout","time":1471.512106137,"data":"{\"run\": \"sp_values_s29\", \"step\": 250, \"val\": 1.375, \"loss\": 4.450610160827637, \"s\": 87}\n"}
36
+ ,{"stream_name":"stdout","time":1559.097106882,"data":"{\"run\": \"sp_values_s29\", \"step\": 500, \"val\": 1.6145833333333335, \"loss\": 2.346332550048828, \"s\": 174}\n"}
37
+ ,{"stream_name":"stdout","time":1645.830903071,"data":"{\"run\": \"sp_values_s29\", \"step\": 750, \"val\": 1.6145833333333335, \"loss\": 2.190868854522705, \"s\": 261}\n"}
38
+ ,{"stream_name":"stdout","time":1732.973225386,"data":"{\"run\": \"sp_values_s29\", \"step\": 1000, \"val\": 1.6666666666666665, \"loss\": 1.3257291316986084, \"s\": 348}\n"}
39
+ ,{"stream_name":"stdout","time":1820.431503621,"data":"{\"run\": \"sp_values_s29\", \"step\": 1250, \"val\": 1.6458333333333335, \"loss\": 2.51035737991333, \"s\": 436}\n"}
40
+ ,{"stream_name":"stdout","time":1907.325803849,"data":"{\"run\": \"sp_values_s29\", \"step\": 1500, \"val\": 1.6875, \"loss\": 2.5555827617645264, \"s\": 523}\n"}
41
+ ,{"stream_name":"stdout","time":1993.856924421,"data":"{\"run\": \"sp_values_s29\", \"step\": 1750, \"val\": 1.6041666666666665, \"loss\": 1.5838260650634766, \"s\": 609}\n"}
42
+ ,{"stream_name":"stdout","time":2081.193454877,"data":"{\"run\": \"sp_values_s29\", \"step\": 2000, \"val\": 1.6770833333333335, \"loss\": 2.4831151962280273, \"s\": 697}\n"}
43
+ ,{"stream_name":"stdout","time":2168.354451021,"data":"{\"run\": \"sp_values_s29\", \"step\": 2250, \"val\": 1.7083333333333335, \"loss\": 1.9796416759490967, \"s\": 784}\n"}
44
+ ,{"stream_name":"stdout","time":2255.405604058,"data":"{\"run\": \"sp_values_s29\", \"step\": 2500, \"val\": 1.6666666666666665, \"loss\": 1.4970152378082275, \"s\": 871}\n"}
45
+ ,{"stream_name":"stdout","time":2342.536059791,"data":"{\"run\": \"sp_values_s29\", \"step\": 2750, \"val\": 1.6875, \"loss\": 2.13800311088562, \"s\": 958}\n"}
46
+ ,{"stream_name":"stdout","time":2428.773263448,"data":"{\"run\": \"sp_values_s29\", \"step\": 3000, \"val\": 1.6875, \"loss\": 2.1700429916381836, \"s\": 1044}\n"}
47
+ ,{"stream_name":"stdout","time":2440.483019732,"data":"{\"sys\": \"sp_values_s29\", \"params\": 3842368, \"selected_step\": 2250, \"val\": 1.7083333333333335, \"soft_prompt_positions\": 8, \"sealed_id\": 1.0, \"sealed_ood\": 1.0, \"sealed_1_new\": 1.0, \"sealed_2_new\": 0.9895833333333334, \"sealed_4_new\": 0.9635416666666666}\n"}
48
+ ,{"stream_name":"stdout","time":2442.496346525,"data":"{\"run\": \"sp_values_s43\", \"step\": 0, \"val\": 0.0}\n"}
49
+ ,{"stream_name":"stdout","time":2529.924771935,"data":"{\"run\": \"sp_values_s43\", \"step\": 250, \"val\": 1.4583333333333333, \"loss\": 2.9020471572875977, \"s\": 87}\n"}
50
+ ,{"stream_name":"stdout","time":2615.888976615,"data":"{\"run\": \"sp_values_s43\", \"step\": 500, \"val\": 1.46875, \"loss\": 2.5136404037475586, \"s\": 173}\n"}
51
+ ,{"stream_name":"stdout","time":2704.372018282,"data":"{\"run\": \"sp_values_s43\", \"step\": 750, \"val\": 1.625, \"loss\": 2.120654821395874, \"s\": 262}\n"}
52
+ ,{"stream_name":"stdout","time":2791.388173071,"data":"{\"run\": \"sp_values_s43\", \"step\": 1000, \"val\": 1.6770833333333335, \"loss\": 3.3165745735168457, \"s\": 349}\n"}
53
+ ,{"stream_name":"stdout","time":2879.397411429,"data":"{\"run\": \"sp_values_s43\", \"step\": 1250, \"val\": 1.78125, \"loss\": 2.017547607421875, \"s\": 437}\n"}
54
+ ,{"stream_name":"stdout","time":2966.957135373,"data":"{\"run\": \"sp_values_s43\", \"step\": 1500, \"val\": 1.78125, \"loss\": 2.347735643386841, \"s\": 524}\n"}
55
+ ,{"stream_name":"stdout","time":3054.617083351,"data":"{\"run\": \"sp_values_s43\", \"step\": 1750, \"val\": 1.7708333333333335, \"loss\": 2.394887924194336, \"s\": 612}\n"}
56
+ ,{"stream_name":"stdout","time":3141.533202514,"data":"{\"run\": \"sp_values_s43\", \"step\": 2000, \"val\": 1.625, \"loss\": 1.9778445959091187, \"s\": 699}\n"}
57
+ ,{"stream_name":"stdout","time":3227.879030741,"data":"{\"run\": \"sp_values_s43\", \"step\": 2250, \"val\": 1.78125, \"loss\": 2.360875129699707, \"s\": 785}\n"}
58
+ ,{"stream_name":"stdout","time":3314.498209102,"data":"{\"run\": \"sp_values_s43\", \"step\": 2500, \"val\": 1.7916666666666665, \"loss\": 1.7670954465866089, \"s\": 872}\n"}
59
+ ,{"stream_name":"stdout","time":3401.342048246,"data":"{\"run\": \"sp_values_s43\", \"step\": 2750, \"val\": 1.75, \"loss\": 2.02799654006958, \"s\": 959}\n"}
60
+ ,{"stream_name":"stdout","time":3488.451204805,"data":"{\"run\": \"sp_values_s43\", \"step\": 3000, \"val\": 1.7395833333333335, \"loss\": 1.7585619688034058, \"s\": 1046}\n"}
61
+ ,{"stream_name":"stdout","time":3500.244749137,"data":"{\"sys\": \"sp_values_s43\", \"params\": 3842368, \"selected_step\": 2500, \"val\": 1.7916666666666665, \"soft_prompt_positions\": 8, \"sealed_id\": 1.0, \"sealed_ood\": 1.0, \"sealed_1_new\": 1.0, \"sealed_2_new\": 0.984375, \"sealed_4_new\": 0.9322916666666666}\n"}
62
+ ,{"stream_name":"stdout","time":3502.496340876,"data":"{\"run\": \"sp_values_csv_s17\", \"step\": 0, \"val\": 0.0}\n"}
63
+ ,{"stream_name":"stdout","time":3597.190372836,"data":"{\"run\": \"sp_values_csv_s17\", \"step\": 250, \"val\": 1.90625, \"loss\": 2.3863179683685303, \"s\": 95}\n"}
64
+ ,{"stream_name":"stdout","time":3691.237500232,"data":"{\"run\": \"sp_values_csv_s17\", \"step\": 500, \"val\": 1.9583333333333335, \"loss\": 2.344421863555908, \"s\": 189}\n"}
65
+ ,{"stream_name":"stdout","time":3785.572447256,"data":"{\"run\": \"sp_values_csv_s17\", \"step\": 750, \"val\": 1.9791666666666665, \"loss\": 1.9774900674819946, \"s\": 283}\n"}
66
+ ,{"stream_name":"stdout","time":3879.655793829,"data":"{\"run\": \"sp_values_csv_s17\", \"step\": 1000, \"val\": 1.9583333333333335, \"loss\": 1.8494818210601807, \"s\": 377}\n"}
67
+ ,{"stream_name":"stdout","time":3974.548123947,"data":"{\"run\": \"sp_values_csv_s17\", \"step\": 1250, \"val\": 1.9479166666666665, \"loss\": 1.641819715499878, \"s\": 472}\n"}
68
+ ,{"stream_name":"stdout","time":4069.833349564,"data":"{\"run\": \"sp_values_csv_s17\", \"step\": 1500, \"val\": 1.9895833333333335, \"loss\": 1.3760128021240234, \"s\": 567}\n"}
69
+ ,{"stream_name":"stdout","time":4164.647401236,"data":"{\"run\": \"sp_values_csv_s17\", \"step\": 1750, \"val\": 1.9895833333333335, \"loss\": 1.9606835842132568, \"s\": 662}\n"}
70
+ ,{"stream_name":"stdout","time":4260.292046457,"data":"{\"run\": \"sp_values_csv_s17\", \"step\": 2000, \"val\": 2.0, \"loss\": 1.9302983283996582, \"s\": 758}\n"}
71
+ ,{"stream_name":"stdout","time":4354.666832377,"data":"{\"run\": \"sp_values_csv_s17\", \"step\": 2250, \"val\": 1.9895833333333335, \"loss\": 1.5998188257217407, \"s\": 852}\n"}
72
+ ,{"stream_name":"stdout","time":4448.229734015,"data":"{\"run\": \"sp_values_csv_s17\", \"step\": 2500, \"val\": 1.9895833333333335, \"loss\": 1.07892906665802, \"s\": 946}\n"}
73
+ ,{"stream_name":"stdout","time":4542.45081406,"data":"{\"run\": \"sp_values_csv_s17\", \"step\": 2750, \"val\": 2.0, \"loss\": 1.3307654857635498, \"s\": 1040}\n"}
74
+ ,{"stream_name":"stdout","time":4636.918178215,"data":"{\"run\": \"sp_values_csv_s17\", \"step\": 3000, \"val\": 2.0, \"loss\": 0.9787193536758423, \"s\": 1134}\n"}
75
+ ,{"stream_name":"stdout","time":4650.271880555,"data":"{\"sys\": \"sp_values_csv_s17\", \"params\": 3842368, \"selected_step\": 2000, \"val\": 2.0, \"soft_prompt_positions\": 8, \"sealed_id\": 1.0, \"sealed_ood\": 1.0, \"sealed_1_new\": 1.0, \"sealed_2_new\": 1.0, \"sealed_4_new\": 1.0}\n"}
76
+ ,{"stream_name":"stdout","time":4652.516545442,"data":"{\"run\": \"sp_values_csv_s29\", \"step\": 0, \"val\": 0.0}\n"}
77
+ ,{"stream_name":"stdout","time":4746.943424249,"data":"{\"run\": \"sp_values_csv_s29\", \"step\": 250, \"val\": 1.9583333333333335, \"loss\": 2.902418851852417, \"s\": 94}\n"}
78
+ ,{"stream_name":"stdout","time":4842.027314419,"data":"{\"run\": \"sp_values_csv_s29\", \"step\": 500, \"val\": 1.9375, \"loss\": 1.829451084136963, \"s\": 190}\n"}
79
+ ,{"stream_name":"stdout","time":4935.870307238,"data":"{\"run\": \"sp_values_csv_s29\", \"step\": 750, \"val\": 1.9375, \"loss\": 1.5734684467315674, \"s\": 283}\n"}
80
+ ,{"stream_name":"stdout","time":5030.65106833,"data":"{\"run\": \"sp_values_csv_s29\", \"step\": 1000, \"val\": 1.96875, \"loss\": 0.8627220988273621, \"s\": 378}\n"}
81
+ ,{"stream_name":"stdout","time":5125.899735903,"data":"{\"run\": \"sp_values_csv_s29\", \"step\": 1250, \"val\": 1.9791666666666665, \"loss\": 1.8343243598937988, \"s\": 473}\n"}
82
+ ,{"stream_name":"stdout","time":5219.968519164,"data":"{\"run\": \"sp_values_csv_s29\", \"step\": 1500, \"val\": 1.9895833333333335, \"loss\": 1.6268024444580078, \"s\": 567}\n"}
83
+ ,{"stream_name":"stdout","time":5313.952497176,"data":"{\"run\": \"sp_values_csv_s29\", \"step\": 1750, \"val\": 1.96875, \"loss\": 1.224697232246399, \"s\": 661}\n"}
84
+ ,{"stream_name":"stdout","time":5408.727257067,"data":"{\"run\": \"sp_values_csv_s29\", \"step\": 2000, \"val\": 1.9583333333333335, \"loss\": 1.8627893924713135, \"s\": 756}\n"}
85
+ ,{"stream_name":"stdout","time":5503.152377398,"data":"{\"run\": \"sp_values_csv_s29\", \"step\": 2250, \"val\": 1.9583333333333335, \"loss\": 1.6173484325408936, \"s\": 851}\n"}
86
+ ,{"stream_name":"stdout","time":5597.532104906,"data":"{\"run\": \"sp_values_csv_s29\", \"step\": 2500, \"val\": 1.9791666666666665, \"loss\": 1.2948013544082642, \"s\": 945}\n"}
87
+ ,{"stream_name":"stdout","time":5692.405593825,"data":"{\"run\": \"sp_values_csv_s29\", \"step\": 2750, \"val\": 1.96875, \"loss\": 1.518986463546753, \"s\": 1040}\n"}
88
+ ,{"stream_name":"stdout","time":5786.105113037,"data":"{\"run\": \"sp_values_csv_s29\", \"step\": 3000, \"val\": 1.96875, \"loss\": 1.6026052236557007, \"s\": 1134}\n"}
89
+ ,{"stream_name":"stdout","time":5799.436388588,"data":"{\"sys\": \"sp_values_csv_s29\", \"params\": 3842368, \"selected_step\": 1500, \"val\": 1.9895833333333335, \"soft_prompt_positions\": 8, \"sealed_id\": 1.0, \"sealed_ood\": 1.0, \"sealed_1_new\": 1.0, \"sealed_2_new\": 1.0, \"sealed_4_new\": 1.0}\n"}
90
+ ,{"stream_name":"stdout","time":5801.656527668,"data":"{\"run\": \"sp_values_csv_s43\", \"step\": 0, \"val\": 0.0}\n"}
91
+ ,{"stream_name":"stdout","time":5896.668236069,"data":"{\"run\": \"sp_values_csv_s43\", \"step\": 250, \"val\": 1.8854166666666665, \"loss\": 2.186886787414551, \"s\": 95}\n"}
92
+ ,{"stream_name":"stdout","time":5990.220831139,"data":"{\"run\": \"sp_values_csv_s43\", \"step\": 500, \"val\": 1.96875, \"loss\": 1.741999864578247, \"s\": 189}\n"}
93
+ ,{"stream_name":"stdout","time":6086.120661244,"data":"{\"run\": \"sp_values_csv_s43\", \"step\": 750, \"val\": 1.8645833333333335, \"loss\": 1.5374534130096436, \"s\": 284}\n"}
94
+ ,{"stream_name":"stdout","time":6181.057805559,"data":"{\"run\": \"sp_values_csv_s43\", \"step\": 1000, \"val\": 1.9583333333333335, \"loss\": 2.0583114624023438, \"s\": 379}\n"}
95
+ ,{"stream_name":"stdout","time":6276.707419186,"data":"{\"run\": \"sp_values_csv_s43\", \"step\": 1250, \"val\": 1.9166666666666665, \"loss\": 1.2399407625198364, \"s\": 475}\n"}
96
+ ,{"stream_name":"stdout","time":6371.443170589,"data":"{\"run\": \"sp_values_csv_s43\", \"step\": 1500, \"val\": 1.96875, \"loss\": 1.6992583274841309, \"s\": 570}\n"}
97
+ ,{"stream_name":"stdout","time":6466.554435181,"data":"{\"run\": \"sp_values_csv_s43\", \"step\": 1750, \"val\": 1.9895833333333335, \"loss\": 1.7322700023651123, \"s\": 665}\n"}
98
+ ,{"stream_name":"stdout","time":6561.062734654,"data":"{\"run\": \"sp_values_csv_s43\", \"step\": 2000, \"val\": 1.9583333333333335, \"loss\": 1.5983898639678955, \"s\": 759}\n"}
99
+ ,{"stream_name":"stdout","time":6654.941557139,"data":"{\"run\": \"sp_values_csv_s43\", \"step\": 2250, \"val\": 1.96875, \"loss\": 1.6161710023880005, \"s\": 853}\n"}
100
+ ,{"stream_name":"stdout","time":6749.082979317,"data":"{\"run\": \"sp_values_csv_s43\", \"step\": 2500, \"val\": 1.96875, \"loss\": 1.3451036214828491, \"s\": 947}\n"}
101
+ ,{"stream_name":"stdout","time":6843.554073075,"data":"{\"run\": \"sp_values_csv_s43\", \"step\": 2750, \"val\": 1.9791666666666665, \"loss\": 1.2667261362075806, \"s\": 1042}\n"}
102
+ ,{"stream_name":"stdout","time":6938.307644552,"data":"{\"run\": \"sp_values_csv_s43\", \"step\": 3000, \"val\": 1.96875, \"loss\": 1.4136821031570435, \"s\": 1137}\n"}
103
+ ,{"stream_name":"stdout","time":6951.710262143,"data":"{\"sys\": \"sp_values_csv_s43\", \"params\": 3842368, \"selected_step\": 1750, \"val\": 1.9895833333333335, \"soft_prompt_positions\": 8, \"sealed_id\": 1.0, \"sealed_ood\": 1.0, \"sealed_1_new\": 1.0, \"sealed_2_new\": 1.0, \"sealed_4_new\": 1.0}\n"}
104
+ ,{"stream_name":"stdout","time":6954.609131963,"data":"{\"run\": \"sp_kv_s17\", \"step\": 0, \"val\": 0.0625}\n"}
105
+ ,{"stream_name":"stdout","time":7069.711213562,"data":"{\"run\": \"sp_kv_s17\", \"step\": 250, \"val\": 2.0, \"loss\": 1.9125851392745972, \"s\": 115}\n"}
106
+ ,{"stream_name":"stdout","time":7185.042959089,"data":"{\"run\": \"sp_kv_s17\", \"step\": 500, \"val\": 2.0, \"loss\": 2.0021212100982666, \"s\": 230}\n"}
107
+ ,{"stream_name":"stdout","time":7300.369875705,"data":"{\"run\": \"sp_kv_s17\", \"step\": 750, \"val\": 2.0, \"loss\": 1.5451703071594238, \"s\": 346}\n"}
108
+ ,{"stream_name":"stdout","time":7415.495030592,"data":"{\"run\": \"sp_kv_s17\", \"step\": 1000, \"val\": 2.0, \"loss\": 1.4742144346237183, \"s\": 461}\n"}
109
+ ,{"stream_name":"stdout","time":7531.547118465,"data":"{\"run\": \"sp_kv_s17\", \"step\": 1250, \"val\": 2.0, \"loss\": 1.4599279165267944, \"s\": 577}\n"}
110
+ ,{"stream_name":"stdout","time":7647.554036009,"data":"{\"run\": \"sp_kv_s17\", \"step\": 1500, \"val\": 2.0, \"loss\": 1.2338306903839111, \"s\": 693}\n"}
111
+ ,{"stream_name":"stdout","time":7763.436058644,"data":"{\"run\": \"sp_kv_s17\", \"step\": 1750, \"val\": 2.0, \"loss\": 1.8669570684432983, \"s\": 809}\n"}
112
+ ,{"stream_name":"stdout","time":7880.156176707,"data":"{\"run\": \"sp_kv_s17\", \"step\": 2000, \"val\": 2.0, \"loss\": 1.7696266174316406, \"s\": 926}\n"}
113
+ ,{"stream_name":"stdout","time":7995.855988061,"data":"{\"run\": \"sp_kv_s17\", \"step\": 2250, \"val\": 2.0, \"loss\": 1.350313425064087, \"s\": 1041}\n"}
114
+ ,{"stream_name":"stdout","time":8110.64570648,"data":"{\"run\": \"sp_kv_s17\", \"step\": 2500, \"val\": 2.0, \"loss\": 0.9131432175636292, \"s\": 1156}\n"}
115
+ ,{"stream_name":"stdout","time":8225.736727802,"data":"{\"run\": \"sp_kv_s17\", \"step\": 2750, \"val\": 2.0, \"loss\": 1.2375553846359253, \"s\": 1271}\n"}
116
+ ,{"stream_name":"stdout","time":8341.095640136,"data":"{\"run\": \"sp_kv_s17\", \"step\": 3000, \"val\": 2.0, \"loss\": 0.8844823837280273, \"s\": 1386}\n"}
117
+ ,{"stream_name":"stdout","time":8358.2884852,"data":"{\"sys\": \"sp_kv_s17\", \"params\": 3842368, \"selected_step\": 250, \"val\": 2.0, \"soft_prompt_positions\": 8, \"sealed_id\": 1.0, \"sealed_ood\": 1.0, \"sealed_1_new\": 1.0, \"sealed_2_new\": 1.0, \"sealed_4_new\": 1.0}\n"}
118
+ ,{"stream_name":"stdout","time":8361.209239923,"data":"{\"run\": \"sp_kv_s29\", \"step\": 0, \"val\": 0.07291666666666667}\n"}
119
+ ,{"stream_name":"stdout","time":8476.470163654,"data":"{\"run\": \"sp_kv_s29\", \"step\": 250, \"val\": 1.9895833333333335, \"loss\": 2.1503524780273438, \"s\": 115}\n"}
120
+ ,{"stream_name":"stdout","time":8592.654786484,"data":"{\"run\": \"sp_kv_s29\", \"step\": 500, \"val\": 2.0, \"loss\": 1.5459184646606445, \"s\": 231}\n"}
121
+ ,{"stream_name":"stdout","time":8707.826241838,"data":"{\"run\": \"sp_kv_s29\", \"step\": 750, \"val\": 1.9895833333333335, \"loss\": 1.2727928161621094, \"s\": 347}\n"}
122
+ ,{"stream_name":"stdout","time":8823.59979388,"data":"{\"run\": \"sp_kv_s29\", \"step\": 1000, \"val\": 2.0, \"loss\": 0.745368242263794, \"s\": 462}\n"}
123
+ ,{"stream_name":"stdout","time":8939.752280122,"data":"{\"run\": \"sp_kv_s29\", \"step\": 1250, \"val\": 2.0, \"loss\": 1.6049551963806152, \"s\": 579}\n"}
124
+ ,{"stream_name":"stdout","time":9054.866158633,"data":"{\"run\": \"sp_kv_s29\", \"step\": 1500, \"val\": 2.0, \"loss\": 1.5594961643218994, \"s\": 694}\n"}
125
+ ,{"stream_name":"stdout","time":9169.872659237,"data":"{\"run\": \"sp_kv_s29\", \"step\": 1750, \"val\": 2.0, \"loss\": 1.1384079456329346, \"s\": 809}\n"}
126
+ ,{"stream_name":"stdout","time":9285.538980744,"data":"{\"run\": \"sp_kv_s29\", \"step\": 2000, \"val\": 1.96875, \"loss\": 1.4778153896331787, \"s\": 924}\n"}
127
+ ,{"stream_name":"stdout","time":9401.162170161,"data":"{\"run\": \"sp_kv_s29\", \"step\": 2250, \"val\": 2.0, \"loss\": 1.5377285480499268, \"s\": 1040}\n"}
128
+ ,{"stream_name":"stdout","time":9516.750259155,"data":"{\"run\": \"sp_kv_s29\", \"step\": 2500, \"val\": 2.0, \"loss\": 1.1550800800323486, \"s\": 1156}\n"}
129
+ ,{"stream_name":"stdout","time":9632.356820272,"data":"{\"run\": \"sp_kv_s29\", \"step\": 2750, \"val\": 2.0, \"loss\": 1.3703510761260986, \"s\": 1271}\n"}
130
+ ,{"stream_name":"stdout","time":9747.15441662,"data":"{\"run\": \"sp_kv_s29\", \"step\": 3000, \"val\": 2.0, \"loss\": 1.3216395378112793, \"s\": 1386}\n"}
131
+ ,{"stream_name":"stdout","time":9764.37948733,"data":"{\"sys\": \"sp_kv_s29\", \"params\": 3842368, \"selected_step\": 500, \"val\": 2.0, \"soft_prompt_positions\": 8, \"sealed_id\": 1.0, \"sealed_ood\": 1.0, \"sealed_1_new\": 1.0, \"sealed_2_new\": 1.0, \"sealed_4_new\": 1.0}\n"}
132
+ ,{"stream_name":"stdout","time":9767.312728899,"data":"{\"run\": \"sp_kv_s43\", \"step\": 0, \"val\": 0.07291666666666667}\n"}
133
+ ,{"stream_name":"stdout","time":9883.208554309,"data":"{\"run\": \"sp_kv_s43\", \"step\": 250, \"val\": 2.0, \"loss\": 1.6502081155776978, \"s\": 116}\n"}
134
+ ,{"stream_name":"stdout","time":9997.853492914,"data":"{\"run\": \"sp_kv_s43\", \"step\": 500, \"val\": 2.0, \"loss\": 1.304010272026062, \"s\": 231}\n"}
135
+ ,{"stream_name":"stdout","time":10114.418303539,"data":"{\"run\": \"sp_kv_s43\", \"step\": 750, \"val\": 2.0, \"loss\": 1.260686993598938, \"s\": 347}\n"}
136
+ ,{"stream_name":"stdout","time":10230.311754109,"data":"{\"run\": \"sp_kv_s43\", \"step\": 1000, \"val\": 2.0, \"loss\": 1.5032280683517456, \"s\": 463}\n"}
137
+ ,{"stream_name":"stdout","time":10346.890779959,"data":"{\"run\": \"sp_kv_s43\", \"step\": 1250, \"val\": 2.0, \"loss\": 1.139223337173462, \"s\": 580}\n"}
138
+ ,{"stream_name":"stdout","time":10462.6621338,"data":"{\"run\": \"sp_kv_s43\", \"step\": 1500, \"val\": 2.0, \"loss\": 1.4122791290283203, \"s\": 695}\n"}
139
+ ,{"stream_name":"stdout","time":10578.37733939,"data":"{\"run\": \"sp_kv_s43\", \"step\": 1750, \"val\": 2.0, \"loss\": 1.3497402667999268, \"s\": 811}\n"}
140
+ ,{"stream_name":"stdout","time":10693.879305726,"data":"{\"run\": \"sp_kv_s43\", \"step\": 2000, \"val\": 2.0, \"loss\": 1.427781581878662, \"s\": 927}\n"}
141
+ ,{"stream_name":"stdout","time":10808.648862386,"data":"{\"run\": \"sp_kv_s43\", \"step\": 2250, \"val\": 2.0, \"loss\": 1.3728115558624268, \"s\": 1041}\n"}
142
+ ,{"stream_name":"stdout","time":10924.288212135,"data":"{\"run\": \"sp_kv_s43\", \"step\": 2500, \"val\": 2.0, \"loss\": 1.1903347969055176, \"s\": 1157}\n"}
143
+ ,{"stream_name":"stdout","time":11039.960710764,"data":"{\"run\": \"sp_kv_s43\", \"step\": 2750, \"val\": 2.0, \"loss\": 1.029153823852539, \"s\": 1273}\n"}
144
+ ,{"stream_name":"stdout","time":11155.485832086,"data":"{\"run\": \"sp_kv_s43\", \"step\": 3000, \"val\": 2.0, \"loss\": 1.0028929710388184, \"s\": 1388}\n"}
145
+ ,{"stream_name":"stdout","time":11172.736693556,"data":"{\"sys\": \"sp_kv_s43\", \"params\": 3842368, \"selected_step\": 250, \"val\": 2.0, \"soft_prompt_positions\": 8, \"sealed_id\": 1.0, \"sealed_ood\": 1.0, \"sealed_1_new\": 1.0, \"sealed_2_new\": 1.0, \"sealed_4_new\": 1.0}\n"}
146
+ ,{"stream_name":"stderr","time":11181.503505782,"data":"/usr/local/lib/python3.13/dist-packages/mistune.py:435: SyntaxWarning: invalid escape sequence '\\|'\n"}
147
+ ,{"stream_name":"stderr","time":11181.503534751,"data":" cells[i][c] = re.sub('\\\\\\\\\\|', '|', cell)\n"}
148
+ ,{"stream_name":"stderr","time":11181.765160135,"data":"/usr/local/lib/python3.13/dist-packages/nbconvert/filters/filter_links.py:36: SyntaxWarning: invalid escape sequence '\\_'\n"}
149
+ ,{"stream_name":"stderr","time":11181.765180439,"data":" text = re.sub(r'_', '\\_', text) # Escape underscores in display text\n"}
150
+ ,{"stream_name":"stderr","time":11182.577499966,"data":"[NbConvertApp] Converting notebook __script__.ipynb to html\n"}
151
+ ,{"stream_name":"stderr","time":11183.090133157,"data":"[NbConvertApp] Writing 278885 bytes to __results__.html\n"}
152
+ ]
021/evalA.py ADDED
@@ -0,0 +1,55 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # CUERDAS-021 A: K=2 retrained with varied pool; sealed eval once; probes; controls.
2
+ from c021 import *
3
+ from transformers import AutoModelForCausalLM, AutoConfig
4
+ log = opener('resultsA.jsonl'); PRED = open(f'{OD}/predictionsA.jsonl', 'w'); SEEDS = [17] if SMOKE else [17, 29, 43]
5
+ log({'sealed_sha256': H, 'torch': torch.__version__, 'pool_sizes': {'words': len(WORDS), 'names': len(NAMES), 'subs': len(SUBS)}})
6
+ def ev(sysname, rowfn):
7
+ res = {}
8
+ for s in SETS:
9
+ r = per_ex(rowfn, S[s], CAND); res[s] = r
10
+ for i, c in enumerate(r): PRED.write(json.dumps({'sys': sysname, 'set': s, 'i': i, 'correct': bool(c)}) + '\n')
11
+ PRED.flush(); log({'sys': sysname, **{s: sum(v) / len(v) for s, v in res.items()}}); return res
12
+ ev('text_template', lambda mb: text_rows(mb, R.SUF_MAIN, R.TPL))
13
+ for sd in SEEDS:
14
+ c = R.PackK(2, sd); c.load_state_dict(torch.load(f'{INP}/k2long_s{sd}.pt', map_location=R.DEV)); c.eval()
15
+ ev(f'k2_018_s{sd}', lambda mb, c=c: k2_rows(c, mb, R.SUF_MAIN))
16
+ cods = {}
17
+ for sd in SEEDS:
18
+ teach = load_teacher(sd); k2 = R.PackK(2, sd); k2.load_state_dict(torch.load(f'{INP}/k2_s{sd}.pt', map_location=R.DEV))
19
+ step, v = train(list(k2.parameters()), lambda mb, tpl: k2_rows(k2, mb, tpl), teach, sd, f'k2_021_s{sd}', log, set_mode=lambda t: k2.train(t))
20
+ k2.eval().requires_grad_(False); torch.save(k2.state_dict(), f'{OD}/k2v_s{sd}.pt'); cods[sd] = k2
21
+ log({'selected': f'k2_021_s{sd}', 'step': step, 'val': v})
22
+ ev(f'k2_021_s{sd}', lambda mb, k2=k2: k2_rows(k2, mb, R.SUF_MAIN))
23
+ # ---- probes on retrained slots ----
24
+ VALS = [V[j] + NEW[j] for j in range(4)]; CEV = [F.normalize(torch.stack([R.lex(w) for w in VALS[j]]), dim=-1) for j in range(4)]
25
+ g = random.Random(2103); AUG = sample_msgs(g, 300 if SMOKE else 20000)
26
+ with torch.no_grad(): YA = torch.stack([torch.stack([R.lex(w) for w in m]) for m in AUG])
27
+ def slots(cod, ms):
28
+ with torch.no_grad(): return torch.cat([cod(torch.stack([torch.stack([R.lex(w) for w in m]) for m in ms[i:i+512]])).flatten(1) for i in range(0, len(ms), 512)])
29
+ def dec(P, ms):
30
+ ok = torch.ones(len(ms), dtype=torch.bool, device='cpu')
31
+ for j in range(4): ok &= (F.normalize(P[:, j], dim=-1) @ CEV[j].T).argmax(1).cpu() == torch.tensor([VALS[j].index(m[j]) for m in ms], device='cpu')
32
+ return ok.float().tolist()
33
+ def train_mlp(X, Y, shuffle=False, steps=None, seed=0):
34
+ steps = steps or (5 if SMOKE else 4000); torch.manual_seed(seed); Y = Y[torch.randperm(len(Y))] if shuffle else Y; mu, sdv = X.mean(0), X.std(0) + 1e-6
35
+ net = nn.Sequential(nn.Linear(X.shape[1], 1024), nn.GELU(), nn.Linear(1024, 1024), nn.GELU(), nn.Linear(1024, 4 * D)); opt = torch.optim.Adam(net.parameters(), 1e-3)
36
+ for _ in range(steps):
37
+ b = torch.randint(0, len(X), (256,)); p = net((X[b] - mu) / sdv).view(-1, 4, D); loss = (1 - F.cosine_similarity(p, Y[b], -1)).mean(); opt.zero_grad(); loss.backward(); opt.step()
38
+ net.eval(); return lambda Z: net((Z - mu) / sdv).view(-1, 4, D).detach()
39
+ for sd in SEEDS:
40
+ cod = cods[sd]; XA = slots(cod, AUG)
41
+ A = torch.linalg.solve(XA.T @ XA + 10 * torch.eye(XA.shape[1]), XA.T @ YA.flatten(1))
42
+ fns = {'ridge_aug': lambda Z: (Z @ A).view(-1, 4, D), 'mlp_aug': train_mlp(XA, YA, seed=sd), 'mlp_ctrl_shuffled': train_mlp(XA, YA, shuffle=True, seed=sd)}
43
+ for name, fn in fns.items():
44
+ for s in ('sealed_ood', 'sealed_1_new', 'sealed_2_new', 'sealed_4_new'):
45
+ r = dec(fn(slots(cod, S[s])), S[s])
46
+ for i, c in enumerate(r): PRED.write(json.dumps({'sys': f'{name}_s{sd}', 'set': s, 'i': i, 'correct': bool(c)}) + '\n')
47
+ log({'probe': name, 'seed': sd, 'set': s, 'acc': sum(r) / len(r)})
48
+ # ---- random-weight receiver ----
49
+ torch.manual_seed(0); rnd = AutoModelForCausalLM.from_config(AutoConfig.from_pretrained(R.MP)).float().eval().requires_grad_(False).to(R.DEV)
50
+ rnd.model.embed_tokens.weight.data.copy_(R.Emb.data); R.lm = rnd
51
+ for s in ('sealed_ood', 'sealed_4_new'):
52
+ e = {'control': 'random_receiver', 'set': s, 'text_template': sum(per_ex(lambda mb: text_rows(mb, R.SUF_MAIN, R.TPL), S[s], CAND)) / len(S[s])}
53
+ for sd in SEEDS: e[f'k2_021_s{sd}'] = sum(per_ex(lambda mb, c=cods[sd]: k2_rows(c, mb, R.SUF_MAIN), S[s], CAND)) / len(S[s])
54
+ log(e)
55
+ PRED.close()
021/evalB.py ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # CUERDAS-021 B: compact text with matched effort: few-shot and reparameterized soft prompt (~3.84M params, same steps/data/loss/selection).
2
+ from c021 import *
3
+ log = opener('resultsB.jsonl'); PRED = open(f'{OD}/predictionsB.jsonl', 'w'); SEEDS = [17] if SMOKE else [17, 29, 43]
4
+ FMT = {'values': "{action} {target} {status} {window}", 'values_csv': "{action}, {target}, {status}, {window}",
5
+ 'kv': "action: {action}; target: {target}; status: {status}; window: {window}.", 'template': R.TPL}
6
+ def npos(fmt, ms): return sum(len(tok.encode(fmt.format(**dict(zip(FIELDS, m))), add_special_tokens=False)) for m in ms) / len(ms) + 15
7
+ def ev(sysname, rowfn, extra):
8
+ res = {}
9
+ for s in SETS:
10
+ r = per_ex(rowfn, S[s], CAND); res[s] = r
11
+ for i, c in enumerate(r): PRED.write(json.dumps({'sys': sysname, 'set': s, 'i': i, 'correct': bool(c)}) + '\n')
12
+ PRED.flush(); log({'sys': sysname, **extra, **{s: sum(v) / len(v) for s, v in res.items()}})
13
+ for fn, fmt in FMT.items():
14
+ log({'positions_per_event': fn, **{s: round(npos(fmt, S[s]), 2) for s in SETS}})
15
+ # few-shot: 2 fixed training demos with all 4 answers (shared, cacheable prefix)
16
+ demos = random.Random(0).sample(R.FIT, 2)
17
+ for fn, fmt in FMT.items():
18
+ pre = ''.join(fmt.format(**dict(zip(FIELDS, d))) + ''.join(R.SUF_MAIN.format(field=FIELDS[j]) + ' ' + d[j] + '.' for j in range(4)) + '\n\n' for d in demos)
19
+ P = len(tok.encode(pre, add_special_tokens=False))
20
+ ev(f'fewshot_{fn}', lambda mb, f=pre + fmt: text_rows(mb, R.SUF_MAIN, f), {'shared_prefix_positions': P})
21
+ class SoftPrompt(nn.Module): # prefix-tuning style reparameterization: P = base + MLP(base)
22
+ def __init__(s, seed, n=8, h=3328):
23
+ super().__init__(); torch.manual_seed(seed)
24
+ s.base = nn.Parameter(Emb[torch.randint(1000, 20000, (n,))].clone())
25
+ s.mlp = nn.Sequential(nn.Linear(D, h), nn.Tanh(), nn.Linear(h, D)); nn.init.zeros_(s.mlp[-1].weight); nn.init.zeros_(s.mlp[-1].bias)
26
+ def forward(s): return s.base + s.mlp(s.base)
27
+ for fn in ('values', 'values_csv', 'kv'):
28
+ fmt = FMT[fn]
29
+ for sd in SEEDS:
30
+ sp = SoftPrompt(sd); npar = sum(p.numel() for p in sp.parameters()); teach = load_teacher(sd)
31
+ step, v = train(list(sp.parameters()), lambda mb, tpl: text_rows(mb, tpl, fmt, prefix=sp()), teach, sd, f'sp_{fn}_s{sd}', log, set_mode=lambda t: sp.train(t))
32
+ sp.eval().requires_grad_(False); torch.save(sp.state_dict(), f'{OD}/sp_{fn}_s{sd}.pt')
33
+ with torch.no_grad(): pr = sp()
34
+ ev(f'sp_{fn}_s{sd}', lambda mb: text_rows(mb, R.SUF_MAIN, fmt, prefix=pr), {'params': npar, 'selected_step': step, 'val': v, 'soft_prompt_positions': 8})
35
+ PRED.close()
021/kaggle_kernel_A.py ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os,subprocess,sys,glob,shutil
2
+ subprocess.run([sys.executable,'-m','pip','install','-q','transformers==4.46.3'],check=True)
3
+ inp=os.path.dirname(glob.glob('/kaggle/input/**/sealed021.json',recursive=True)[0])
4
+ S='/kaggle/working/src'; os.makedirs(S+'/data',exist_ok=True); shutil.copy(f'{inp}/expanded.json',f'{S}/data/')
5
+ W='/kaggle/working/code'; os.makedirs(W,exist_ok=True)
6
+ for f in ('run018.py','c021.py','evalA.py'): shutil.copy(f'{inp}/{f}',W)
7
+ from transformers import AutoTokenizer,AutoModelForCausalLM
8
+ mp='/kaggle/working/m135'; R='12fd25f77366fa6b3b4b768ec3050bf629380bac'
9
+ AutoTokenizer.from_pretrained('HuggingFaceTB/SmolLM2-135M-Instruct',revision=R).save_pretrained(mp); AutoModelForCausalLM.from_pretrained('HuggingFaceTB/SmolLM2-135M-Instruct',revision=R).save_pretrained(mp)
10
+ subprocess.run([sys.executable,f'{W}/evalA.py'],env=dict(os.environ,SRC=S,MP=mp,OD='/kaggle/working',INP=inp),check=True)
11
+ shutil.rmtree(mp); shutil.rmtree(S); shutil.rmtree(W)
021/kaggle_kernel_B.py ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os,subprocess,sys,glob,shutil
2
+ subprocess.run([sys.executable,'-m','pip','install','-q','transformers==4.46.3'],check=True)
3
+ inp=os.path.dirname(glob.glob('/kaggle/input/**/sealed021.json',recursive=True)[0])
4
+ S='/kaggle/working/src'; os.makedirs(S+'/data',exist_ok=True); shutil.copy(f'{inp}/expanded.json',f'{S}/data/')
5
+ W='/kaggle/working/code'; os.makedirs(W,exist_ok=True)
6
+ for f in ('run018.py','c021.py','evalB.py'): shutil.copy(f'{inp}/{f}',W)
7
+ from transformers import AutoTokenizer,AutoModelForCausalLM
8
+ mp='/kaggle/working/m135'; R='12fd25f77366fa6b3b4b768ec3050bf629380bac'
9
+ AutoTokenizer.from_pretrained('HuggingFaceTB/SmolLM2-135M-Instruct',revision=R).save_pretrained(mp); AutoModelForCausalLM.from_pretrained('HuggingFaceTB/SmolLM2-135M-Instruct',revision=R).save_pretrained(mp)
10
+ subprocess.run([sys.executable,f'{W}/evalB.py'],env=dict(os.environ,SRC=S,MP=mp,OD='/kaggle/working',INP=inp),check=True)
11
+ shutil.rmtree(mp); shutil.rmtree(S); shutil.rmtree(W)
021/make_sealed021.py ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # CUERDAS-021 sealed test generator (new seed). Unseen values: words, capitalized names, multi-token values.
2
+ import json, random, itertools, re, hashlib, sys
3
+ from transformers import AutoTokenizer
4
+ tok = AutoTokenizer.from_pretrained(sys.argv[1]); SEED = 20261017
5
+ meta = json.load(open(sys.argv[2])); V = meta['vocabulary']; rows = meta['rows']; SPL = meta['splits']
6
+ s020 = json.load(open(sys.argv[3]))
7
+ def enc(w): return tok.encode(' ' + w, add_special_tokens=False)
8
+ used = {tuple(r['values']) for r in rows}
9
+ fit = [tuple(rows[i]['values']) for i in SPL['fit']]; seen_at = {v[:2] for v in fit}; seen_sw = {v[2:] for v in fit}
10
+ isid = lambda v: v[:2] in seen_at and v[2:] in seen_sw
11
+ rng = random.Random(15015); uni = [v for v in itertools.product(*[range(len(x)) for x in V]) if v not in used]; rng.shuffle(uni)
12
+ dev = set([v for v in uni if isid(v)][:200] + [v for v in uni if not isid(v)][:400])
13
+ idx = lambda m: tuple(V[j].index(m[j]) for j in range(4))
14
+ s20 = {idx(m) for k in ('sealed_id', 'sealed_ood') for m in s020[k]}
15
+ elig = [v for v in itertools.product(*[range(len(x)) for x in V]) if v not in used and v not in dev and v not in s20 and not all(v[j] < (7, 7, 8, 5)[j] for j in range(4))]
16
+ r = random.Random(SEED); r.shuffle(elig)
17
+ sid = [v for v in elig if isid(v)][:200]; sood = [v for v in elig if not isid(v)][:400]
18
+ W = lambda v: [V[j][v[j]] for j in range(4)]
19
+ voc = tok.get_vocab()
20
+ words = sorted(t[1:] for t in voc if re.fullmatch(r'Ġ[a-z]{3,14}', t)); names = sorted(t[1:] for t in voc if re.fullmatch(r'Ġ[A-Z][a-z]{2,10}', t))
21
+ subs = sorted(t for t in voc if re.fullmatch(r'[a-z]{2,5}', t))
22
+ r.shuffle(words); r.shuffle(names); r.shuffle(subs)
23
+ old = set(sum(s020['new_words'], [])) | set('inspect repair cancel approve archive scan print send sensor battery printer server truck door screen cable broken stable late empty full warm missing urgent night evening weekend noon spring summer winter midnight'.split())
24
+ taken = {enc(w)[0] for vs in V for w in vs} | {enc(w)[0] for w in old}
25
+ NEW = [[] for _ in range(4)]; TYPE = {}
26
+ wi = iter(words); ni = iter(names)
27
+ def take(it):
28
+ while True:
29
+ w = next(it)
30
+ if len(enc(w)) == 1 and enc(w)[0] not in taken and w not in old: taken.add(enc(w)[0]); return w
31
+ for j in range(4):
32
+ for _ in range(3): w = take(wi); NEW[j].append(w); TYPE[w] = 'word'
33
+ for _ in range(3): w = take(ni); NEW[j].append(w); TYPE[w] = 'name'
34
+ a = take(wi); b = take(wi); p = a + ' ' + b; assert len(enc(p)) == 2; NEW[j].append(p); TYPE[p] = 'phrase'
35
+ while True:
36
+ a = take(wi); sb = subs.pop(); pw = a + sb
37
+ if enc(pw) == [enc(a)[0], tok.convert_tokens_to_ids(sb)]: NEW[j].append(pw); TYPE[pw] = 'multitoken_word'; break
38
+ excl = sorted({t for j in range(4) for w in NEW[j] for t in enc(w)})
39
+ def novel(n_new, n=192):
40
+ out = []
41
+ for k in range(n):
42
+ b = W(sood[k % 400])
43
+ for jj in r.sample(range(4), n_new): b[jj] = r.choice(NEW[jj])
44
+ out.append(b)
45
+ return out
46
+ S = {'seed': SEED, 'new_values': NEW, 'types': TYPE, 'exclude_token_ids': excl,
47
+ 'sealed_id': [W(v) for v in sid], 'sealed_ood': [W(v) for v in sood], 'sealed_1_new': novel(1), 'sealed_2_new': novel(2), 'sealed_4_new': novel(4)}
48
+ s = json.dumps(S, sort_keys=True); open(sys.argv[4], 'w').write(s)
49
+ print('sha256', hashlib.sha256(s.encode()).hexdigest(), 'eligible', len(elig))
50
+ print(NEW)
021/predictionsA.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
021/predictionsB.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
021/preregistro021.md ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # CUERDAS-021 — Pre-registro (escrito ANTES de cualquier evaluación sobre el conjunto sellado)
2
+
3
+ Fecha: 2026-10-16 (America/Mexico_City). Aprobado por Erik Sanchez. Solo recursos privados (Kaggle T4 gratis).
4
+
5
+ ## Conjunto sellado nuevo
6
+ - `make_sealed021.py`, semilla nueva **20261017**. Universo elegible: combinaciones del vocabulario 014 que no están en filas 014, ni en dev-ID/dev-CH (semilla 15015), ni en el sellado 020, ni en el universo 013 → 16,894.
7
+ - `sealed_id` (200), `sealed_ood` (400, con al menos un par retenido), `sealed_1_new`/`2_new`/`4_new` (192 c/u).
8
+ - Valores nuevos: 8 por campo = 3 palabras minúsculas de 1 token, 3 nombres/palabras capitalizadas de 1 token, 1 frase de 2 palabras y 1 palabra de varios sub-tokens. Opciones por campo = vocabulario 014 + esos 8 (primer token único). Los números NO entran al test sellado: en este tokenizador todos empiezan con el token " ", y la lectura de primer token no los distingue; sí entran al entrenamiento.
9
+ - Archivo `sealed021.json`, **SHA-256 a86a6a5ee375ae67a3e79f969c9553b0e5548ed3b0e80e5315862e6552ae72bc**; el kernel lo verifica. Todos los tokens de los valores sellados se excluyen del pool de entrenamiento y de validación.
10
+ - Se evalúa **una sola vez**. Una corrida previa de prueba (SMOKE) usó solo mensajes de desarrollo, nunca el sellado.
11
+
12
+ ## Exp. 1: K=2 reentrenado con pool variado
13
+ - Mismo protocolo que 018 (init desde K=2 de 017, maestro binder 015, CE+KL+0.5·coseno, 3,000 pasos, lr 5e-4, OneCycle, lote 12×4, 4 paráfrasis) pero el reemplazo (p=0.4 por campo) toma valores de 5 tipos: palabra minúscula (3–14 letras), nombre capitalizado, número 0–9999, frase de 2–3 palabras, pseudo-palabra multi-token.
14
+ - Semillas 17, 29, 43. Selección cada 250 pasos SOLO con validación: contrato (96) + transferencia restringida nueva (96, valores reservados de palabras/nombres/frases, disjuntos de entrenamiento y sellado).
15
+ - Reporte: exactitud por # de palabras nuevas (1/2/4) vs plantilla de texto, Wilson 95%, bootstrap pareado 10,000. Referencia: checkpoints 018 en el mismo sellado.
16
+ - Sondas sin LLM (ridge_aug λ=10, mlp_aug, mlp con objetivos barajados), entrenadas con 20,000 mensajes del pool de entrenamiento nuevo; decodificación por coseno contra el embedding léxico del valor completo. Receptor con pesos aleatorios.
17
+
18
+ ## Exp. 2: texto compacto con esfuerzo equivalente
19
+ - Formatos: `values`, `values_csv`, `kv` (+ plantilla como referencia).
20
+ - Few-shot: 2 ejemplos fijos de FIT con las 4 respuestas (prefijo compartido, se reporta su largo aparte).
21
+ - Soft prompt reparametrizado (estilo prefix-tuning): 8 vectores = base + MLP(576→3328→576), **3,842,368 parámetros** (codificador K=2: 3,842,952). Mismos datos, pasos, pérdida (con el mismo maestro), optimizador y selección por validación que K=2. 3 semillas por formato.
22
+ - Posiciones por evento = tokens del registro + 15 (+8 del soft prompt, compartibles; se reportan ambos).
23
+
24
+ ## Reglas de decisión
25
+ 1. **Transferencia**: el reentrenamiento "mejora" si la media de 3 semillas en `sealed_4_new` supera a los checkpoints 018 con IC bootstrap > 0; "alcanza al texto" si queda a ≥ −2 pp de la plantilla.
26
+ 2. **Ventaja vs texto compacto**: K=2 conserva ventaja si ningún formato compacto entrenado (soft prompt) o few-shot queda a ≥ −2 pp de K=2 en `sealed_ood` con menos de 17+8 posiciones totales; si alguno empata o gana, se dice que la ventaja de posiciones se reduce a la diferencia de posiciones medida.
27
+ 3. **Lookup**: si una sonda sin LLM iguala al LLM en `sealed_4_new`, se reporta que la información es recuperable sin LLM.
28
+ Todo resultado se reporta.
021/resultados021.md ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # CUERDAS-021 — Resultados (completo)
2
+
3
+ Fecha: 16 oct 2026 (hora CDMX). Pre-registro: `preregistro021.md` (SHA-256 1f2a27c7…). Conjunto sellado nuevo `sealed021.json` (semilla 20261017, SHA-256 a86a6a5e…72bc), verificado por el kernel y **evaluado una sola vez**. Kernel privado `cuerdas021-a` (Tesla T4). Datos: `gpu/resultsA.jsonl`, `gpu/predictionsA.jsonl`, checkpoints `gpu/k2v_s*.pt`.
4
+
5
+ Valores nuevos del sellado (8 por campo): 3 palabras, 3 nombres capitalizados, 1 frase de 2 palabras, 1 palabra multi-token. Exactitud exact-match (%), lectura de primer token; Wilson 95% y bootstrap pareado (10,000) de la media de 3 semillas.
6
+
7
+ ## Parte A — K=2 reentrenado con pool variado (palabras, nombres, números, frases, multi-token)
8
+
9
+ Selección solo por validación (pasos elegidos: 2750 / 1500 / 1250).
10
+
11
+ | Conjunto (n) | Texto plantilla | K=2 018 (s17/s29/s43) | **K=2 021 (s17/s29/s43)** | Δ 021 vs texto | Δ 021 vs 018 |
12
+ |---|---|---|---|---|---|
13
+ | sealed_id (200) | 100 [98.1, 100] | 100 / 99.5 / 97.5 | 100 / 99.0 / 99.5 | −0.50 [−1.17, 0.00] | +0.50 |
14
+ | sealed_ood (400) | 98.5 [96.8, 99.3] | 99.0 / 99.0 / 97.8 | 99.75 / 98.5 / 97.75 (media 98.67) | **+0.17 [−1.25, +1.67]** | +0.08 [−0.92, +1.08] |
15
+ | 1 nueva (192) | 96.9 [93.4, 98.6] | 87.5 / 87.5 / 75.5 | 90.6 / 88.5 / 66.1 (81.8) | −15.10 [−19.97, −10.24] | −1.74 [−4.86, +1.39] |
16
+ | 2 nuevas (192) | 97.4 [94.0, 98.9] | 62.5 / 57.8 / 45.8 | 63.0 / 56.8 / 38.0 (52.6) | −44.79 [−50.87, −38.89] | −2.78 [−6.42, +0.87] |
17
+ | 4 nuevas (192) | 98.4 [95.5, 99.5] | 28.6 / 26.6 / 25.0 | 31.2 / 30.2 / 26.6 (29.3) | **−69.10 [−75.00, −63.02]** | +2.60 [−1.04, +6.25] |
18
+
19
+ Por tipo (1 nueva, s17): palabras 97%, nombres 93%, frases 79%, multi-token 69% (texto: 100/95/92/100%).
20
+
21
+ **Regla 1: NO se cumple.** El pool más grande y variado no mejora la transferencia (4 nuevas: +2.6 pp, IC cruza 0) y queda muy lejos del texto (−69 pp). Con nombres y valores multi-token, el codificador K=2 falla fuerte. La paridad en combinaciones nuevas de palabras conocidas se mantiene (98.67 vs 98.5%).
22
+
23
+ ### Sondas sin LLM sobre las ranuras 021 (media 3 semillas)
24
+
25
+ | Decodificador | sealed_ood | 1 nueva | 2 nuevas | 4 nuevas |
26
+ |---|---|---|---|---|
27
+ | LLM congelado (K=2 021) | 98.67 | 81.8 | 52.6 | 29.3 |
28
+ | ridge_aug | 100 | 89.4 | 72.7 | **44.4** (+15.1 [+6.8, +23.4] vs LLM) |
29
+ | mlp_aug | 100 | **100** | **97.7** | 30.9 |
30
+ | mlp objetivos barajados (control) | 0 | 0 | 0 | 0 |
31
+ | Receptor pesos aleatorios (texto / K=2) | 0 / 0 | — | — | 0 / 0 |
32
+
33
+ **Regla 3: se cumple.** Sondas sin LLM igualan o superan al LLM leyendo las mismas ranuras (ridge +15 pp con 4 nuevas; MLP 97.7% vs 52.6% con 2 nuevas). La información de los valores nuevos **sí está en las ranuras**; el cuello de botella es que el LLM congelado no la lee (sobre todo nombres y multi-token). Las ranuras se comportan como un código casi léxico decodificable sin el LLM.
34
+
35
+ ## Parte B — texto compacto con esfuerzo equivalente (kernel privado `cuerdas021-b`, ≈6 h de T4)
36
+
37
+ Soft prompt reparametrizado: 8 vectores, 3,842,368 parámetros (K=2: 3,842,952). Mismos datos, pasos, pérdida, maestro y selección por validación que K=2. Few-shot: 2 ejemplos fijos (prefijo compartido de 99–131 posiciones). Posiciones por evento = registro + 15 (el soft prompt suma 8 posiciones compartidas).
38
+
39
+ | Ruta | Posiciones/evento (+compartidas) | sealed_ood | 1 nueva | 2 nuevas | 4 nuevas |
40
+ |---|---|---|---|---|---|
41
+ | **K=2 021** (media 3 semillas) | **17** | 98.67 | 81.8 | 52.6 | 29.3 |
42
+ | Plantilla texto zero-shot | 35 | 98.5 | 96.9 | 97.4 | 98.4 |
43
+ | Soft prompt + `values` (s17/s29/s43) | 19.0–20.9 (+8) | 100/100/100 | 100 | 98.6 | 95.5 (96.9/96.4/93.2) |
44
+ | **Soft prompt + `values_csv`** | **22.0–23.9 (+8)** | **100** | **100** | **100** | **100** |
45
+ | Soft prompt + `kv` | 31.0–32.0 (+8) | 100 | 100 | 100 | 100 |
46
+ | Few-shot `kv` | 31 (+123) | 91.75 | 84.4 | 86.5 | 89.6 |
47
+ | Few-shot `values_csv` | 22 (+105) | 90.75 | 50.0 | 24.5 | 4.7 |
48
+ | Few-shot `values` | 19 (+99) | 32.5 | 3.1 | 0 | 0 |
49
+ | Few-shot plantilla | 35 (+131) | 86.5 | 66.7 | 58.3 | 38.5 |
50
+
51
+ Soft prompt `values_csv` − K=2 (bootstrap pareado, media de semillas): sealed_ood +1.33 [+0.58, +2.17]; 1 nueva +18.2 [+14.4, +22.4]; 2 nuevas +47.4 [+41.7, +53.1]; 4 nuevas **+70.7 [+64.9, +76.2]** pp.
52
+
53
+ **Regla 2: K=2 NO conserva ventaja.** Con el mismo presupuesto (parámetros, datos, pasos, maestro), un soft prompt delante de texto compacto con comas da **100% en todo** con 22 posiciones por evento (+8 compartidas), y con solo valores llega a 100% en conocidos y 93–97% con 4 nuevas a 19 posiciones. K=2 ahorra apenas 2–5 posiciones por evento (17 vs 19–22) y pierde hasta 70 pp con valores nuevos. Few-shot sin entrenar no alcanza.
54
+
55
+ ## Veredicto honesto
56
+ - La ventaja de Cuerdas desaparece frente a un baseline justo. El "51% menos posiciones" era contra una plantilla verbosa; contra texto compacto entrenado el ahorro es de ~2–5 posiciones por evento (−11% a −23%), a cambio de perder mucha exactitud con valores nuevos.
57
+ - La paridad en combinaciones de palabras conocidas es real (dos tests sellados), pero el texto compacto entrenado la iguala o supera (100%).
58
+ - Las ranuras sí contienen la información (las sondas la leen), pero el receptor congelado no las lee bien; el texto compacto sí.
59
+ - Conclusión: como método, Cuerdas no tiene ventaja práctica en este escenario. Como reporte técnico, el resultado negativo, bien documentado y pre-registrado, es valioso.
021/resultsA.jsonl ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"sealed_sha256": "a86a6a5ee375ae67a3e79f969c9553b0e5548ed3b0e80e5315862e6552ae72bc", "torch": "2.11.0+cu128", "pool_sizes": {"words": 20654, "names": 8188, "subs": 7189}}
2
+ {"sys": "text_template", "sealed_id": 1.0, "sealed_ood": 0.985, "sealed_1_new": 0.96875, "sealed_2_new": 0.9739583333333334, "sealed_4_new": 0.984375}
3
+ {"sys": "k2_018_s17", "sealed_id": 1.0, "sealed_ood": 0.99, "sealed_1_new": 0.875, "sealed_2_new": 0.625, "sealed_4_new": 0.2864583333333333}
4
+ {"sys": "k2_018_s29", "sealed_id": 0.995, "sealed_ood": 0.99, "sealed_1_new": 0.875, "sealed_2_new": 0.578125, "sealed_4_new": 0.265625}
5
+ {"sys": "k2_018_s43", "sealed_id": 0.975, "sealed_ood": 0.9775, "sealed_1_new": 0.7552083333333334, "sealed_2_new": 0.4583333333333333, "sealed_4_new": 0.25}
6
+ {"run": "k2_021_s17", "step": 0, "val": 1.1145833333333333}
7
+ {"run": "k2_021_s17", "step": 250, "val": 1.0833333333333333, "loss": 4.719815254211426, "s": 51}
8
+ {"run": "k2_021_s17", "step": 500, "val": 1.0833333333333333, "loss": 5.05856990814209, "s": 101}
9
+ {"run": "k2_021_s17", "step": 750, "val": 1.0625, "loss": 4.387829780578613, "s": 154}
10
+ {"run": "k2_021_s17", "step": 1000, "val": 1.09375, "loss": 4.11569881439209, "s": 206}
11
+ {"run": "k2_021_s17", "step": 1250, "val": 1.1354166666666667, "loss": 4.992859363555908, "s": 259}
12
+ {"run": "k2_021_s17", "step": 1500, "val": 1.1458333333333335, "loss": 4.392324447631836, "s": 311}
13
+ {"run": "k2_021_s17", "step": 1750, "val": 1.1354166666666667, "loss": 4.7229413986206055, "s": 364}
14
+ {"run": "k2_021_s17", "step": 2000, "val": 1.1458333333333333, "loss": 4.124819278717041, "s": 417}
15
+ {"run": "k2_021_s17", "step": 2250, "val": 1.1666666666666667, "loss": 3.794137477874756, "s": 469}
16
+ {"run": "k2_021_s17", "step": 2500, "val": 1.1666666666666667, "loss": 2.527881622314453, "s": 521}
17
+ {"run": "k2_021_s17", "step": 2750, "val": 1.1770833333333333, "loss": 3.1677093505859375, "s": 573}
18
+ {"run": "k2_021_s17", "step": 3000, "val": 1.1666666666666667, "loss": 1.943210482597351, "s": 626}
19
+ {"selected": "k2_021_s17", "step": 2750, "val": 1.1770833333333333}
20
+ {"sys": "k2_021_s17", "sealed_id": 1.0, "sealed_ood": 0.9975, "sealed_1_new": 0.90625, "sealed_2_new": 0.6302083333333334, "sealed_4_new": 0.3125}
21
+ {"run": "k2_021_s29", "step": 0, "val": 1.0729166666666667}
22
+ {"run": "k2_021_s29", "step": 250, "val": 1.0520833333333335, "loss": 4.769656181335449, "s": 52}
23
+ {"run": "k2_021_s29", "step": 500, "val": 1.0833333333333333, "loss": 4.504295825958252, "s": 105}
24
+ {"run": "k2_021_s29", "step": 750, "val": 1.15625, "loss": 3.0690648555755615, "s": 157}
25
+ {"run": "k2_021_s29", "step": 1000, "val": 1.125, "loss": 2.2323856353759766, "s": 210}
26
+ {"run": "k2_021_s29", "step": 1250, "val": 1.1666666666666667, "loss": 4.31823205947876, "s": 263}
27
+ {"run": "k2_021_s29", "step": 1500, "val": 1.1770833333333333, "loss": 4.3195061683654785, "s": 315}
28
+ {"run": "k2_021_s29", "step": 1750, "val": 1.1354166666666667, "loss": 3.4602582454681396, "s": 367}
29
+ {"run": "k2_021_s29", "step": 2000, "val": 1.1041666666666667, "loss": 3.7912955284118652, "s": 419}
30
+ {"run": "k2_021_s29", "step": 2250, "val": 1.09375, "loss": 3.81158709526062, "s": 472}
31
+ {"run": "k2_021_s29", "step": 2500, "val": 1.1354166666666667, "loss": 2.8178346157073975, "s": 524}
32
+ {"run": "k2_021_s29", "step": 2750, "val": 1.1145833333333333, "loss": 3.228869676589966, "s": 577}
33
+ {"run": "k2_021_s29", "step": 3000, "val": 1.1145833333333333, "loss": 3.688642978668213, "s": 629}
34
+ {"selected": "k2_021_s29", "step": 1500, "val": 1.1770833333333333}
35
+ {"sys": "k2_021_s29", "sealed_id": 0.99, "sealed_ood": 0.985, "sealed_1_new": 0.8854166666666666, "sealed_2_new": 0.5677083333333334, "sealed_4_new": 0.3020833333333333}
36
+ {"run": "k2_021_s43", "step": 0, "val": 1.1041666666666667}
37
+ {"run": "k2_021_s43", "step": 250, "val": 1.0625, "loss": 3.4185612201690674, "s": 52}
38
+ {"run": "k2_021_s43", "step": 500, "val": 1.0, "loss": 4.755449295043945, "s": 104}
39
+ {"run": "k2_021_s43", "step": 750, "val": 1.0520833333333335, "loss": 4.684804916381836, "s": 157}
40
+ {"run": "k2_021_s43", "step": 1000, "val": 1.09375, "loss": 4.735645771026611, "s": 210}
41
+ {"run": "k2_021_s43", "step": 1250, "val": 1.1458333333333333, "loss": 3.7620046138763428, "s": 263}
42
+ {"run": "k2_021_s43", "step": 1500, "val": 1.1354166666666667, "loss": 4.758123397827148, "s": 315}
43
+ {"run": "k2_021_s43", "step": 1750, "val": 1.1145833333333333, "loss": 4.265966892242432, "s": 368}
44
+ {"run": "k2_021_s43", "step": 2000, "val": 1.0625, "loss": 3.849872589111328, "s": 419}
45
+ {"run": "k2_021_s43", "step": 2250, "val": 1.1041666666666665, "loss": 4.972720146179199, "s": 471}
46
+ {"run": "k2_021_s43", "step": 2500, "val": 1.125, "loss": 2.9099667072296143, "s": 523}
47
+ {"run": "k2_021_s43", "step": 2750, "val": 1.125, "loss": 3.000356674194336, "s": 575}
48
+ {"run": "k2_021_s43", "step": 3000, "val": 1.125, "loss": 2.915379762649536, "s": 626}
49
+ {"selected": "k2_021_s43", "step": 1250, "val": 1.1458333333333333}
50
+ {"sys": "k2_021_s43", "sealed_id": 0.995, "sealed_ood": 0.9775, "sealed_1_new": 0.6614583333333334, "sealed_2_new": 0.3802083333333333, "sealed_4_new": 0.265625}
51
+ {"probe": "ridge_aug", "seed": 17, "set": "sealed_ood", "acc": 1.0}
52
+ {"probe": "ridge_aug", "seed": 17, "set": "sealed_1_new", "acc": 0.890625}
53
+ {"probe": "ridge_aug", "seed": 17, "set": "sealed_2_new", "acc": 0.7135416666666666}
54
+ {"probe": "ridge_aug", "seed": 17, "set": "sealed_4_new", "acc": 0.4322916666666667}
55
+ {"probe": "mlp_aug", "seed": 17, "set": "sealed_ood", "acc": 1.0}
56
+ {"probe": "mlp_aug", "seed": 17, "set": "sealed_1_new", "acc": 1.0}
57
+ {"probe": "mlp_aug", "seed": 17, "set": "sealed_2_new", "acc": 0.984375}
58
+ {"probe": "mlp_aug", "seed": 17, "set": "sealed_4_new", "acc": 0.24479166666666666}
59
+ {"probe": "mlp_ctrl_shuffled", "seed": 17, "set": "sealed_ood", "acc": 0.0}
60
+ {"probe": "mlp_ctrl_shuffled", "seed": 17, "set": "sealed_1_new", "acc": 0.0}
61
+ {"probe": "mlp_ctrl_shuffled", "seed": 17, "set": "sealed_2_new", "acc": 0.0}
62
+ {"probe": "mlp_ctrl_shuffled", "seed": 17, "set": "sealed_4_new", "acc": 0.0}
63
+ {"probe": "ridge_aug", "seed": 29, "set": "sealed_ood", "acc": 1.0}
64
+ {"probe": "ridge_aug", "seed": 29, "set": "sealed_1_new", "acc": 0.9114583333333334}
65
+ {"probe": "ridge_aug", "seed": 29, "set": "sealed_2_new", "acc": 0.7604166666666666}
66
+ {"probe": "ridge_aug", "seed": 29, "set": "sealed_4_new", "acc": 0.421875}
67
+ {"probe": "mlp_aug", "seed": 29, "set": "sealed_ood", "acc": 1.0}
68
+ {"probe": "mlp_aug", "seed": 29, "set": "sealed_1_new", "acc": 1.0}
69
+ {"probe": "mlp_aug", "seed": 29, "set": "sealed_2_new", "acc": 0.984375}
70
+ {"probe": "mlp_aug", "seed": 29, "set": "sealed_4_new", "acc": 0.359375}
71
+ {"probe": "mlp_ctrl_shuffled", "seed": 29, "set": "sealed_ood", "acc": 0.0}
72
+ {"probe": "mlp_ctrl_shuffled", "seed": 29, "set": "sealed_1_new", "acc": 0.0}
73
+ {"probe": "mlp_ctrl_shuffled", "seed": 29, "set": "sealed_2_new", "acc": 0.0}
74
+ {"probe": "mlp_ctrl_shuffled", "seed": 29, "set": "sealed_4_new", "acc": 0.0}
75
+ {"probe": "ridge_aug", "seed": 43, "set": "sealed_ood", "acc": 1.0}
76
+ {"probe": "ridge_aug", "seed": 43, "set": "sealed_1_new", "acc": 0.8802083333333334}
77
+ {"probe": "ridge_aug", "seed": 43, "set": "sealed_2_new", "acc": 0.7083333333333334}
78
+ {"probe": "ridge_aug", "seed": 43, "set": "sealed_4_new", "acc": 0.4791666666666667}
79
+ {"probe": "mlp_aug", "seed": 43, "set": "sealed_ood", "acc": 1.0}
80
+ {"probe": "mlp_aug", "seed": 43, "set": "sealed_1_new", "acc": 1.0}
81
+ {"probe": "mlp_aug", "seed": 43, "set": "sealed_2_new", "acc": 0.9635416666666666}
82
+ {"probe": "mlp_aug", "seed": 43, "set": "sealed_4_new", "acc": 0.3229166666666667}
83
+ {"probe": "mlp_ctrl_shuffled", "seed": 43, "set": "sealed_ood", "acc": 0.0}
84
+ {"probe": "mlp_ctrl_shuffled", "seed": 43, "set": "sealed_1_new", "acc": 0.0}
85
+ {"probe": "mlp_ctrl_shuffled", "seed": 43, "set": "sealed_2_new", "acc": 0.0}
86
+ {"probe": "mlp_ctrl_shuffled", "seed": 43, "set": "sealed_4_new", "acc": 0.0}
87
+ {"control": "random_receiver", "set": "sealed_ood", "text_template": 0.0, "k2_021_s17": 0.0, "k2_021_s29": 0.0, "k2_021_s43": 0.0}
88
+ {"control": "random_receiver", "set": "sealed_4_new", "text_template": 0.0, "k2_021_s17": 0.0, "k2_021_s29": 0.0, "k2_021_s43": 0.0}
021/resultsB.jsonl ADDED
@@ -0,0 +1,134 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"positions_per_event": "values", "sealed_id": 19.0, "sealed_ood": 19.0, "sealed_1_new": 19.4, "sealed_2_new": 19.95, "sealed_4_new": 20.85}
2
+ {"positions_per_event": "values_csv", "sealed_id": 22.0, "sealed_ood": 22.0, "sealed_1_new": 22.4, "sealed_2_new": 22.95, "sealed_4_new": 23.85}
3
+ {"positions_per_event": "kv", "sealed_id": 31.0, "sealed_ood": 31.0, "sealed_1_new": 31.21, "sealed_2_new": 31.41, "sealed_4_new": 32.01}
4
+ {"positions_per_event": "template", "sealed_id": 35.0, "sealed_ood": 35.0, "sealed_1_new": 35.21, "sealed_2_new": 35.41, "sealed_4_new": 36.01}
5
+ {"sys": "fewshot_values", "shared_prefix_positions": 99, "sealed_id": 0.345, "sealed_ood": 0.325, "sealed_1_new": 0.03125, "sealed_2_new": 0.0, "sealed_4_new": 0.0}
6
+ {"sys": "fewshot_values_csv", "shared_prefix_positions": 105, "sealed_id": 0.89, "sealed_ood": 0.9075, "sealed_1_new": 0.5, "sealed_2_new": 0.24479166666666666, "sealed_4_new": 0.046875}
7
+ {"sys": "fewshot_kv", "shared_prefix_positions": 123, "sealed_id": 0.945, "sealed_ood": 0.9175, "sealed_1_new": 0.84375, "sealed_2_new": 0.8645833333333334, "sealed_4_new": 0.8958333333333334}
8
+ {"sys": "fewshot_template", "shared_prefix_positions": 131, "sealed_id": 0.9, "sealed_ood": 0.865, "sealed_1_new": 0.6666666666666666, "sealed_2_new": 0.5833333333333334, "sealed_4_new": 0.3854166666666667}
9
+ {"run": "sp_values_s17", "step": 0, "val": 0.0}
10
+ {"run": "sp_values_s17", "step": 250, "val": 1.6145833333333333, "loss": 3.044680595397949, "s": 87}
11
+ {"run": "sp_values_s17", "step": 500, "val": 1.6666666666666667, "loss": 3.143272638320923, "s": 174}
12
+ {"run": "sp_values_s17", "step": 750, "val": 1.6770833333333335, "loss": 2.6209170818328857, "s": 261}
13
+ {"run": "sp_values_s17", "step": 1000, "val": 1.6041666666666665, "loss": 3.025043487548828, "s": 348}
14
+ {"run": "sp_values_s17", "step": 1250, "val": 1.7083333333333335, "loss": 2.137320041656494, "s": 435}
15
+ {"run": "sp_values_s17", "step": 1500, "val": 1.71875, "loss": 1.954566240310669, "s": 523}
16
+ {"run": "sp_values_s17", "step": 1750, "val": 1.6354166666666665, "loss": 3.326815128326416, "s": 610}
17
+ {"run": "sp_values_s17", "step": 2000, "val": 1.7291666666666665, "loss": 2.481618642807007, "s": 699}
18
+ {"run": "sp_values_s17", "step": 2250, "val": 1.7083333333333335, "loss": 2.0035409927368164, "s": 786}
19
+ {"run": "sp_values_s17", "step": 2500, "val": 1.7395833333333335, "loss": 1.278429627418518, "s": 872}
20
+ {"run": "sp_values_s17", "step": 2750, "val": 1.7708333333333335, "loss": 1.8047574758529663, "s": 960}
21
+ {"run": "sp_values_s17", "step": 3000, "val": 1.7395833333333335, "loss": 1.2187349796295166, "s": 1047}
22
+ {"sys": "sp_values_s17", "params": 3842368, "selected_step": 2750, "val": 1.7708333333333335, "soft_prompt_positions": 8, "sealed_id": 1.0, "sealed_ood": 1.0, "sealed_1_new": 1.0, "sealed_2_new": 0.984375, "sealed_4_new": 0.96875}
23
+ {"run": "sp_values_s29", "step": 0, "val": 0.0}
24
+ {"run": "sp_values_s29", "step": 250, "val": 1.375, "loss": 4.450610160827637, "s": 87}
25
+ {"run": "sp_values_s29", "step": 500, "val": 1.6145833333333335, "loss": 2.346332550048828, "s": 174}
26
+ {"run": "sp_values_s29", "step": 750, "val": 1.6145833333333335, "loss": 2.190868854522705, "s": 261}
27
+ {"run": "sp_values_s29", "step": 1000, "val": 1.6666666666666665, "loss": 1.3257291316986084, "s": 348}
28
+ {"run": "sp_values_s29", "step": 1250, "val": 1.6458333333333335, "loss": 2.51035737991333, "s": 436}
29
+ {"run": "sp_values_s29", "step": 1500, "val": 1.6875, "loss": 2.5555827617645264, "s": 523}
30
+ {"run": "sp_values_s29", "step": 1750, "val": 1.6041666666666665, "loss": 1.5838260650634766, "s": 609}
31
+ {"run": "sp_values_s29", "step": 2000, "val": 1.6770833333333335, "loss": 2.4831151962280273, "s": 697}
32
+ {"run": "sp_values_s29", "step": 2250, "val": 1.7083333333333335, "loss": 1.9796416759490967, "s": 784}
33
+ {"run": "sp_values_s29", "step": 2500, "val": 1.6666666666666665, "loss": 1.4970152378082275, "s": 871}
34
+ {"run": "sp_values_s29", "step": 2750, "val": 1.6875, "loss": 2.13800311088562, "s": 958}
35
+ {"run": "sp_values_s29", "step": 3000, "val": 1.6875, "loss": 2.1700429916381836, "s": 1044}
36
+ {"sys": "sp_values_s29", "params": 3842368, "selected_step": 2250, "val": 1.7083333333333335, "soft_prompt_positions": 8, "sealed_id": 1.0, "sealed_ood": 1.0, "sealed_1_new": 1.0, "sealed_2_new": 0.9895833333333334, "sealed_4_new": 0.9635416666666666}
37
+ {"run": "sp_values_s43", "step": 0, "val": 0.0}
38
+ {"run": "sp_values_s43", "step": 250, "val": 1.4583333333333333, "loss": 2.9020471572875977, "s": 87}
39
+ {"run": "sp_values_s43", "step": 500, "val": 1.46875, "loss": 2.5136404037475586, "s": 173}
40
+ {"run": "sp_values_s43", "step": 750, "val": 1.625, "loss": 2.120654821395874, "s": 262}
41
+ {"run": "sp_values_s43", "step": 1000, "val": 1.6770833333333335, "loss": 3.3165745735168457, "s": 349}
42
+ {"run": "sp_values_s43", "step": 1250, "val": 1.78125, "loss": 2.017547607421875, "s": 437}
43
+ {"run": "sp_values_s43", "step": 1500, "val": 1.78125, "loss": 2.347735643386841, "s": 524}
44
+ {"run": "sp_values_s43", "step": 1750, "val": 1.7708333333333335, "loss": 2.394887924194336, "s": 612}
45
+ {"run": "sp_values_s43", "step": 2000, "val": 1.625, "loss": 1.9778445959091187, "s": 699}
46
+ {"run": "sp_values_s43", "step": 2250, "val": 1.78125, "loss": 2.360875129699707, "s": 785}
47
+ {"run": "sp_values_s43", "step": 2500, "val": 1.7916666666666665, "loss": 1.7670954465866089, "s": 872}
48
+ {"run": "sp_values_s43", "step": 2750, "val": 1.75, "loss": 2.02799654006958, "s": 959}
49
+ {"run": "sp_values_s43", "step": 3000, "val": 1.7395833333333335, "loss": 1.7585619688034058, "s": 1046}
50
+ {"sys": "sp_values_s43", "params": 3842368, "selected_step": 2500, "val": 1.7916666666666665, "soft_prompt_positions": 8, "sealed_id": 1.0, "sealed_ood": 1.0, "sealed_1_new": 1.0, "sealed_2_new": 0.984375, "sealed_4_new": 0.9322916666666666}
51
+ {"run": "sp_values_csv_s17", "step": 0, "val": 0.0}
52
+ {"run": "sp_values_csv_s17", "step": 250, "val": 1.90625, "loss": 2.3863179683685303, "s": 95}
53
+ {"run": "sp_values_csv_s17", "step": 500, "val": 1.9583333333333335, "loss": 2.344421863555908, "s": 189}
54
+ {"run": "sp_values_csv_s17", "step": 750, "val": 1.9791666666666665, "loss": 1.9774900674819946, "s": 283}
55
+ {"run": "sp_values_csv_s17", "step": 1000, "val": 1.9583333333333335, "loss": 1.8494818210601807, "s": 377}
56
+ {"run": "sp_values_csv_s17", "step": 1250, "val": 1.9479166666666665, "loss": 1.641819715499878, "s": 472}
57
+ {"run": "sp_values_csv_s17", "step": 1500, "val": 1.9895833333333335, "loss": 1.3760128021240234, "s": 567}
58
+ {"run": "sp_values_csv_s17", "step": 1750, "val": 1.9895833333333335, "loss": 1.9606835842132568, "s": 662}
59
+ {"run": "sp_values_csv_s17", "step": 2000, "val": 2.0, "loss": 1.9302983283996582, "s": 758}
60
+ {"run": "sp_values_csv_s17", "step": 2250, "val": 1.9895833333333335, "loss": 1.5998188257217407, "s": 852}
61
+ {"run": "sp_values_csv_s17", "step": 2500, "val": 1.9895833333333335, "loss": 1.07892906665802, "s": 946}
62
+ {"run": "sp_values_csv_s17", "step": 2750, "val": 2.0, "loss": 1.3307654857635498, "s": 1040}
63
+ {"run": "sp_values_csv_s17", "step": 3000, "val": 2.0, "loss": 0.9787193536758423, "s": 1134}
64
+ {"sys": "sp_values_csv_s17", "params": 3842368, "selected_step": 2000, "val": 2.0, "soft_prompt_positions": 8, "sealed_id": 1.0, "sealed_ood": 1.0, "sealed_1_new": 1.0, "sealed_2_new": 1.0, "sealed_4_new": 1.0}
65
+ {"run": "sp_values_csv_s29", "step": 0, "val": 0.0}
66
+ {"run": "sp_values_csv_s29", "step": 250, "val": 1.9583333333333335, "loss": 2.902418851852417, "s": 94}
67
+ {"run": "sp_values_csv_s29", "step": 500, "val": 1.9375, "loss": 1.829451084136963, "s": 190}
68
+ {"run": "sp_values_csv_s29", "step": 750, "val": 1.9375, "loss": 1.5734684467315674, "s": 283}
69
+ {"run": "sp_values_csv_s29", "step": 1000, "val": 1.96875, "loss": 0.8627220988273621, "s": 378}
70
+ {"run": "sp_values_csv_s29", "step": 1250, "val": 1.9791666666666665, "loss": 1.8343243598937988, "s": 473}
71
+ {"run": "sp_values_csv_s29", "step": 1500, "val": 1.9895833333333335, "loss": 1.6268024444580078, "s": 567}
72
+ {"run": "sp_values_csv_s29", "step": 1750, "val": 1.96875, "loss": 1.224697232246399, "s": 661}
73
+ {"run": "sp_values_csv_s29", "step": 2000, "val": 1.9583333333333335, "loss": 1.8627893924713135, "s": 756}
74
+ {"run": "sp_values_csv_s29", "step": 2250, "val": 1.9583333333333335, "loss": 1.6173484325408936, "s": 851}
75
+ {"run": "sp_values_csv_s29", "step": 2500, "val": 1.9791666666666665, "loss": 1.2948013544082642, "s": 945}
76
+ {"run": "sp_values_csv_s29", "step": 2750, "val": 1.96875, "loss": 1.518986463546753, "s": 1040}
77
+ {"run": "sp_values_csv_s29", "step": 3000, "val": 1.96875, "loss": 1.6026052236557007, "s": 1134}
78
+ {"sys": "sp_values_csv_s29", "params": 3842368, "selected_step": 1500, "val": 1.9895833333333335, "soft_prompt_positions": 8, "sealed_id": 1.0, "sealed_ood": 1.0, "sealed_1_new": 1.0, "sealed_2_new": 1.0, "sealed_4_new": 1.0}
79
+ {"run": "sp_values_csv_s43", "step": 0, "val": 0.0}
80
+ {"run": "sp_values_csv_s43", "step": 250, "val": 1.8854166666666665, "loss": 2.186886787414551, "s": 95}
81
+ {"run": "sp_values_csv_s43", "step": 500, "val": 1.96875, "loss": 1.741999864578247, "s": 189}
82
+ {"run": "sp_values_csv_s43", "step": 750, "val": 1.8645833333333335, "loss": 1.5374534130096436, "s": 284}
83
+ {"run": "sp_values_csv_s43", "step": 1000, "val": 1.9583333333333335, "loss": 2.0583114624023438, "s": 379}
84
+ {"run": "sp_values_csv_s43", "step": 1250, "val": 1.9166666666666665, "loss": 1.2399407625198364, "s": 475}
85
+ {"run": "sp_values_csv_s43", "step": 1500, "val": 1.96875, "loss": 1.6992583274841309, "s": 570}
86
+ {"run": "sp_values_csv_s43", "step": 1750, "val": 1.9895833333333335, "loss": 1.7322700023651123, "s": 665}
87
+ {"run": "sp_values_csv_s43", "step": 2000, "val": 1.9583333333333335, "loss": 1.5983898639678955, "s": 759}
88
+ {"run": "sp_values_csv_s43", "step": 2250, "val": 1.96875, "loss": 1.6161710023880005, "s": 853}
89
+ {"run": "sp_values_csv_s43", "step": 2500, "val": 1.96875, "loss": 1.3451036214828491, "s": 947}
90
+ {"run": "sp_values_csv_s43", "step": 2750, "val": 1.9791666666666665, "loss": 1.2667261362075806, "s": 1042}
91
+ {"run": "sp_values_csv_s43", "step": 3000, "val": 1.96875, "loss": 1.4136821031570435, "s": 1137}
92
+ {"sys": "sp_values_csv_s43", "params": 3842368, "selected_step": 1750, "val": 1.9895833333333335, "soft_prompt_positions": 8, "sealed_id": 1.0, "sealed_ood": 1.0, "sealed_1_new": 1.0, "sealed_2_new": 1.0, "sealed_4_new": 1.0}
93
+ {"run": "sp_kv_s17", "step": 0, "val": 0.0625}
94
+ {"run": "sp_kv_s17", "step": 250, "val": 2.0, "loss": 1.9125851392745972, "s": 115}
95
+ {"run": "sp_kv_s17", "step": 500, "val": 2.0, "loss": 2.0021212100982666, "s": 230}
96
+ {"run": "sp_kv_s17", "step": 750, "val": 2.0, "loss": 1.5451703071594238, "s": 346}
97
+ {"run": "sp_kv_s17", "step": 1000, "val": 2.0, "loss": 1.4742144346237183, "s": 461}
98
+ {"run": "sp_kv_s17", "step": 1250, "val": 2.0, "loss": 1.4599279165267944, "s": 577}
99
+ {"run": "sp_kv_s17", "step": 1500, "val": 2.0, "loss": 1.2338306903839111, "s": 693}
100
+ {"run": "sp_kv_s17", "step": 1750, "val": 2.0, "loss": 1.8669570684432983, "s": 809}
101
+ {"run": "sp_kv_s17", "step": 2000, "val": 2.0, "loss": 1.7696266174316406, "s": 926}
102
+ {"run": "sp_kv_s17", "step": 2250, "val": 2.0, "loss": 1.350313425064087, "s": 1041}
103
+ {"run": "sp_kv_s17", "step": 2500, "val": 2.0, "loss": 0.9131432175636292, "s": 1156}
104
+ {"run": "sp_kv_s17", "step": 2750, "val": 2.0, "loss": 1.2375553846359253, "s": 1271}
105
+ {"run": "sp_kv_s17", "step": 3000, "val": 2.0, "loss": 0.8844823837280273, "s": 1386}
106
+ {"sys": "sp_kv_s17", "params": 3842368, "selected_step": 250, "val": 2.0, "soft_prompt_positions": 8, "sealed_id": 1.0, "sealed_ood": 1.0, "sealed_1_new": 1.0, "sealed_2_new": 1.0, "sealed_4_new": 1.0}
107
+ {"run": "sp_kv_s29", "step": 0, "val": 0.07291666666666667}
108
+ {"run": "sp_kv_s29", "step": 250, "val": 1.9895833333333335, "loss": 2.1503524780273438, "s": 115}
109
+ {"run": "sp_kv_s29", "step": 500, "val": 2.0, "loss": 1.5459184646606445, "s": 231}
110
+ {"run": "sp_kv_s29", "step": 750, "val": 1.9895833333333335, "loss": 1.2727928161621094, "s": 347}
111
+ {"run": "sp_kv_s29", "step": 1000, "val": 2.0, "loss": 0.745368242263794, "s": 462}
112
+ {"run": "sp_kv_s29", "step": 1250, "val": 2.0, "loss": 1.6049551963806152, "s": 579}
113
+ {"run": "sp_kv_s29", "step": 1500, "val": 2.0, "loss": 1.5594961643218994, "s": 694}
114
+ {"run": "sp_kv_s29", "step": 1750, "val": 2.0, "loss": 1.1384079456329346, "s": 809}
115
+ {"run": "sp_kv_s29", "step": 2000, "val": 1.96875, "loss": 1.4778153896331787, "s": 924}
116
+ {"run": "sp_kv_s29", "step": 2250, "val": 2.0, "loss": 1.5377285480499268, "s": 1040}
117
+ {"run": "sp_kv_s29", "step": 2500, "val": 2.0, "loss": 1.1550800800323486, "s": 1156}
118
+ {"run": "sp_kv_s29", "step": 2750, "val": 2.0, "loss": 1.3703510761260986, "s": 1271}
119
+ {"run": "sp_kv_s29", "step": 3000, "val": 2.0, "loss": 1.3216395378112793, "s": 1386}
120
+ {"sys": "sp_kv_s29", "params": 3842368, "selected_step": 500, "val": 2.0, "soft_prompt_positions": 8, "sealed_id": 1.0, "sealed_ood": 1.0, "sealed_1_new": 1.0, "sealed_2_new": 1.0, "sealed_4_new": 1.0}
121
+ {"run": "sp_kv_s43", "step": 0, "val": 0.07291666666666667}
122
+ {"run": "sp_kv_s43", "step": 250, "val": 2.0, "loss": 1.6502081155776978, "s": 116}
123
+ {"run": "sp_kv_s43", "step": 500, "val": 2.0, "loss": 1.304010272026062, "s": 231}
124
+ {"run": "sp_kv_s43", "step": 750, "val": 2.0, "loss": 1.260686993598938, "s": 347}
125
+ {"run": "sp_kv_s43", "step": 1000, "val": 2.0, "loss": 1.5032280683517456, "s": 463}
126
+ {"run": "sp_kv_s43", "step": 1250, "val": 2.0, "loss": 1.139223337173462, "s": 580}
127
+ {"run": "sp_kv_s43", "step": 1500, "val": 2.0, "loss": 1.4122791290283203, "s": 695}
128
+ {"run": "sp_kv_s43", "step": 1750, "val": 2.0, "loss": 1.3497402667999268, "s": 811}
129
+ {"run": "sp_kv_s43", "step": 2000, "val": 2.0, "loss": 1.427781581878662, "s": 927}
130
+ {"run": "sp_kv_s43", "step": 2250, "val": 2.0, "loss": 1.3728115558624268, "s": 1041}
131
+ {"run": "sp_kv_s43", "step": 2500, "val": 2.0, "loss": 1.1903347969055176, "s": 1157}
132
+ {"run": "sp_kv_s43", "step": 2750, "val": 2.0, "loss": 1.029153823852539, "s": 1273}
133
+ {"run": "sp_kv_s43", "step": 3000, "val": 2.0, "loss": 1.0028929710388184, "s": 1388}
134
+ {"sys": "sp_kv_s43", "params": 3842368, "selected_step": 250, "val": 2.0, "soft_prompt_positions": 8, "sealed_id": 1.0, "sealed_ood": 1.0, "sealed_1_new": 1.0, "sealed_2_new": 1.0, "sealed_4_new": 1.0}
021/sealed021.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"exclude_token_ids": [477, 2269, 4187, 5649, 8223, 9685, 10373, 10848, 14127, 15864, 15882, 17744, 19583, 19842, 19853, 19965, 21450, 22736, 24332, 24334, 24459, 24519, 25322, 25360, 25459, 25527, 29795, 30747, 30967, 32142, 33430, 34171, 36526, 36976, 37622, 38509, 43642, 46191, 46832, 47575], "new_values": [["tube", "remotely", "calming", "Insect", "Benin", "Mom", "purchases nat", "comfortablyifera"], ["interpolation", "heroine", "equine", "Bodies", "Fischer", "Gaussian", "electrolyte benefited", "geospatialard"], ["sarc", "downloads", "tackle", "Archived", "Filipino", "Fred", "linestyle seen", "identifierbud"], ["aspire", "careful", "manners", "Answer", "Director", "Prepar", "choice suffrage", "brickbc"]], "sealed_1_new": [["report", "customer", "active", "manners"], ["query", "Gaussian", "active", "minute"], ["delete", "machine", "closed", "choice suffrage"], ["create", "valve", "running", "Prepar"], ["update", "machine", "tackle", "today"], ["stop", "equine", "paid", "hour"], ["Benin", "motor", "ready", "hour"], ["report", "motor", "open", "aspire"], ["write", "electrolyte benefited", "active", "day"], ["create", "Fischer", "high", "hour"], ["purchases nat", "valve", "ready", "now"], ["read", "heroine", "pending", "day"], ["query", "electrolyte benefited", "idle", "week"], ["report", "Bodies", "active", "minute"], ["execute", "motor", "paid", "Director"], ["Insect", "motor", "running", "none"], ["tube", "pump", "ready", "hour"], ["read", "invoice", "tackle", "today"], ["stop", "Gaussian", "open", "minute"], ["notify", "customer", "pending", "aspire"], ["write", "Gaussian", "failed", "hour"], ["write", "valve", "complete", "Director"], ["write", "heroine", "high", "hour"], ["query", "file", "linestyle seen", "today"], ["update", "order", "open", "Director"], ["write", "invoice", "paid", "aspire"], ["stop", "email", "active", "aspire"], ["Insect", "inventory", "high", "today"], ["stop", "motor", "blocked", "manners"], ["read", "geospatialard", "active", "year"], ["Benin", "motor", "blocked", "none"], ["stop", "valve", "high", "Prepar"], ["Insect", "meeting", "overload", "month"], ["purchases nat", "machine", "paid", "today"], ["Insect", "meeting", "failed", "hour"], ["create", "heroine", "complete", "year"], ["Insect", "meeting", "failed", "today"], ["update", "motor", "closed", "Prepar"], ["write", "machine", "ready", "choice suffrage"], ["update", "file", "blocked", "careful"], ["write", "file", "overload", "careful"], ["stop", "equine", "overload", "month"], ["create", "Gaussian", "high", "hour"], ["read", "job", "tackle", "day"], ["stop", "heroine", "overload", "tomorrow"], ["create", "Fischer", "ready", "now"], ["read", "customer", "linestyle seen", "tomorrow"], ["read", "geospatialard", "closed", "day"], ["stop", "Bodies", "complete", "today"], ["comfortablyifera", "valve", "active", "day"], ["notify", "heroine", "none", "hour"], ["purchases nat", "customer", "failed", "today"], ["execute", "order", "sarc", "minute"], ["stop", "invoice", "Archived", "year"], ["execute", "Bodies", "ready", "hour"], ["calming", "inventory", "open", "tomorrow"], ["check", "file", "linestyle seen", "month"], ["query", "file", "Filipino", "minute"], ["tube", "motor", "low", "year"], ["notify", "job", "active", "aspire"], ["notify", "pump", "active", "careful"], ["start", "invoice", "tackle", "tomorrow"], ["purchases nat", "meeting", "normal", "year"], ["write", "customer", "Filipino", "year"], ["check", "customer", "tackle", "year"], ["delete", "motor", "tackle", "year"], ["query", "file", "overload", "manners"], ["notify", "pump", "idle", "careful"], ["read", "valve", "open", "Director"], ["check", "motor", "Filipino", "week"], ["Mom", "valve", "ready", "month"], ["check", "heroine", "running", "hour"], ["start", "customer", "Archived", "day"], ["read", "pump", "pending", "brickbc"], ["read", "Bodies", "active", "year"], ["write", "valve", "closed", "choice suffrage"], ["comfortablyifera", "customer", "ready", "month"], ["check", "motor", "idle", "careful"], ["start", "job", "Filipino", "year"], ["update", "file", "high", "Director"], ["execute", "interpolation", "active", "now"], ["execute", "invoice", "Filipino", "minute"], ["stop", "inventory", "idle", "Director"], ["comfortablyifera", "inventory", "open", "tomorrow"], ["write", "valve", "closed", "brickbc"], ["Mom", "order", "closed", "day"], ["Mom", "inventory", "running", "now"], ["query", "meeting", "active", "Prepar"], ["write", "machine", "Filipino", "none"], ["Insect", "file", "high", "today"], ["create", "order", "linestyle seen", "year"], ["query", "valve", "failed", "brickbc"], ["write", "interpolation", "ready", "day"], ["write", "customer", "Fred", "minute"], ["write", "heroine", "open", "year"], ["start", "interpolation", "none", "month"], ["execute", "valve", "paid", "careful"], ["check", "Gaussian", "open", "tomorrow"], ["report", "motor", "linestyle seen", "hour"], ["purchases nat", "inventory", "active", "week"], ["notify", "customer", "Fred", "hour"], ["execute", "order", "Filipino", "now"], ["stop", "heroine", "blocked", "none"], ["stop", "email", "closed", "manners"], ["stop", "heroine", "ready", "now"], ["write", "Bodies", "normal", "year"], ["calming", "file", "none", "hour"], ["Mom", "invoice", "closed", "now"], ["update", "invoice", "closed", "Prepar"], ["check", "geospatialard", "idle", "tomorrow"], ["delete", "Gaussian", "ready", "hour"], ["create", "electrolyte benefited", "ready", "tomorrow"], ["create", "order", "low", "manners"], ["write", "valve", "Archived", "minute"], ["notify", "pump", "Filipino", "minute"], ["Benin", "inventory", "closed", "week"], ["query", "machine", "open", "manners"], ["stop", "email", "tackle", "year"], ["write", "email", "Filipino", "none"], ["delete", "inventory", "failed", "Answer"], ["read", "geospatialard", "overload", "hour"], ["update", "email", "Filipino", "tomorrow"], ["check", "pump", "ready", "Prepar"], ["purchases nat", "order", "ready", "month"], ["query", "machine", "downloads", "today"], ["write", "pump", "pending", "Prepar"], ["check", "interpolation", "none", "day"], ["start", "machine", "overload", "aspire"], ["delete", "order", "low", "choice suffrage"], ["comfortablyifera", "meeting", "complete", "hour"], ["start", "job", "closed", "manners"], ["execute", "file", "none", "Answer"], ["Benin", "motor", "overload", "month"], ["check", "file", "downloads", "week"], ["create", "equine", "closed", "day"], ["check", "motor", "idle", "careful"], ["check", "customer", "Filipino", "week"], ["delete", "Fischer", "blocked", "day"], ["Mom", "valve", "pending", "day"], ["execute", "order", "Archived", "minute"], ["create", "inventory", "failed", "careful"], ["delete", "equine", "active", "day"], ["read", "invoice", "Archived", "tomorrow"], ["query", "Fischer", "active", "year"], ["read", "heroine", "low", "day"], ["delete", "machine", "running", "brickbc"], ["purchases nat", "inventory", "active", "month"], ["Benin", "order", "complete", "day"], ["write", "email", "tackle", "week"], ["create", "job", "tackle", "week"], ["delete", "electrolyte benefited", "active", "none"], ["stop", "Gaussian", "ready", "now"], ["tube", "pump", "idle", "day"], ["update", "order", "Filipino", "hour"], ["Mom", "invoice", "open", "month"], ["notify", "motor", "normal", "Director"], ["Mom", "valve", "ready", "hour"], ["read", "invoice", "overload", "aspire"], ["update", "Bodies", "idle", "none"], ["comfortablyifera", "email", "normal", "week"], ["execute", "geospatialard", "high", "minute"], ["read", "equine", "high", "month"], ["report", "pump", "identifierbud", "tomorrow"], ["query", "electrolyte benefited", "blocked", "tomorrow"], ["report", "email", "Archived", "hour"], ["purchases nat", "invoice", "none", "day"], ["Insect", "email", "pending", "hour"], ["start", "file", "downloads", "tomorrow"], ["report", "motor", "Filipino", "day"], ["stop", "inventory", "linestyle seen", "tomorrow"], ["start", "Gaussian", "high", "hour"], ["check", "machine", "sarc", "tomorrow"], ["start", "meeting", "blocked", "brickbc"], ["query", "heroine", "ready", "week"], ["query", "machine", "closed", "manners"], ["notify", "job", "open", "aspire"], ["notify", "meeting", "failed", "Answer"], ["write", "motor", "sarc", "minute"], ["start", "meeting", "idle", "Prepar"], ["write", "valve", "Filipino", "minute"], ["read", "pump", "linestyle seen", "now"], ["read", "order", "pending", "Answer"], ["Benin", "machine", "low", "year"], ["create", "valve", "blocked", "aspire"], ["read", "pump", "open", "careful"], ["start", "meeting", "Fred", "now"], ["calming", "pump", "pending", "day"], ["query", "motor", "high", "careful"], ["query", "equine", "paid", "hour"], ["execute", "valve", "normal", "Answer"], ["Mom", "pump", "none", "minute"], ["report", "equine", "complete", "day"]], "sealed_2_new": [["report", "equine", "active", "Director"], ["Benin", "valve", "active", "careful"], ["delete", "electrolyte benefited", "closed", "manners"], ["create", "Fischer", "running", "Answer"], ["update", "Fischer", "high", "aspire"], ["comfortablyifera", "Gaussian", "paid", "hour"], ["write", "electrolyte benefited", "ready", "brickbc"], ["Mom", "motor", "identifierbud", "none"], ["comfortablyifera", "equine", "active", "day"], ["Mom", "job", "high", "choice suffrage"], ["calming", "valve", "ready", "choice suffrage"], ["Insect", "motor", "sarc", "day"], ["comfortablyifera", "machine", "Filipino", "week"], ["report", "electrolyte benefited", "Fred", "minute"], ["remotely", "motor", "identifierbud", "month"], ["report", "Gaussian", "running", "Prepar"], ["stop", "heroine", "Archived", "hour"], ["read", "electrolyte benefited", "low", "Answer"], ["Insect", "interpolation", "open", "minute"], ["tube", "customer", "Archived", "day"], ["write", "Bodies", "failed", "manners"], ["write", "interpolation", "Filipino", "today"], ["write", "Fischer", "Fred", "hour"], ["query", "Bodies", "linestyle seen", "today"], ["update", "order", "linestyle seen", "careful"], ["write", "Bodies", "tackle", "month"], ["stop", "geospatialard", "downloads", "hour"], ["query", "equine", "Archived", "today"], ["purchases nat", "Gaussian", "blocked", "none"], ["tube", "pump", "sarc", "year"], ["calming", "equine", "blocked", "none"], ["Mom", "valve", "high", "manners"], ["remotely", "meeting", "overload", "Director"], ["Mom", "machine", "Archived", "today"], ["notify", "Bodies", "failed", "manners"], ["Benin", "job", "complete", "Answer"], ["write", "meeting", "Archived", "Director"], ["tube", "motor", "closed", "manners"], ["write", "machine", "Filipino", "Prepar"], ["purchases nat", "Bodies", "blocked", "none"], ["Mom", "heroine", "overload", "now"], ["stop", "job", "identifierbud", "Answer"], ["create", "pump", "Fred", "choice suffrage"], ["read", "interpolation", "downloads", "day"], ["Mom", "email", "Filipino", "tomorrow"], ["calming", "job", "ready", "Director"], ["read", "customer", "tackle", "careful"], ["calming", "meeting", "closed", "Answer"], ["stop", "Bodies", "complete", "careful"], ["comfortablyifera", "valve", "linestyle seen", "day"], ["notify", "Fischer", "none", "Director"], ["Mom", "customer", "failed", "Answer"], ["comfortablyifera", "order", "normal", "choice suffrage"], ["stop", "heroine", "Archived", "year"], ["execute", "equine", "linestyle seen", "hour"], ["execute", "Bodies", "open", "aspire"], ["check", "Bodies", "identifierbud", "month"], ["query", "file", "Filipino", "Prepar"], ["create", "Gaussian", "low", "manners"], ["remotely", "job", "tackle", "minute"], ["tube", "pump", "active", "Answer"], ["start", "Gaussian", "Filipino", "tomorrow"], ["Mom", "Fischer", "normal", "year"], ["Mom", "Gaussian", "low", "year"], ["check", "Bodies", "normal", "manners"], ["delete", "heroine", "normal", "choice suffrage"], ["query", "file", "Fred", "Prepar"], ["notify", "Fischer", "tackle", "now"], ["read", "equine", "open", "aspire"], ["check", "Bodies", "tackle", "week"], ["stop", "valve", "Filipino", "brickbc"], ["Benin", "motor", "Archived", "hour"], ["tube", "customer", "Archived", "day"], ["calming", "pump", "pending", "Answer"], ["read", "valve", "Fred", "aspire"], ["write", "heroine", "Archived", "none"], ["start", "geospatialard", "Archived", "month"], ["Benin", "motor", "idle", "careful"], ["purchases nat", "job", "active", "Answer"], ["Benin", "file", "high", "Answer"], ["Insect", "order", "Archived", "now"], ["comfortablyifera", "invoice", "identifierbud", "minute"], ["calming", "inventory", "Fred", "week"], ["remotely", "inventory", "open", "Director"], ["comfortablyifera", "valve", "sarc", "today"], ["create", "interpolation", "Filipino", "day"], ["update", "Bodies", "linestyle seen", "now"], ["remotely", "meeting", "active", "Answer"], ["Insect", "Fischer", "open", "none"], ["execute", "interpolation", "high", "careful"], ["create", "order", "tackle", "brickbc"], ["query", "electrolyte benefited", "failed", "Director"], ["purchases nat", "electrolyte benefited", "ready", "day"], ["Benin", "customer", "Archived", "minute"], ["comfortablyifera", "Fischer", "open", "year"], ["Mom", "Gaussian", "none", "month"], ["calming", "valve", "paid", "Answer"], ["check", "Fischer", "Fred", "tomorrow"], ["calming", "motor", "tackle", "hour"], ["stop", "equine", "tackle", "week"], ["notify", "interpolation", "linestyle seen", "hour"], ["remotely", "order", "normal", "careful"], ["Benin", "machine", "sarc", "none"], ["remotely", "email", "closed", "Director"], ["stop", "electrolyte benefited", "tackle", "now"], ["Mom", "pump", "normal", "Prepar"], ["stop", "electrolyte benefited", "none", "Answer"], ["calming", "invoice", "closed", "Prepar"], ["update", "Fischer", "closed", "brickbc"], ["check", "heroine", "identifierbud", "tomorrow"], ["delete", "email", "identifierbud", "Prepar"], ["create", "geospatialard", "ready", "careful"], ["tube", "Bodies", "low", "year"], ["purchases nat", "valve", "running", "Answer"], ["purchases nat", "electrolyte benefited", "ready", "minute"], ["read", "interpolation", "closed", "careful"], ["Insect", "machine", "open", "Prepar"], ["stop", "interpolation", "failed", "careful"], ["write", "email", "downloads", "aspire"], ["comfortablyifera", "inventory", "failed", "choice suffrage"], ["Insect", "geospatialard", "overload", "hour"], ["update", "email", "Filipino", "aspire"], ["Mom", "pump", "sarc", "hour"], ["start", "order", "downloads", "choice suffrage"], ["query", "machine", "Filipino", "manners"], ["remotely", "interpolation", "pending", "day"], ["check", "motor", "tackle", "Director"], ["calming", "electrolyte benefited", "overload", "now"], ["delete", "order", "downloads", "Prepar"], ["start", "Bodies", "complete", "choice suffrage"], ["start", "Gaussian", "closed", "Prepar"], ["calming", "interpolation", "none", "hour"], ["Benin", "motor", "overload", "Answer"], ["check", "Gaussian", "Fred", "week"], ["purchases nat", "machine", "Archived", "day"], ["check", "geospatialard", "Archived", "tomorrow"], ["check", "customer", "linestyle seen", "Answer"], ["delete", "machine", "identifierbud", "Director"], ["create", "valve", "Filipino", "careful"], ["Insect", "Gaussian", "failed", "minute"], ["create", "Fischer", "sarc", "hour"], ["delete", "customer", "downloads", "manners"], ["tube", "invoice", "complete", "Prepar"], ["query", "geospatialard", "active", "Prepar"], ["remotely", "meeting", "sarc", "day"], ["comfortablyifera", "Fischer", "running", "none"], ["stop", "interpolation", "identifierbud", "month"], ["calming", "heroine", "complete", "day"], ["write", "Gaussian", "linestyle seen", "week"], ["Benin", "job", "running", "manners"], ["remotely", "Fischer", "active", "none"], ["stop", "email", "tackle", "aspire"], ["tube", "pump", "Filipino", "day"], ["calming", "order", "identifierbud", "hour"], ["calming", "Fischer", "open", "month"], ["purchases nat", "motor", "tackle", "day"], ["remotely", "valve", "tackle", "hour"], ["comfortablyifera", "Fischer", "overload", "week"], ["Mom", "order", "idle", "manners"], ["remotely", "Gaussian", "normal", "week"], ["execute", "order", "linestyle seen", "Answer"], ["purchases nat", "pump", "downloads", "month"], ["calming", "pump", "identifierbud", "tomorrow"], ["remotely", "electrolyte benefited", "blocked", "tomorrow"], ["remotely", "email", "high", "aspire"], ["Mom", "invoice", "none", "Answer"], ["calming", "interpolation", "pending", "hour"], ["start", "interpolation", "identifierbud", "tomorrow"], ["purchases nat", "motor", "ready", "Director"], ["comfortablyifera", "inventory", "failed", "brickbc"], ["start", "motor", "Filipino", "manners"], ["check", "Fischer", "blocked", "Prepar"], ["calming", "meeting", "tackle", "none"], ["query", "Bodies", "tackle", "week"], ["Insect", "machine", "closed", "aspire"], ["Benin", "job", "open", "Answer"], ["notify", "interpolation", "Fred", "today"], ["purchases nat", "motor", "Fred", "minute"], ["start", "interpolation", "sarc", "hour"], ["write", "valve", "Filipino", "Prepar"], ["read", "pump", "Fred", "Prepar"], ["Insect", "order", "pending", "brickbc"], ["notify", "Bodies", "sarc", "year"], ["comfortablyifera", "valve", "blocked", "manners"], ["Benin", "Bodies", "open", "hour"], ["Insect", "meeting", "failed", "Answer"], ["Benin", "pump", "tackle", "day"], ["query", "heroine", "high", "Prepar"], ["tube", "file", "paid", "careful"], ["execute", "Gaussian", "linestyle seen", "today"], ["comfortablyifera", "pump", "none", "choice suffrage"], ["report", "heroine", "downloads", "day"]], "sealed_4_new": [["Mom", "electrolyte benefited", "sarc", "choice suffrage"], ["Mom", "heroine", "Archived", "brickbc"], ["Mom", "electrolyte benefited", "sarc", "choice suffrage"], ["tube", "Gaussian", "downloads", "brickbc"], ["Insect", "electrolyte benefited", "sarc", "choice suffrage"], ["comfortablyifera", "equine", "Filipino", "Answer"], ["Insect", "electrolyte benefited", "identifierbud", "manners"], ["tube", "Bodies", "Filipino", "manners"], ["remotely", "Fischer", "Archived", "brickbc"], ["purchases nat", "electrolyte benefited", "sarc", "Director"], ["purchases nat", "Bodies", "downloads", "manners"], ["tube", "electrolyte benefited", "downloads", "Prepar"], ["comfortablyifera", "interpolation", "downloads", "Answer"], ["Mom", "Fischer", "linestyle seen", "Director"], ["Insect", "Bodies", "tackle", "aspire"], ["Mom", "geospatialard", "Filipino", "Answer"], ["Benin", "Gaussian", "Filipino", "Prepar"], ["purchases nat", "heroine", "Filipino", "manners"], ["tube", "interpolation", "identifierbud", "brickbc"], ["tube", "Fischer", "Archived", "brickbc"], ["comfortablyifera", "Gaussian", "tackle", "manners"], ["Mom", "electrolyte benefited", "sarc", "brickbc"], ["Mom", "electrolyte benefited", "sarc", "Prepar"], ["Insect", "Bodies", "downloads", "Answer"], ["calming", "equine", "Filipino", "careful"], ["remotely", "Gaussian", "sarc", "Director"], ["calming", "heroine", "sarc", "Prepar"], ["Mom", "interpolation", "identifierbud", "Prepar"], ["calming", "Fischer", "Filipino", "careful"], ["calming", "electrolyte benefited", "downloads", "Answer"], ["comfortablyifera", "electrolyte benefited", "Fred", "brickbc"], ["Mom", "heroine", "tackle", "Prepar"], ["comfortablyifera", "Fischer", "linestyle seen", "careful"], ["comfortablyifera", "equine", "Filipino", "choice suffrage"], ["tube", "Bodies", "identifierbud", "Answer"], ["Insect", "equine", "Filipino", "aspire"], ["Benin", "Gaussian", "identifierbud", "brickbc"], ["comfortablyifera", "Bodies", "sarc", "Prepar"], ["Mom", "geospatialard", "linestyle seen", "choice suffrage"], ["Mom", "heroine", "sarc", "Director"], ["Benin", "Fischer", "downloads", "manners"], ["remotely", "Fischer", "sarc", "choice suffrage"], ["Mom", "geospatialard", "Fred", "Director"], ["purchases nat", "Fischer", "Fred", "Prepar"], ["Insect", "geospatialard", "sarc", "aspire"], ["comfortablyifera", "heroine", "downloads", "choice suffrage"], ["purchases nat", "Fischer", "Filipino", "manners"], ["comfortablyifera", "Bodies", "downloads", "Director"], ["comfortablyifera", "Fischer", "Fred", "aspire"], ["Mom", "heroine", "identifierbud", "Answer"], ["tube", "geospatialard", "Filipino", "careful"], ["purchases nat", "Gaussian", "identifierbud", "choice suffrage"], ["Mom", "electrolyte benefited", "linestyle seen", "Director"], ["Benin", "Fischer", "tackle", "choice suffrage"], ["purchases nat", "equine", "Archived", "Director"], ["tube", "Gaussian", "tackle", "Prepar"], ["comfortablyifera", "heroine", "Fred", "Prepar"], ["Benin", "Fischer", "tackle", "aspire"], ["purchases nat", "geospatialard", "linestyle seen", "Director"], ["Mom", "Fischer", "Filipino", "aspire"], ["tube", "equine", "tackle", "brickbc"], ["comfortablyifera", "equine", "sarc", "aspire"], ["Mom", "interpolation", "Filipino", "Director"], ["Insect", "electrolyte benefited", "linestyle seen", "manners"], ["tube", "Fischer", "Filipino", "Answer"], ["comfortablyifera", "Fischer", "Filipino", "manners"], ["remotely", "geospatialard", "downloads", "brickbc"], ["Benin", "heroine", "sarc", "Answer"], ["Insect", "heroine", "sarc", "choice suffrage"], ["Benin", "interpolation", "Archived", "Prepar"], ["tube", "interpolation", "Fred", "Director"], ["Mom", "Bodies", "Fred", "Director"], ["tube", "Fischer", "Filipino", "brickbc"], ["Mom", "equine", "downloads", "aspire"], ["Mom", "Gaussian", "sarc", "Director"], ["tube", "Bodies", "downloads", "Answer"], ["Mom", "geospatialard", "Filipino", "manners"], ["Benin", "electrolyte benefited", "Filipino", "manners"], ["Mom", "electrolyte benefited", "Fred", "Director"], ["comfortablyifera", "equine", "linestyle seen", "Prepar"], ["Benin", "Fischer", "downloads", "careful"], ["Mom", "equine", "sarc", "aspire"], ["tube", "Bodies", "tackle", "aspire"], ["Insect", "electrolyte benefited", "Fred", "careful"], ["calming", "interpolation", "Filipino", "manners"], ["Insect", "Bodies", "tackle", "brickbc"], ["Benin", "electrolyte benefited", "identifierbud", "brickbc"], ["Mom", "geospatialard", "Filipino", "Director"], ["calming", "Fischer", "sarc", "careful"], ["comfortablyifera", "Gaussian", "sarc", "Director"], ["tube", "Fischer", "downloads", "Answer"], ["comfortablyifera", "equine", "Fred", "careful"], ["comfortablyifera", "geospatialard", "Fred", "Answer"], ["comfortablyifera", "Gaussian", "Archived", "manners"], ["Mom", "geospatialard", "Filipino", "Director"], ["tube", "geospatialard", "linestyle seen", "manners"], ["calming", "Bodies", "Fred", "choice suffrage"], ["Insect", "Gaussian", "tackle", "Answer"], ["tube", "electrolyte benefited", "downloads", "manners"], ["Insect", "interpolation", "downloads", "choice suffrage"], ["Benin", "heroine", "Archived", "careful"], ["tube", "heroine", "sarc", "Answer"], ["tube", "geospatialard", "downloads", "brickbc"], ["Benin", "heroine", "sarc", "careful"], ["tube", "Fischer", "downloads", "careful"], ["calming", "heroine", "Archived", "Answer"], ["Benin", "heroine", "Fred", "choice suffrage"], ["Insect", "equine", "Filipino", "aspire"], ["remotely", "equine", "Filipino", "Director"], ["remotely", "geospatialard", "tackle", "Prepar"], ["calming", "electrolyte benefited", "identifierbud", "Answer"], ["comfortablyifera", "Fischer", "downloads", "Director"], ["Mom", "Bodies", "downloads", "brickbc"], ["Benin", "Fischer", "Archived", "manners"], ["comfortablyifera", "electrolyte benefited", "identifierbud", "Prepar"], ["calming", "Gaussian", "Fred", "aspire"], ["remotely", "electrolyte benefited", "identifierbud", "choice suffrage"], ["calming", "heroine", "sarc", "careful"], ["Benin", "interpolation", "downloads", "manners"], ["tube", "Gaussian", "Archived", "Director"], ["tube", "equine", "tackle", "brickbc"], ["calming", "electrolyte benefited", "Filipino", "Answer"], ["Insect", "Gaussian", "downloads", "Answer"], ["tube", "electrolyte benefited", "identifierbud", "choice suffrage"], ["tube", "geospatialard", "downloads", "Prepar"], ["remotely", "equine", "Archived", "Prepar"], ["calming", "electrolyte benefited", "sarc", "careful"], ["Mom", "heroine", "linestyle seen", "careful"], ["comfortablyifera", "electrolyte benefited", "Archived", "brickbc"], ["comfortablyifera", "geospatialard", "Filipino", "aspire"], ["Insect", "Bodies", "downloads", "choice suffrage"], ["Mom", "Fischer", "tackle", "choice suffrage"], ["tube", "equine", "tackle", "careful"], ["calming", "Fischer", "Fred", "Answer"], ["tube", "interpolation", "downloads", "choice suffrage"], ["purchases nat", "interpolation", "tackle", "brickbc"], ["purchases nat", "Fischer", "linestyle seen", "choice suffrage"], ["comfortablyifera", "interpolation", "identifierbud", "manners"], ["comfortablyifera", "Bodies", "Filipino", "Prepar"], ["purchases nat", "heroine", "sarc", "manners"], ["calming", "heroine", "linestyle seen", "aspire"], ["tube", "equine", "Fred", "careful"], ["purchases nat", "geospatialard", "Fred", "Prepar"], ["purchases nat", "equine", "downloads", "choice suffrage"], ["Mom", "Fischer", "Archived", "manners"], ["comfortablyifera", "equine", "Fred", "careful"], ["Benin", "Gaussian", "Archived", "careful"], ["Insect", "interpolation", "sarc", "choice suffrage"], ["comfortablyifera", "geospatialard", "linestyle seen", "careful"], ["remotely", "Gaussian", "Archived", "Answer"], ["purchases nat", "Bodies", "Archived", "Answer"], ["tube", "Gaussian", "linestyle seen", "Answer"], ["comfortablyifera", "heroine", "linestyle seen", "Prepar"], ["remotely", "electrolyte benefited", "linestyle seen", "Director"], ["purchases nat", "Gaussian", "downloads", "Director"], ["purchases nat", "heroine", "tackle", "brickbc"], ["calming", "interpolation", "sarc", "Prepar"], ["purchases nat", "geospatialard", "downloads", "manners"], ["Benin", "geospatialard", "downloads", "brickbc"], ["comfortablyifera", "Fischer", "Fred", "choice suffrage"], ["Insect", "Fischer", "downloads", "careful"], ["tube", "Fischer", "Archived", "careful"], ["Benin", "heroine", "identifierbud", "choice suffrage"], ["calming", "Gaussian", "identifierbud", "manners"], ["calming", "heroine", "tackle", "manners"], ["Benin", "Bodies", "tackle", "Prepar"], ["Benin", "Fischer", "identifierbud", "choice suffrage"], ["Mom", "electrolyte benefited", "Fred", "Answer"], ["comfortablyifera", "Gaussian", "downloads", "brickbc"], ["remotely", "equine", "Archived", "manners"], ["calming", "Bodies", "sarc", "Director"], ["purchases nat", "interpolation", "linestyle seen", "choice suffrage"], ["comfortablyifera", "interpolation", "downloads", "careful"], ["calming", "interpolation", "Fred", "brickbc"], ["remotely", "heroine", "sarc", "brickbc"], ["Mom", "geospatialard", "Filipino", "Prepar"], ["Benin", "geospatialard", "identifierbud", "brickbc"], ["remotely", "Fischer", "downloads", "manners"], ["Mom", "Fischer", "Filipino", "aspire"], ["calming", "Gaussian", "tackle", "careful"], ["comfortablyifera", "Fischer", "sarc", "choice suffrage"], ["Mom", "Bodies", "Fred", "choice suffrage"], ["remotely", "geospatialard", "tackle", "manners"], ["Insect", "Fischer", "tackle", "Director"], ["Insect", "Fischer", "identifierbud", "Prepar"], ["Benin", "Gaussian", "Archived", "choice suffrage"], ["tube", "Bodies", "sarc", "brickbc"], ["Insect", "electrolyte benefited", "Fred", "careful"], ["comfortablyifera", "geospatialard", "Fred", "brickbc"], ["comfortablyifera", "geospatialard", "sarc", "brickbc"], ["Insect", "interpolation", "sarc", "aspire"], ["comfortablyifera", "geospatialard", "sarc", "Answer"]], "sealed_id": [["delete", "email", "active", "tomorrow"], ["check", "invoice", "failed", "now"], ["stop", "motor", "active", "week"], ["stop", "motor", "normal", "hour"], ["notify", "invoice", "normal", "now"], ["read", "meeting", "closed", "none"], ["stop", "meeting", "complete", "now"], ["check", "meeting", "ready", "minute"], ["query", "meeting", "idle", "minute"], ["notify", "meeting", "high", "day"], ["report", "job", "running", "none"], ["read", "meeting", "none", "day"], ["update", "job", "high", "none"], ["stop", "meeting", "complete", "tomorrow"], ["notify", "machine", "blocked", "minute"], ["start", "valve", "high", "month"], ["write", "motor", "blocked", "minute"], ["execute", "machine", "none", "year"], ["check", "email", "pending", "minute"], ["delete", "invoice", "ready", "minute"], ["report", "machine", "open", "today"], ["execute", "pump", "closed", "minute"], ["execute", "pump", "blocked", "year"], ["query", "meeting", "open", "hour"], ["check", "inventory", "open", "week"], ["report", "email", "idle", "week"], ["start", "motor", "normal", "minute"], ["delete", "meeting", "low", "tomorrow"], ["query", "inventory", "complete", "minute"], ["write", "order", "open", "hour"], ["write", "machine", "overload", "none"], ["query", "meeting", "blocked", "today"], ["notify", "email", "overload", "today"], ["execute", "pump", "none", "none"], ["check", "invoice", "running", "tomorrow"], ["check", "inventory", "paid", "tomorrow"], ["delete", "order", "none", "minute"], ["check", "pump", "high", "year"], ["update", "valve", "blocked", "day"], ["delete", "motor", "blocked", "now"], ["stop", "order", "normal", "today"], ["report", "invoice", "pending", "hour"], ["read", "machine", "none", "week"], ["read", "inventory", "idle", "year"], ["write", "invoice", "normal", "hour"], ["start", "email", "failed", "minute"], ["write", "pump", "high", "minute"], ["check", "inventory", "complete", "today"], ["notify", "valve", "open", "week"], ["read", "customer", "low", "hour"], ["execute", "file", "open", "year"], ["start", "motor", "ready", "tomorrow"], ["update", "motor", "low", "hour"], ["notify", "motor", "closed", "hour"], ["report", "meeting", "pending", "today"], ["execute", "pump", "paid", "day"], ["start", "order", "ready", "tomorrow"], ["check", "meeting", "failed", "day"], ["delete", "motor", "closed", "tomorrow"], ["notify", "invoice", "ready", "tomorrow"], ["notify", "machine", "active", "tomorrow"], ["create", "file", "normal", "today"], ["execute", "customer", "paid", "minute"], ["create", "valve", "high", "none"], ["report", "meeting", "paid", "tomorrow"], ["start", "customer", "failed", "now"], ["query", "meeting", "active", "none"], ["read", "email", "pending", "today"], ["notify", "machine", "complete", "tomorrow"], ["check", "valve", "running", "none"], ["write", "invoice", "low", "today"], ["update", "machine", "ready", "today"], ["execute", "meeting", "none", "minute"], ["execute", "customer", "closed", "month"], ["query", "email", "blocked", "day"], ["read", "motor", "low", "hour"], ["update", "machine", "idle", "none"], ["write", "order", "failed", "now"], ["report", "file", "idle", "none"], ["query", "customer", "low", "month"], ["check", "email", "low", "tomorrow"], ["query", "invoice", "closed", "today"], ["create", "machine", "ready", "now"], ["start", "motor", "none", "today"], ["check", "customer", "overload", "day"], ["execute", "inventory", "blocked", "hour"], ["start", "machine", "closed", "none"], ["write", "customer", "running", "hour"], ["stop", "valve", "ready", "none"], ["notify", "order", "idle", "day"], ["update", "customer", "open", "now"], ["delete", "meeting", "complete", "tomorrow"], ["delete", "order", "normal", "today"], ["read", "customer", "high", "minute"], ["write", "meeting", "idle", "none"], ["start", "email", "closed", "hour"], ["report", "valve", "blocked", "year"], ["write", "pump", "complete", "minute"], ["stop", "meeting", "complete", "hour"], ["stop", "invoice", "closed", "now"], ["execute", "meeting", "closed", "now"], ["execute", "motor", "active", "hour"], ["start", "email", "failed", "tomorrow"], ["create", "valve", "overload", "day"], ["read", "file", "failed", "year"], ["delete", "file", "blocked", "minute"], ["report", "order", "blocked", "year"], ["check", "job", "complete", "now"], ["update", "job", "paid", "today"], ["read", "valve", "pending", "tomorrow"], ["notify", "file", "normal", "month"], ["update", "motor", "active", "tomorrow"], ["write", "job", "normal", "week"], ["notify", "meeting", "closed", "minute"], ["execute", "machine", "normal", "tomorrow"], ["check", "order", "normal", "none"], ["execute", "job", "active", "week"], ["create", "motor", "failed", "week"], ["write", "pump", "ready", "none"], ["delete", "order", "active", "week"], ["query", "job", "closed", "year"], ["report", "order", "ready", "tomorrow"], ["delete", "invoice", "ready", "tomorrow"], ["write", "motor", "running", "hour"], ["delete", "order", "open", "hour"], ["stop", "invoice", "active", "hour"], ["start", "valve", "complete", "week"], ["check", "valve", "running", "today"], ["notify", "motor", "blocked", "now"], ["read", "customer", "blocked", "day"], ["start", "email", "blocked", "now"], ["write", "customer", "failed", "year"], ["check", "valve", "complete", "tomorrow"], ["report", "machine", "complete", "tomorrow"], ["stop", "invoice", "high", "now"], ["notify", "customer", "idle", "none"], ["stop", "job", "normal", "now"], ["stop", "order", "complete", "now"], ["notify", "machine", "ready", "now"], ["create", "inventory", "blocked", "year"], ["create", "pump", "blocked", "now"], ["start", "order", "blocked", "now"], ["read", "motor", "ready", "year"], ["start", "pump", "running", "tomorrow"], ["delete", "valve", "low", "none"], ["read", "job", "idle", "minute"], ["start", "customer", "overload", "day"], ["execute", "job", "paid", "year"], ["read", "meeting", "overload", "minute"], ["read", "motor", "overload", "hour"], ["update", "meeting", "overload", "year"], ["read", "machine", "complete", "minute"], ["execute", "machine", "high", "day"], ["write", "motor", "blocked", "now"], ["query", "inventory", "failed", "tomorrow"], ["query", "order", "open", "month"], ["delete", "job", "complete", "minute"], ["start", "customer", "idle", "month"], ["notify", "email", "normal", "today"], ["update", "motor", "high", "minute"], ["delete", "invoice", "none", "year"], ["notify", "file", "blocked", "year"], ["delete", "email", "low", "hour"], ["stop", "valve", "running", "minute"], ["query", "motor", "normal", "week"], ["notify", "order", "blocked", "hour"], ["create", "invoice", "active", "week"], ["stop", "invoice", "blocked", "month"], ["execute", "meeting", "ready", "week"], ["delete", "motor", "pending", "now"], ["create", "valve", "complete", "today"], ["report", "email", "pending", "year"], ["start", "pump", "active", "day"], ["report", "pump", "overload", "tomorrow"], ["write", "meeting", "paid", "day"], ["create", "motor", "complete", "minute"], ["create", "inventory", "running", "now"], ["read", "valve", "idle", "year"], ["query", "inventory", "normal", "hour"], ["execute", "inventory", "pending", "minute"], ["notify", "order", "open", "week"], ["start", "email", "closed", "none"], ["read", "machine", "failed", "now"], ["check", "pump", "high", "none"], ["notify", "machine", "paid", "minute"], ["stop", "job", "ready", "week"], ["report", "order", "high", "week"], ["report", "pump", "normal", "month"], ["notify", "order", "complete", "tomorrow"], ["update", "pump", "open", "month"], ["delete", "file", "idle", "day"], ["create", "customer", "overload", "week"], ["report", "machine", "paid", "minute"], ["read", "email", "none", "day"], ["check", "order", "none", "now"], ["create", "inventory", "ready", "tomorrow"], ["write", "customer", "paid", "today"], ["report", "file", "normal", "today"], ["create", "pump", "paid", "tomorrow"], ["update", "meeting", "blocked", "month"]], "sealed_ood": [["report", "customer", "active", "now"], ["query", "valve", "active", "minute"], ["delete", "machine", "closed", "today"], ["create", "valve", "running", "year"], ["update", "machine", "high", "today"], ["stop", "email", "paid", "hour"], ["write", "motor", "ready", "hour"], ["report", "motor", "open", "none"], ["write", "email", "active", "day"], ["create", "job", "high", "hour"], ["execute", "valve", "ready", "now"], ["read", "motor", "pending", "day"], ["query", "machine", "idle", "week"], ["report", "machine", "active", "minute"], ["execute", "motor", "paid", "month"], ["report", "motor", "running", "none"], ["stop", "pump", "ready", "hour"], ["read", "invoice", "low", "today"], ["stop", "email", "open", "minute"], ["notify", "customer", "pending", "day"], ["write", "valve", "failed", "hour"], ["write", "valve", "complete", "today"], ["write", "job", "high", "hour"], ["query", "file", "low", "today"], ["update", "order", "open", "now"], ["write", "invoice", "paid", "month"], ["stop", "email", "active", "hour"], ["query", "inventory", "high", "today"], ["stop", "motor", "blocked", "none"], ["read", "pump", "active", "year"], ["update", "motor", "blocked", "none"], ["stop", "valve", "high", "hour"], ["check", "meeting", "overload", "month"], ["query", "machine", "paid", "today"], ["notify", "meeting", "failed", "hour"], ["create", "job", "complete", "year"], ["write", "meeting", "failed", "today"], ["update", "motor", "closed", "day"], ["write", "machine", "ready", "hour"], ["update", "file", "blocked", "none"], ["write", "file", "overload", "now"], ["stop", "job", "overload", "month"], ["create", "pump", "high", "hour"], ["read", "job", "pending", "day"], ["stop", "email", "overload", "tomorrow"], ["create", "job", "ready", "now"], ["read", "customer", "blocked", "tomorrow"], ["read", "meeting", "closed", "day"], ["stop", "email", "complete", "today"], ["write", "valve", "active", "day"], ["notify", "order", "none", "hour"], ["query", "customer", "failed", "today"], ["execute", "order", "normal", "minute"], ["stop", "invoice", "running", "year"], ["execute", "pump", "ready", "hour"], ["execute", "inventory", "open", "tomorrow"], ["check", "file", "failed", "month"], ["query", "file", "ready", "minute"], ["create", "motor", "low", "year"], ["notify", "job", "active", "minute"], ["notify", "pump", "active", "today"], ["start", "invoice", "high", "tomorrow"], ["execute", "meeting", "normal", "year"], ["write", "customer", "low", "year"], ["check", "customer", "normal", "year"], ["delete", "motor", "normal", "year"], ["query", "file", "overload", "minute"], ["notify", "pump", "idle", "now"], ["read", "valve", "open", "tomorrow"], ["check", "motor", "paid", "week"], ["stop", "valve", "ready", "month"], ["check", "motor", "running", "hour"], ["start", "customer", "complete", "day"], ["read", "pump", "pending", "hour"], ["read", "valve", "active", "year"], ["write", "valve", "closed", "none"], ["start", "customer", "ready", "month"], ["check", "motor", "idle", "month"], ["start", "job", "active", "year"], ["update", "file", "high", "today"], ["execute", "order", "active", "now"], ["execute", "invoice", "active", "minute"], ["stop", "inventory", "idle", "week"], ["notify", "inventory", "open", "tomorrow"], ["write", "valve", "closed", "today"], ["create", "order", "closed", "day"], ["update", "inventory", "running", "now"], ["query", "meeting", "active", "minute"], ["write", "machine", "open", "none"], ["execute", "file", "high", "today"], ["create", "order", "active", "year"], ["query", "valve", "failed", "today"], ["write", "valve", "ready", "day"], ["write", "customer", "active", "minute"], ["write", "valve", "open", "year"], ["start", "valve", "none", "month"], ["execute", "valve", "paid", "week"], ["check", "valve", "open", "tomorrow"], ["report", "motor", "normal", "hour"], ["stop", "inventory", "active", "week"], ["notify", "customer", "paid", "hour"], ["execute", "order", "normal", "now"], ["stop", "machine", "blocked", "none"], ["stop", "email", "closed", "week"], ["stop", "inventory", "ready", "now"], ["write", "pump", "normal", "year"], ["stop", "file", "none", "hour"], ["start", "invoice", "closed", "now"], ["update", "invoice", "closed", "day"], ["check", "order", "idle", "tomorrow"], ["delete", "email", "ready", "hour"], ["create", "meeting", "ready", "tomorrow"], ["create", "order", "low", "year"], ["write", "valve", "running", "minute"], ["notify", "pump", "ready", "minute"], ["read", "inventory", "closed", "week"], ["query", "machine", "open", "year"], ["stop", "email", "failed", "year"], ["write", "email", "none", "none"], ["delete", "inventory", "failed", "today"], ["read", "pump", "overload", "hour"], ["update", "email", "blocked", "tomorrow"], ["check", "pump", "ready", "hour"], ["start", "order", "ready", "month"], ["query", "machine", "open", "today"], ["write", "pump", "pending", "day"], ["check", "motor", "none", "day"], ["start", "machine", "overload", "now"], ["delete", "order", "low", "year"], ["start", "meeting", "complete", "hour"], ["start", "job", "closed", "day"], ["execute", "file", "none", "hour"], ["update", "motor", "overload", "month"], ["check", "file", "high", "week"], ["create", "machine", "closed", "day"], ["check", "motor", "idle", "tomorrow"], ["check", "customer", "pending", "week"], ["delete", "machine", "blocked", "day"], ["create", "valve", "pending", "day"], ["execute", "order", "failed", "minute"], ["create", "inventory", "failed", "hour"], ["delete", "customer", "active", "day"], ["read", "invoice", "complete", "tomorrow"], ["query", "machine", "active", "year"], ["read", "meeting", "low", "day"], ["delete", "machine", "running", "none"], ["stop", "inventory", "active", "month"], ["report", "order", "complete", "day"], ["write", "email", "failed", "week"], ["create", "job", "running", "week"], ["delete", "machine", "active", "none"], ["stop", "email", "ready", "now"], ["read", "pump", "idle", "day"], ["update", "order", "paid", "hour"], ["read", "invoice", "open", "month"], ["notify", "motor", "normal", "day"], ["check", "valve", "ready", "hour"], ["read", "invoice", "overload", "week"], ["update", "order", "idle", "none"], ["write", "email", "normal", "week"], ["execute", "order", "high", "minute"], ["read", "pump", "high", "month"], ["report", "pump", "idle", "tomorrow"], ["query", "motor", "blocked", "tomorrow"], ["report", "email", "high", "hour"], ["read", "invoice", "none", "day"], ["write", "email", "pending", "hour"], ["start", "file", "idle", "tomorrow"], ["report", "motor", "ready", "day"], ["stop", "inventory", "failed", "tomorrow"], ["start", "motor", "high", "hour"], ["check", "machine", "blocked", "tomorrow"], ["start", "meeting", "blocked", "none"], ["query", "file", "ready", "week"], ["query", "machine", "closed", "week"], ["notify", "job", "open", "hour"], ["notify", "meeting", "failed", "today"], ["write", "motor", "active", "minute"], ["start", "meeting", "idle", "hour"], ["write", "valve", "closed", "minute"], ["read", "pump", "ready", "now"], ["read", "order", "pending", "day"], ["notify", "machine", "low", "year"], ["create", "valve", "blocked", "none"], ["read", "pump", "open", "hour"], ["start", "meeting", "failed", "now"], ["update", "pump", "pending", "day"], ["query", "motor", "high", "hour"], ["query", "file", "paid", "hour"], ["execute", "valve", "normal", "today"], ["read", "pump", "none", "minute"], ["report", "pump", "complete", "day"], ["check", "pump", "overload", "now"], ["start", "meeting", "none", "month"], ["create", "order", "high", "today"], ["execute", "valve", "open", "year"], ["create", "pump", "open", "none"], ["write", "email", "paid", "week"], ["report", "file", "complete", "year"], ["check", "motor", "idle", "day"], ["notify", "customer", "running", "year"], ["start", "pump", "active", "minute"], ["update", "order", "complete", "year"], ["delete", "machine", "blocked", "year"], ["stop", "pump", "running", "year"], ["check", "valve", "paid", "hour"], ["write", "customer", "open", "none"], ["check", "motor", "overload", "month"], ["check", "motor", "high", "tomorrow"], ["read", "invoice", "low", "month"], ["check", "motor", "none", "tomorrow"], ["update", "inventory", "normal", "month"], ["query", "order", "low", "year"], ["write", "email", "overload", "year"], ["report", "meeting", "blocked", "tomorrow"], ["start", "invoice", "ready", "today"], ["write", "email", "complete", "week"], ["write", "job", "active", "year"], ["read", "pump", "failed", "none"], ["write", "email", "active", "week"], ["start", "motor", "overload", "month"], ["read", "invoice", "normal", "none"], ["report", "customer", "ready", "none"], ["notify", "pump", "paid", "minute"], ["delete", "order", "ready", "hour"], ["read", "pump", "low", "minute"], ["create", "job", "normal", "none"], ["read", "machine", "active", "minute"], ["report", "motor", "open", "today"], ["stop", "inventory", "running", "week"], ["delete", "machine", "idle", "year"], ["execute", "valve", "running", "year"], ["write", "meeting", "overload", "month"], ["notify", "valve", "running", "day"], ["write", "motor", "low", "day"], ["read", "pump", "ready", "today"], ["check", "file", "paid", "tomorrow"], ["write", "machine", "closed", "week"], ["update", "motor", "none", "hour"], ["update", "order", "none", "now"], ["create", "customer", "pending", "week"], ["check", "customer", "open", "none"], ["write", "machine", "failed", "today"], ["create", "job", "closed", "week"], ["query", "file", "pending", "today"], ["check", "valve", "blocked", "tomorrow"], ["stop", "email", "failed", "minute"], ["notify", "job", "open", "tomorrow"], ["execute", "order", "idle", "hour"], ["start", "job", "low", "day"], ["check", "email", "open", "none"], ["start", "job", "pending", "week"], ["update", "inventory", "open", "hour"], ["stop", "order", "paid", "hour"], ["start", "meeting", "idle", "none"], ["write", "valve", "paid", "year"], ["report", "customer", "active", "week"], ["start", "pump", "running", "day"], ["start", "meeting", "normal", "hour"], ["report", "customer", "complete", "week"], ["update", "inventory", "complete", "tomorrow"], ["write", "email", "overload", "day"], ["execute", "valve", "idle", "month"], ["delete", "invoice", "ready", "hour"], ["check", "meeting", "normal", "year"], ["delete", "inventory", "high", "hour"], ["stop", "order", "failed", "hour"], ["write", "file", "paid", "hour"], ["delete", "inventory", "paid", "hour"], ["notify", "pump", "paid", "today"], ["query", "file", "failed", "tomorrow"], ["delete", "valve", "high", "today"], ["read", "motor", "normal", "day"], ["start", "invoice", "low", "now"], ["query", "file", "normal", "minute"], ["stop", "pump", "high", "hour"], ["notify", "order", "none", "month"], ["stop", "inventory", "none", "month"], ["report", "pump", "high", "today"], ["notify", "pump", "paid", "hour"], ["stop", "inventory", "paid", "hour"], ["report", "customer", "none", "year"], ["execute", "inventory", "failed", "today"], ["delete", "machine", "normal", "none"], ["stop", "email", "active", "today"], ["stop", "inventory", "low", "today"], ["stop", "meeting", "complete", "year"], ["create", "customer", "high", "today"], ["stop", "inventory", "complete", "today"], ["read", "pump", "none", "year"], ["report", "motor", "complete", "today"], ["update", "inventory", "paid", "minute"], ["report", "customer", "overload", "hour"], ["delete", "customer", "paid", "minute"], ["notify", "pump", "pending", "now"], ["delete", "file", "ready", "month"], ["stop", "email", "overload", "year"], ["write", "invoice", "complete", "day"], ["delete", "customer", "closed", "month"], ["stop", "invoice", "paid", "hour"], ["check", "file", "failed", "day"], ["query", "file", "normal", "month"], ["query", "file", "blocked", "month"], ["start", "valve", "overload", "month"], ["start", "meeting", "failed", "tomorrow"], ["read", "invoice", "open", "week"], ["execute", "pump", "normal", "year"], ["notify", "pump", "pending", "month"], ["write", "inventory", "ready", "month"], ["report", "motor", "none", "hour"], ["write", "pump", "normal", "day"], ["notify", "pump", "low", "hour"], ["execute", "file", "closed", "week"], ["update", "valve", "normal", "year"], ["stop", "customer", "complete", "day"], ["stop", "inventory", "ready", "hour"], ["start", "email", "overload", "month"], ["read", "invoice", "open", "now"], ["write", "pump", "active", "year"], ["create", "meeting", "running", "year"], ["write", "motor", "pending", "day"], ["notify", "order", "ready", "month"], ["check", "file", "blocked", "day"], ["query", "machine", "paid", "hour"], ["start", "meeting", "low", "none"], ["create", "inventory", "failed", "today"], ["write", "order", "pending", "day"], ["delete", "customer", "running", "now"], ["query", "job", "failed", "hour"], ["report", "motor", "ready", "none"], ["check", "meeting", "running", "day"], ["execute", "invoice", "closed", "day"], ["delete", "valve", "blocked", "tomorrow"], ["execute", "valve", "paid", "now"], ["notify", "pump", "pending", "hour"], ["stop", "inventory", "pending", "now"], ["execute", "email", "complete", "year"], ["query", "file", "paid", "tomorrow"], ["delete", "machine", "high", "year"], ["execute", "motor", "complete", "day"], ["create", "machine", "high", "today"], ["notify", "valve", "failed", "hour"], ["check", "file", "active", "week"], ["delete", "file", "low", "year"], ["report", "customer", "running", "today"], ["write", "valve", "active", "hour"], ["update", "email", "running", "year"], ["write", "pump", "open", "none"], ["report", "motor", "failed", "year"], ["write", "invoice", "low", "day"], ["update", "order", "high", "none"], ["write", "inventory", "none", "hour"], ["write", "email", "pending", "week"], ["create", "job", "low", "minute"], ["read", "invoice", "blocked", "month"], ["create", "customer", "failed", "today"], ["notify", "email", "low", "year"], ["stop", "order", "none", "hour"], ["read", "invoice", "high", "year"], ["start", "inventory", "closed", "week"], ["write", "order", "none", "month"], ["check", "file", "complete", "month"], ["create", "pump", "none", "hour"], ["execute", "motor", "complete", "year"], ["check", "inventory", "complete", "year"], ["report", "motor", "open", "hour"], ["read", "pump", "closed", "year"], ["execute", "inventory", "active", "minute"], ["start", "invoice", "failed", "tomorrow"], ["start", "email", "blocked", "tomorrow"], ["start", "valve", "ready", "month"], ["write", "job", "active", "minute"], ["execute", "valve", "blocked", "now"], ["execute", "motor", "running", "year"], ["start", "pump", "none", "month"], ["create", "meeting", "paid", "hour"], ["update", "motor", "idle", "tomorrow"], ["read", "job", "pending", "week"], ["execute", "order", "pending", "hour"], ["read", "pump", "none", "now"], ["create", "motor", "failed", "today"], ["update", "inventory", "running", "day"], ["start", "pump", "overload", "now"], ["execute", "meeting", "active", "minute"], ["notify", "order", "high", "today"], ["notify", "pump", "high", "week"], ["check", "file", "idle", "year"], ["query", "customer", "running", "day"], ["start", "valve", "high", "today"], ["query", "job", "ready", "month"], ["start", "invoice", "idle", "none"], ["query", "valve", "complete", "year"], ["start", "motor", "ready", "hour"], ["read", "meeting", "blocked", "tomorrow"], ["execute", "order", "open", "year"], ["read", "order", "blocked", "tomorrow"], ["query", "order", "overload", "month"], ["delete", "invoice", "blocked", "none"], ["update", "inventory", "idle", "tomorrow"], ["query", "file", "low", "tomorrow"]], "seed": 20261017, "types": {"Answer": "name", "Archived": "name", "Benin": "name", "Bodies": "name", "Director": "name", "Filipino": "name", "Fischer": "name", "Fred": "name", "Gaussian": "name", "Insect": "name", "Mom": "name", "Prepar": "name", "aspire": "word", "brickbc": "multitoken_word", "calming": "word", "careful": "word", "choice suffrage": "phrase", "comfortablyifera": "multitoken_word", "downloads": "word", "electrolyte benefited": "phrase", "equine": "word", "geospatialard": "multitoken_word", "heroine": "word", "identifierbud": "multitoken_word", "interpolation": "word", "linestyle seen": "phrase", "manners": "word", "purchases nat": "phrase", "remotely": "word", "sarc": "word", "tackle": "word", "tube": "word"}}
021/stats021.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import json, math, random, collections, sys
2
+ P = collections.defaultdict(dict)
3
+ for f in ('gpu/predictionsA.jsonl', 'gpu/predictionsB.jsonl'):
4
+ import os
5
+ if not os.path.exists(f): continue
6
+ for l in open(f):
7
+ d = json.loads(l); P[d['sys']].setdefault(d['set'], {})[d['i']] = float(d['correct'])
8
+ L = lambda s, st: [P[s][st][i] for i in sorted(P[s][st])]
9
+ def wil(a):
10
+ n = len(a); p = sum(a)/n; z = 1.96; d = 1+z*z/n; c = (p+z*z/(2*n))/d; h = z*math.sqrt(p*(1-p)/n+z*z/(4*n*n))/d; return f"{100*p:.2f} [{100*(c-h):.1f}, {100*(c+h):.1f}]"
11
+ def boot(a, b, seed=0, B=10000):
12
+ r = random.Random(seed); n = len(a); d = [x-y for x, y in zip(a, b)]; bs = sorted(sum(d[r.randrange(n)] for _ in range(n))/n*100 for _ in range(B))
13
+ return f"{sum(d)/n*100:+.2f} [{bs[int(.025*B)]:+.2f}, {bs[int(.975*B)]:+.2f}]"
14
+ mean = lambda names, st: [sum(x)/len(x) for x in zip(*[L(n, st) for n in names])]
15
+ SETS = ['sealed_id', 'sealed_ood', 'sealed_1_new', 'sealed_2_new', 'sealed_4_new']
16
+ groups = {'k2_021': [f'k2_021_s{s}' for s in (17, 29, 43)], 'k2_018': [f'k2_018_s{s}' for s in (17, 29, 43)]}
17
+ for fn in ('values', 'values_csv', 'kv'): groups[f'sp_{fn}'] = [f'sp_{fn}_s{s}' for s in (17, 29, 43)]
18
+ for p in ('ridge_aug', 'mlp_aug', 'mlp_ctrl_shuffled'): groups[p] = [f'{p}_s{s}' for s in (17, 29, 43)]
19
+ singles = ['text_template'] + [k for k in P if k.startswith('fewshot_')]
20
+ for st in SETS:
21
+ print('==', st); T = L('text_template', st)
22
+ for s in singles: print(f' {s}: {wil(L(s, st))} vs template {boot(L(s, st), T)}')
23
+ for g, ns in groups.items():
24
+ if not all(st in P[n] for n in ns): continue
25
+ per = ' / '.join(f"{100*sum(L(n, st))/len(L(n, st)):.1f}" for n in ns); m = mean(ns, st)
26
+ extra = f" vs k2_021 {boot(m, mean(groups['k2_021'], st), 2)}" if g != 'k2_021' else ''
27
+ print(f' {g}: {per} | mean {100*sum(m)/len(m):.2f} vs template {boot(m, T, 1)}{extra}')
021/stats021.txt ADDED
@@ -0,0 +1,67 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ == sealed_id
2
+ text_template: 100.00 [98.1, 100.0] vs template +0.00 [+0.00, +0.00]
3
+ fewshot_values: 34.50 [28.3, 41.3] vs template -65.50 [-72.00, -59.00]
4
+ fewshot_values_csv: 89.00 [83.9, 92.6] vs template -11.00 [-15.50, -7.00]
5
+ fewshot_kv: 94.50 [90.4, 96.9] vs template -5.50 [-9.00, -2.50]
6
+ fewshot_template: 90.00 [85.1, 93.4] vs template -10.00 [-14.50, -6.00]
7
+ k2_021: 100.0 / 99.0 / 99.5 | mean 99.50 vs template -0.50 [-1.17, +0.00]
8
+ k2_018: 100.0 / 99.5 / 97.5 | mean 99.00 vs template -1.00 [-1.83, -0.33] vs k2_021 -0.50 [-1.50, +0.50]
9
+ sp_values: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +0.00 [+0.00, +0.00] vs k2_021 +0.50 [+0.00, +1.17]
10
+ sp_values_csv: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +0.00 [+0.00, +0.00] vs k2_021 +0.50 [+0.00, +1.17]
11
+ sp_kv: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +0.00 [+0.00, +0.00] vs k2_021 +0.50 [+0.00, +1.17]
12
+ == sealed_ood
13
+ text_template: 98.50 [96.8, 99.3] vs template +0.00 [+0.00, +0.00]
14
+ fewshot_values: 32.50 [28.1, 37.2] vs template -66.00 [-70.75, -61.25]
15
+ fewshot_values_csv: 90.75 [87.5, 93.2] vs template -7.75 [-11.00, -4.75]
16
+ fewshot_kv: 91.75 [88.6, 94.1] vs template -6.75 [-9.75, -3.75]
17
+ fewshot_template: 86.50 [82.8, 89.5] vs template -12.00 [-15.50, -8.75]
18
+ k2_021: 99.8 / 98.5 / 97.8 | mean 98.67 vs template +0.17 [-1.25, +1.67]
19
+ k2_018: 99.0 / 99.0 / 97.8 | mean 98.58 vs template +0.08 [-1.17, +1.42] vs k2_021 -0.08 [-1.08, +0.92]
20
+ sp_values: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +1.50 [+0.50, +2.75] vs k2_021 +1.33 [+0.58, +2.17]
21
+ sp_values_csv: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +1.50 [+0.50, +2.75] vs k2_021 +1.33 [+0.58, +2.17]
22
+ sp_kv: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +1.50 [+0.50, +2.75] vs k2_021 +1.33 [+0.58, +2.17]
23
+ ridge_aug: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +1.50 [+0.50, +2.75] vs k2_021 +1.33 [+0.58, +2.17]
24
+ mlp_aug: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +1.50 [+0.50, +2.75] vs k2_021 +1.33 [+0.58, +2.17]
25
+ mlp_ctrl_shuffled: 0.0 / 0.0 / 0.0 | mean 0.00 vs template -98.50 [-99.50, -97.25] vs k2_021 -98.67 [-99.42, -97.83]
26
+ == sealed_1_new
27
+ text_template: 96.88 [93.4, 98.6] vs template +0.00 [+0.00, +0.00]
28
+ fewshot_values: 3.12 [1.4, 6.6] vs template -93.75 [-96.88, -90.10]
29
+ fewshot_values_csv: 50.00 [43.0, 57.0] vs template -46.88 [-54.17, -39.58]
30
+ fewshot_kv: 84.38 [78.6, 88.8] vs template -12.50 [-17.71, -7.81]
31
+ fewshot_template: 66.67 [59.7, 73.0] vs template -30.21 [-36.98, -23.44]
32
+ k2_021: 90.6 / 88.5 / 66.1 | mean 81.77 vs template -15.10 [-19.97, -10.24]
33
+ k2_018: 87.5 / 87.5 / 75.5 | mean 83.51 vs template -13.37 [-18.06, -9.03] vs k2_021 +1.74 [-1.39, +4.86]
34
+ sp_values: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +3.12 [+1.04, +5.73] vs k2_021 +18.23 [+14.41, +22.40]
35
+ sp_values_csv: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +3.12 [+1.04, +5.73] vs k2_021 +18.23 [+14.41, +22.40]
36
+ sp_kv: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +3.12 [+1.04, +5.73] vs k2_021 +18.23 [+14.41, +22.40]
37
+ ridge_aug: 89.1 / 91.1 / 88.0 | mean 89.41 vs template -7.47 [-11.98, -2.78] vs k2_021 +7.64 [+1.74, +13.54]
38
+ mlp_aug: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +3.12 [+1.04, +5.73] vs k2_021 +18.23 [+14.41, +22.40]
39
+ mlp_ctrl_shuffled: 0.0 / 0.0 / 0.0 | mean 0.00 vs template -96.88 [-98.96, -94.27] vs k2_021 -81.77 [-85.59, -77.60]
40
+ == sealed_2_new
41
+ text_template: 97.40 [94.0, 98.9] vs template +0.00 [+0.00, +0.00]
42
+ fewshot_values: 0.00 [0.0, 2.0] vs template -97.40 [-99.48, -94.79]
43
+ fewshot_values_csv: 24.48 [18.9, 31.0] vs template -72.92 [-79.17, -66.67]
44
+ fewshot_kv: 86.46 [80.9, 90.6] vs template -10.94 [-16.15, -5.73]
45
+ fewshot_template: 58.33 [51.3, 65.1] vs template -39.06 [-45.83, -32.29]
46
+ k2_021: 63.0 / 56.8 / 38.0 | mean 52.60 vs template -44.79 [-50.87, -38.89]
47
+ k2_018: 62.5 / 57.8 / 45.8 | mean 55.38 vs template -42.01 [-48.26, -35.94] vs k2_021 +2.78 [-0.87, +6.42]
48
+ sp_values: 98.4 / 99.0 / 98.4 | mean 98.61 vs template +1.22 [-1.22, +3.99] vs k2_021 +46.01 [+40.28, +51.74]
49
+ sp_values_csv: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +2.60 [+0.52, +5.21] vs k2_021 +47.40 [+41.67, +53.12]
50
+ sp_kv: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +2.60 [+0.52, +5.21] vs k2_021 +47.40 [+41.67, +53.12]
51
+ ridge_aug: 71.4 / 76.0 / 70.8 | mean 72.74 vs template -24.65 [-30.90, -18.58] vs k2_021 +20.14 [+11.98, +28.30]
52
+ mlp_aug: 98.4 / 98.4 / 96.4 | mean 97.74 vs template +0.35 [-2.08, +3.12] vs k2_021 +45.14 [+39.41, +50.69]
53
+ mlp_ctrl_shuffled: 0.0 / 0.0 / 0.0 | mean 0.00 vs template -97.40 [-99.48, -94.79] vs k2_021 -52.60 [-58.33, -46.88]
54
+ == sealed_4_new
55
+ text_template: 98.44 [95.5, 99.5] vs template +0.00 [+0.00, +0.00]
56
+ fewshot_values: 0.00 [0.0, 2.0] vs template -98.44 [-100.00, -96.35]
57
+ fewshot_values_csv: 4.69 [2.5, 8.7] vs template -93.75 [-96.88, -90.10]
58
+ fewshot_kv: 89.58 [84.5, 93.2] vs template -8.85 [-13.54, -4.69]
59
+ fewshot_template: 38.54 [31.9, 45.6] vs template -59.90 [-66.67, -53.12]
60
+ k2_021: 31.2 / 30.2 / 26.6 | mean 29.34 vs template -69.10 [-75.00, -63.02]
61
+ k2_018: 28.6 / 26.6 / 25.0 | mean 26.74 vs template -71.70 [-77.26, -65.97] vs k2_021 -2.60 [-6.25, +1.04]
62
+ sp_values: 96.9 / 96.4 / 93.2 | mean 95.49 vs template -2.95 [-5.56, -0.52] vs k2_021 +66.15 [+59.90, +72.05]
63
+ sp_values_csv: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +1.56 [+0.00, +3.65] vs k2_021 +70.66 [+64.93, +76.22]
64
+ sp_kv: 100.0 / 100.0 / 100.0 | mean 100.00 vs template +1.56 [+0.00, +3.65] vs k2_021 +70.66 [+64.93, +76.22]
65
+ ridge_aug: 43.2 / 42.2 / 47.9 | mean 44.44 vs template -53.99 [-60.24, -47.57] vs k2_021 +15.10 [+6.77, +23.44]
66
+ mlp_aug: 24.5 / 35.9 / 32.3 | mean 30.90 vs template -67.53 [-72.57, -62.15] vs k2_021 +1.56 [-5.73, +8.85]
67
+ mlp_ctrl_shuffled: 0.0 / 0.0 / 0.0 | mean 0.00 vs template -98.44 [-100.00, -96.35] vs k2_021 -29.34 [-35.07, -23.78]
CITATION.cff CHANGED
@@ -1,38 +1,42 @@
1
  cff-version: 1.2.0
2
  message: "If you use this work, please cite it as below."
3
- title: "Cuerdas (Semantic Strands): Compressing Four-Field Structured Messages into Two Latent Slots Read by a Frozen 135M-Parameter Language Model"
4
  type: software
5
  authors:
6
  - family-names: "Sanchez"
7
  given-names: "Erik"
8
  affiliation: "Independent researcher"
9
  email: "sanchezerik836@gmail.com"
10
- version: "2026-10-16"
11
  license:
12
  - MIT
13
  - CC-BY-4.0
14
  repository-artifact: "https://huggingface.co/Eriksanchezolvera/cuerdas"
15
  url: "https://huggingface.co/Eriksanchezolvera/cuerdas"
16
  abstract: >-
17
- Technical report and code for Cuerdas, a small experiment in which a
18
- trainable encoder maps four-field structured messages to K=2 latent slots
19
- read by a frozen SmolLM2-135M-Instruct. On a synthetic task, K=2 matches a
20
- full-text baseline on held-out field combinations across 3 seeds
21
- (98.75-99.75% vs 98.75%) with 51% fewer input positions. Anti-lookup tests
22
- show that decoding the slots without the LLM yields 0-2% on messages with
23
- four unseen words, against 55-73% with the LLM (text: about 97%). Fair GPU
24
- latency changes by -32.8% at batch 64 and about 0% at batch 1 and 16.
25
- A random-weight receiver scores 0%. The dataset is regenerated by make_dataset.py.
 
 
 
26
  keywords:
27
  - soft prompts
28
  - context compression
 
29
  - frozen language model
30
  - knowledge distillation
31
  - SmolLM2
32
  repository-code: "https://huggingface.co/Eriksanchezolvera/cuerdas"
33
  preferred-citation:
34
  type: report
35
- title: "Cuerdas (Semantic Strands): Compressing Four-Field Structured Messages into Two Latent Slots Read by a Frozen 135M-Parameter Language Model"
36
  authors:
37
  - family-names: "Sanchez"
38
  given-names: "Erik"
@@ -41,7 +45,7 @@ preferred-citation:
41
  year: 2026
42
  month: 10
43
  status: preprint
44
- notes: "Technical Report. Not peer reviewed. DOI pending."
45
  url: "https://huggingface.co/Eriksanchezolvera/cuerdas"
46
  license: CC-BY-4.0
47
  references:
 
1
  cff-version: 1.2.0
2
  message: "If you use this work, please cite it as below."
3
+ title: "Semantic Strands: A Pre-registered Negative Result on Learned Soft-Slot Context Compression — Trained Compact Text Wins"
4
  type: software
5
  authors:
6
  - family-names: "Sanchez"
7
  given-names: "Erik"
8
  affiliation: "Independent researcher"
9
  email: "sanchezerik836@gmail.com"
10
+ version: "2026-10-16-rev5"
11
  license:
12
  - MIT
13
  - CC-BY-4.0
14
  repository-artifact: "https://huggingface.co/Eriksanchezolvera/cuerdas"
15
  url: "https://huggingface.co/Eriksanchezolvera/cuerdas"
16
  abstract: >-
17
+ Technical report (revision 5) presenting a pre-registered negative result.
18
+ A 3.84M-parameter encoder maps four-field structured records to K=2 soft
19
+ input slots for a frozen SmolLM2-135M-Instruct (17 vs 35 positions per
20
+ event against a verbal template). On unseen combinations of known words it
21
+ matches the template on two hash-sealed tests evaluated once (99.0 vs
22
+ 99.25%; 98.7 vs 98.5%). It fails on unseen values (29% vs 98% with four
23
+ unseen names, phrases or multi-token values), a larger training pool does
24
+ not help, and compact text with a matched trained soft prompt (3.84M
25
+ parameters) scores 100% at 22 (+8 shared) positions. Probes read the slots
26
+ better than the frozen LLM. Recommendation: compare soft-slot encoders
27
+ against compact text with a matched trainable budget before claiming
28
+ compression.
29
  keywords:
30
  - soft prompts
31
  - context compression
32
+ - prompt compression
33
  - frozen language model
34
  - knowledge distillation
35
  - SmolLM2
36
  repository-code: "https://huggingface.co/Eriksanchezolvera/cuerdas"
37
  preferred-citation:
38
  type: report
39
+ title: "Semantic Strands: A Pre-registered Negative Result on Learned Soft-Slot Context Compression — Trained Compact Text Wins"
40
  authors:
41
  - family-names: "Sanchez"
42
  given-names: "Erik"
 
45
  year: 2026
46
  month: 10
47
  status: preprint
48
+ notes: "Technical Report, revision 5. Not peer reviewed. DOI pending."
49
  url: "https://huggingface.co/Eriksanchezolvera/cuerdas"
50
  license: CC-BY-4.0
51
  references:
README.md CHANGED
@@ -1,74 +1,134 @@
1
  ---
2
  license: mit
 
 
 
3
  base_model: HuggingFaceTB/SmolLM2-135M-Instruct
4
- tags: [soft-prompt, compression, research, cuerdas]
 
 
 
 
 
 
5
  ---
6
- # Cuerdas / Semantic Strands: compresión de mensajes estructurados para un LLM congelado (experimentos 013–018)
7
-
8
- ## Resumen en palabras simples
9
- Queríamos ver si un modelo de lenguaje **congelado** (SmolLM2-135M, sin reentrenarlo) puede leer un mensaje de 4 datos (acción, objetivo, estado, ventana) que llega comprimido en **muy pocas posiciones "inventadas"** en lugar de una frase de 35 posiciones.
10
- - **El mejor resultado (018):** con **2 posiciones** en vez de 35, el modelo acierta **98.75–99.75%**, igual que el texto completo (98.75%), en 3 de 3 semillas, incluso con combinaciones nunca vistas.
11
- - **Es compresión real:** sin el LLM, esas 2 posiciones no se pueden leer cuando hay palabras nuevas (0–2%), y con el LLM sí (55–73%).
12
- - **Velocidad:** con lotes grandes (64 mensajes) en GPU es **33% más rápido**. Con pocos mensajes por lote no hay ahorro.
13
- - **Límites:** con 4 palabras nuevas a la vez está lejos del texto (55–73% contra ~97%). Además, es un vocabulario pequeño y cerrado, con un solo modelo y en inglés. **No** es un protocolo semántico universal.
14
-
15
- ## Método
16
- - **Receptor:** SmolLM2-135M-Instruct, revisión `12fd25f77366fa6b3b4b768ec3050bf629380bac`, siempre congelado.
17
- - **Respuesta:** se hacen 4 consultas ("According to the record, the {field} is"). La respuesta es la ficha del propio LM con mayor puntaje entre las opciones válidas.
18
- - **Espacio:**
19
- - 013: 7×7×8×5 mensajes.
20
- - 014–018: 12×12×16×10.
21
- - OOD = parejas acción–objetivo o estado–ventana nunca vistas en el entrenamiento.
22
- - **Codecs:**
23
- - 015 "binder" K=4: una ranura por campo con el embedding congelado de la palabra del valor; aprende un *binding* de rol con vocabulario aleatorizado.
24
- - 016–018 "PackK" K=1/2: superposición aprendida de los 4 campos (rol de bajo rango + MLP conjunta), destilada del texto (016) o del binder de 015 (curriculum K=4→2→1, 017–018).
25
- - **Rigor:**
26
- - Splits disjuntos y test fresco reservado.
27
- - Checkpoints elegidos solo por validación (en 018, también con una validación de transferencia restringida a opciones válidas).
28
- - 3 semillas.
29
- - Anti-lookup: coseno contra las palabras y sonda lineal sin LLM.
30
- - Latencia con prefijo KV compartido en ambas rutas.
31
-
32
- ## Resultados (test fresco; semillas 17/29/43)
33
- | Ronda | Codec | Posiciones/evento | Fresh OOD (texto 98.75%) | 4 palabras nuevas (texto ~97%) | Sin LLM, 4 nuevas | Latencia |
34
- |---|---|---|---|---|---|---|
35
- | 014 (auditoría de la corrida de ChatGPT) | K=4 léxico | 19 | 100% | 0–2% | 100% (literal) | CPU −16% |
36
- | 015 | binder K=4 | 19 | 99.75–100% | 72–82% | 100% (literal) | CPU −10.5%; GPU b1 +1% |
37
- | 016 | K=1 directo | 16 | 94–99% | 0–45% (inestable) | 0–17% | CPU −11%; GPU b1 ±0 |
38
- | 017 | K=1 curriculum | 16 | 96–98% | 41–47% | 33–46% | — |
39
- | 017 | K=2 curriculum | 17 | 95.5–97.25% | 68–81% | 1–2% | (medida en batch sin caché compartida, inflada) |
40
- | **018** | **K=2, 3000 pasos más** | **17** | **99.5 / 99.75 / 98.75%** | 55 / 73 / 73% | **0–2%** | GPU, caché compartida: b1 +1.3%, b16 +1.6%, **b64 −32.8%** |
41
-
42
- El detalle completo está en cada `resultados-01X.md`. 013 usa otro espacio y otro modelo base (SmolLM2-135M base); ver `013/resultados-013.md`.
43
-
44
- ## Limitaciones honestas
45
- - Vocabulario pequeño y cerrado. Con valores conocidos, una sonda lineal también lee las ranuras al 100%.
46
- - Con 4 palabras nuevas o frases de varios tokens, los resultados quedan muy por debajo del texto.
47
- - Un solo receptor (con una prueba de 360M en 016, que no ayudó), solo en inglés, eventos ya estructurados (sin extracción desde lenguaje natural).
48
- - Latencia: solo hay ahorro con lotes grandes en GPU. No se midieron energía ni VRAM.
49
- - 96 mensajes por prueba de transferencia, así que diferencias de 1 a 3 puntos son ruido.
50
-
51
- ## Cómo reproducir
52
- 1. Python 3.11+, `torch` y `transformers==4.46.3`.
53
- 2. Regenera el dataset con `python make_dataset.py --out $SRC/data/expanded.json` (solo biblioteca estándar). El `expanded.json` original (hecho en la iteración 014 con ChatGPT) **no se redistribuye**; el script reimplementa el protocolo documentado y produce las mismas tuplas y los mismos índices de split (compruébalo con `--check`).
54
- 3. Descarga el receptor en la revisión indicada en `$MP`. **No uses** una copia re-guardada con transformers ≥5 si corres con 4.46.3: el `config.json` nuevo usa `rope_parameters` y 4.46.3 cae en silencio a `rope_theta=10000` (ver `019/resultados-019.md`).
55
- 4. Corre: `SRC=... MP=... OD=salida python 015/run015.py base binder4:1000`. Después `016/run016.py`, `017/run017.py` y `018/run018.py` con `INP=checkpoints`. Los `kaggle_kernel.py` muestran exactamente cómo se corrieron en una GPU Tesla T4 de Kaggle.
56
- 5. Checkpoints en `checkpoints/`: `binder4_1000_s*.pt` (015), `k1_s*.pt` (017 K=1) y `k2long_s*.pt` (018 K=2, el mejor). Se cargan con las clases `Binder` y `PackK` de los scripts.
57
- 6. Re-evaluación 019 (sin entrenar): `019/eval019.py` con `INP=checkpoints`; guarda predicciones por ejemplo, IC bootstrap pareados de las diferencias y el control de receptor con pesos aleatorios. Versiones registradas en Kaggle T4: PyTorch 2.11.0+cu128, CUDA 12.8, transformers 4.46.3, Python 3.13.15.
58
-
59
- ## Re-evaluación 019 (resumen)
60
- - Reprodujo exactamente todas las cifras de 018.
61
- - Fresh OOD, K=2 − texto (media de 3 semillas): **+0.58 pp [−0.42, +1.83]** (bootstrap pareado 95%). Con 4 palabras nuevas: −29.86 pp [−38.19, −21.53].
62
- - Receptor con pesos aleatorios: **0%** en texto y en K=2 (3 semillas).
63
- - Detalle en `019/resultados-019.md`.
64
-
65
- ## Reporte técnico
66
- `paper/cuerdas-technical-report.pdf` (y `.md`), en inglés. Cita: `CITATION.cff`.
67
-
68
- ## Licencia
69
- - **Código** (scripts, kernels, `make_dataset.py`) y checkpoints del codec: **MIT** (ver `LICENSE`).
70
- - **Texto del reporte técnico:** CC BY 4.0.
71
- - El receptor SmolLM2-135M-Instruct no se incluye; conserva su licencia Apache-2.0.
72
-
73
- ## Contacto
74
- Erik Sanchez (investigador independiente) · sanchezerik836@gmail.com
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: mit
3
+ language:
4
+ - en
5
+ - es
6
  base_model: HuggingFaceTB/SmolLM2-135M-Instruct
7
+ tags:
8
+ - soft-prompt
9
+ - context-compression
10
+ - prompt-compression
11
+ - negative-result
12
+ - pre-registered
13
+ - research
14
  ---
15
+
16
+ # Semantic Strands ("Cuerdas"): a pre-registered negative result on learned soft-slot context compression
17
+
18
+ **Technical Report, Revision 5 (16 October 2026). Not peer reviewed.**
19
+ Erik Sanchez, independent researcher · sanchezerik836@gmail.com
20
+
21
+ > **ES.** Les presento los resultados de mi primer reto personal en ciencia computacional. En este punto mi conocimiento todavía es nulo, pero lo compenso con una gran curiosidad. Espero que les sea de utilidad.
22
+ >
23
+ > **EN.** These are the results of my first personal challenge in computer science. At this point my knowledge is still close to zero, but I make up for it with a lot of curiosity. I hope you find it useful.
24
+
25
+ "Cuerdas" (Spanish for "strings") is the internal project name. It has nothing to do with string theory; the English title is *Semantic Strands*.
26
+
27
+ **Bottom line: this approach did not pay off.** A competent compact-text baseline with the same trainable budget beats it. The repository is published so that others can skip this path and reuse the controls that exposed the problem.
28
+
29
+ - Full report: [`paper/cuerdas-technical-report.pdf`](paper/cuerdas-technical-report.pdf) (Markdown source: [`paper/cuerdas-technical-report.md`](paper/cuerdas-technical-report.md))
30
+ - Citation metadata: [`CITATION.cff`](CITATION.cff)
31
+
32
+ > Note: README files from earlier revisions (1–4) presented the method as a success. That framing is withdrawn. Revision 5 supersedes it.
33
+
34
+ ## What was tried
35
+
36
+ We trained a small encoder ("PackK", K=2, 3,842,952 trainable parameters) to map four-field structured records (action, target, status, window) to **2 learned input vectors ("soft slots")** read by a **frozen** SmolLM2-135M-Instruct (revision `12fd25f77366fa6b3b4b768ec3050bf629380bac`; no weights updated, no LoRA). The goal was to replace a 20-token verbal template and so shorten the sequence from **35 to 17 positions per event** (counting a shared query prefix and four query tails).
37
+
38
+ Accuracy is closed-set, first-token, exact match over all four fields. Nothing is generated.
39
+
40
+ ## What held
41
+
42
+ On unseen combinations of **known** words, the slots match the verbal template on two pre-registered, hash-sealed tests evaluated once (seed means, n=400, 3 seeds):
43
+
44
+ | Sealed test | Verbal template (35 pos.) | K=2 slots (17 pos.) |
45
+ |---|---:|---:|
46
+ | Iteration 020 | 99.25% | 99.0% |
47
+ | Iteration 021 | 98.5% | 98.7% |
48
+
49
+ ## What failed
50
+
51
+ 1. **Unseen values.** On the 021 sealed test (unseen words, names, phrases and multi-token values), accuracy with four unseen values falls to **29%**, against **98%** for the template. Retraining with a much larger, more varied replacement pool did not help (+2.6 pp, interval includes zero).
52
+ 2. **A fair compact-text baseline wins.** With a matched trainable budget (a reparameterized soft prompt of 3,842,368 parameters; same data, steps, loss, teacher and validation-only selection), **values separated by commas score 100% on every sealed set**, including four unseen values, at 22 positions per event plus 8 shareable prompt positions. Against this baseline the slots save only 2–5 positions per event and lose up to 70.7 pp (four unseen values).
53
+ 3. **The receiver is the bottleneck.** Probes trained with replacement words read unseen values from the slots better than the frozen LLM does (for example 97.7% vs 52.6% with two unseen values). The information is in the slots; the receiver does not read it.
54
+
55
+ Zero-shot compact text fails with this receiver (0–8%), which is why comparing only against zero-shot or verbose text overstates the benefit of soft slots.
56
+
57
+ Other documented pitfalls: an inflated latency figure from a non-shared prefix (017, corrected in 018), weak probes that made the slots look non-decodable, and a re-saved `config.json` that silently changed the RoPE base under transformers 4.46.3 (see `019/resultados-019.md` and Section 7.6 of the report).
58
+
59
+ ## Recommendations for others
60
+
61
+ Before claiming compression from learned soft slots or soft tokens:
62
+
63
+ 1. Compare against **compact text with a matched trainable budget** (soft prompt, prefix or LoRA of equal size, same data, steps, loss and selection). Zero-shot or verbose text is not a fair comparison.
64
+ 2. Count positions honestly, including shared prefixes, query tails and soft-prompt positions.
65
+ 3. Seal the test: pre-register metrics and decision rules, generate the test with a new seed, hash it and evaluate once.
66
+ 4. Test unseen values of several kinds (words, names, phrases, multi-token values) and report accuracy by the number of unseen fields.
67
+ 5. Train probes on the same augmented distribution as the encoder, with shuffled-target controls.
68
+ 6. Keep a random-weight receiver control, and report per-example predictions with paired bootstrap intervals.
69
+ 7. Measure latency with a shared KV prefix on every route.
70
+ 8. Pin model revisions and configuration files.
71
+
72
+ Scope: one synthetic, closed-vocabulary, four-field task, one 135M receiver, English only. No claim is made beyond it.
73
+
74
+ ## Repository contents
75
+
76
+ | Path | What it is |
77
+ |---|---|
78
+ | `013/` | Iteration 013 (CPU, SmolLM2-135M base, smaller universe): turned out to be a lookup table |
79
+ | `014-auditoria/` | Audit of iteration 014 (run by ChatGPT) |
80
+ | `015/` … `018/` | Training scripts (`run01X.py`), exact Kaggle kernels, raw logs (`results01X.jsonl`) and result notes |
81
+ | `019/` | Re-evaluation of the 018 checkpoints: per-example predictions, paired bootstrap, random-weight receiver, RoPE-config diagnosis |
82
+ | `020/` | Pre-registration, sealed test (`sealed020.json`, SHA-256 `4c7737f6…3393`), compact-text baselines and stronger probes |
83
+ | `021/` | Pre-registration, second sealed test (`sealed021.json`, SHA-256 `a86a6a5e…72bc`), retraining with a varied pool (`c021.py`, `evalA.py`) and matched soft-prompt baselines (`evalB.py`), logs, predictions and statistics |
84
+ | `checkpoints/` | `binder4_1000_s*.pt` (015 teacher), `k1_s*.pt` (017 K=1), `k2_s*.pt` (017 K=2, start point for 018 and 021), `k2long_s*.pt` (018 K=2), `k2v_s*.pt` (021 K=2 retrained), `sp_{values,values_csv,kv}_s*.pt` (021 soft-prompt baselines) |
85
+ | `make_dataset.py` | Regenerates the 014 dataset (standard library only) |
86
+ | `paper/` | Technical report, revision 5 (PDF, Markdown, Figure 1) |
87
+
88
+ The checkpoints are small encoder state dicts, **not** standalone language models. They are loaded with the `Binder` / `PackK` classes in `018/run018.py` (and the soft-prompt class in `021/evalB.py`), together with the frozen receiver. Seeds: 17, 29, 43.
89
+
90
+ ## How to reproduce
91
+
92
+ 1. Python 3.11+, PyTorch and `transformers==4.46.3`. The 019 re-evaluation on a Kaggle Tesla T4 recorded PyTorch 2.11.0+cu128, CUDA 12.8, Python 3.13.15, and reproduced every 018 accuracy exactly.
93
+ 2. Regenerate the dataset: `python make_dataset.py --out $SRC/data/expanded.json` (verify with `--check`). The original `expanded.json`, generated in iteration 014 with ChatGPT, is **not redistributed**; the script produces the same tuples and split indices.
94
+ 3. Download the receiver at the pinned revision into `$MP`. Do **not** load a copy re-saved by transformers ≥ 5 with transformers 4.46.3: the RoPE base silently falls back to 10000.
95
+ 4. Training: `SRC=... MP=... OD=out python 015/run015.py base binder4:1000`, then `016/run016.py`, `017/run017.py` and `018/run018.py` with `INP=checkpoints`.
96
+ 5. Evaluations without training: `019/eval019.py` and `020/eval020.py` (`INP` must contain the checkpoints, `run018.py` and the sealed file). Iteration 021: `021/evalA.py` (retraining, needs `c021.py`) and `021/evalB.py` (soft-prompt baselines). The `kaggle_kernel*.py` files show exactly how each was run on a Kaggle T4.
97
+
98
+ Compute: about 2.6 Kaggle T4 GPU-hours for 015–020 (free tier; Section 4.7 of the report), plus CPU runs for 013–015.
99
+
100
+ ## Use of AI tools
101
+
102
+ AI systems are not authors. Grok Bot (AI assistant) helped design the experiments, wrote and ran most of the code and drafted the report. ChatGPT ran iteration 014, generated its dataset and reviewed earlier revisions. The author set the goals and acceptance criteria, chose what to run next and approved the claims. Details: Section 11 of the report.
103
+
104
+ ## License
105
+
106
+ - **Code** (scripts, kernels, `make_dataset.py`) and **encoder checkpoints**: MIT (see `LICENSE`).
107
+ - **Report text** (including Figure 1): CC BY 4.0.
108
+ - SmolLM2-135M-Instruct is not included and keeps its Apache-2.0 license. The original ChatGPT-generated 014 dataset is not redistributed.
109
+
110
+ ## How to cite
111
+
112
+ > Sanchez, E. (2026). *Semantic Strands: A Pre-registered Negative Result on Learned Soft-Slot Context Compression — Trained Compact Text Wins.* Technical Report, revision 5. https://huggingface.co/Eriksanchezolvera/cuerdas
113
+
114
+ DOI: pending
115
+
116
+ ```bibtex
117
+ @techreport{sanchez2026cuerdas,
118
+ author = {Sanchez, Erik},
119
+ title = {Semantic Strands: A Pre-registered Negative Result on
120
+ Learned Soft-Slot Context Compression --- Trained
121
+ Compact Text Wins},
122
+ institution = {Independent researcher},
123
+ type = {Technical Report},
124
+ number = {Revision 5},
125
+ year = {2026},
126
+ month = oct,
127
+ url = {https://huggingface.co/Eriksanchezolvera/cuerdas},
128
+ note = {DOI pending}
129
+ }
130
+ ```
131
+
132
+ ## Contact
133
+
134
+ Erik Sanchez (independent researcher) · sanchezerik836@gmail.com
checkpoints/k2_s17.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b7deb5aead22c265b1fd11bba19c00768e1f5d1d0d63faa92f0555fc8883c659
3
+ size 15375601
checkpoints/k2_s29.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a06de687aa3de8e5252f58f6b403119be2810b0b157a1c85541b70e55496566e
3
+ size 15375601
checkpoints/k2_s43.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:19f665e0d16688e253b7a1c98c48803fd6cd23babc10f02f39586e3b98676418
3
+ size 15375601
checkpoints/k2v_s17.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e9666602831ae1954dfa26871ec6918670f5171f0a340d5f95168d9da4192788
3
+ size 15375746
checkpoints/k2v_s29.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6fe4e70e747e29f3f7f2b57f3766489eecc5c74d03d2010b8a1402b647faa5b2
3
+ size 15375746
checkpoints/k2v_s43.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e9df3ebcfaeea902c3d05e4c8b736d366b4cf260b5773e4ed31d08d586e76e0c
3
+ size 15375746
checkpoints/sp_kv_s17.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a64f53e9416380b0064e4b2b79446ff48027867c4b2902f570358099366857d6
3
+ size 15372335
checkpoints/sp_kv_s29.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:22cbed2f6f231d2c6f17d94d81ce69a8ffd6234ec9f20e895cef136c0155e82e
3
+ size 15372335
checkpoints/sp_kv_s43.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3568c35b04a42bdb2f80838137987674e9fe69526db090b5e5047ed178b5b2b1
3
+ size 15372335
checkpoints/sp_values_csv_s17.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:99e8b81d05f974f38cc464c4c72c3ae2ed6941b61be77c2a0b31428e3ee0aaf6
3
+ size 15372423
checkpoints/sp_values_csv_s29.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9a7e69db920cc50f1b8f17f1afab805ba3e85b645f0348718b4dbbf5722f3b7e
3
+ size 15372423
checkpoints/sp_values_csv_s43.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:13979deb620a6836805c017c1263016a85a857ff420dbbd0e34d51cc7a6d01e0
3
+ size 15372423
checkpoints/sp_values_s17.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b57f329cef963875aaceaeab5c4492baec2a136c98751cb4eb5145cd73f6ab8b
3
+ size 15372379
checkpoints/sp_values_s29.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6f4b9b8ee1a766a9756446063f77fa63e28ce54de7606073997195e68f80d561
3
+ size 15372379
checkpoints/sp_values_s43.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0db03dbc657a69071a20edf34c5595678e5cb6df7ce509646ec2a4f548dfcd5d
3
+ size 15372379
paper/cuerdas-technical-report.md CHANGED
The diff for this file is too large to render. See raw diff
 
paper/cuerdas-technical-report.pdf CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d4a2b8a24692cf004b7f158e2c66e0d73dd128419ed7ae714d6966c910581421
3
- size 121577
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c2ac543fc4f49adf33d90c598aa1f0332bcf75e7070f462a98c241a68e217797
3
+ size 181206
paper/fig_architecture.svg ADDED