wuff-mann commited on
Commit
eae4d2a
·
verified ·
1 Parent(s): 60daba0

Stage2R-B2 bilingual Global Memory requalification 20260906T122804Z — B2_REMOTE_RULER_INVALID_ENGLISH_CONTROL_FAILED

Browse files
experiments/stage2r/b2/20260906T122804Z/README_STAGE2R_B2.md ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # MANN-GPT Stage2R-B2 — Bilingual Global Memory Requalification
2
+
3
+ Run: 20260906T122804Z
4
+
5
+ Formal parent:
6
+ - checkpoint/stage2r/a4b_q2/20260905T090918Z/winner/model.safetensors
7
+ - SHA256: bd0d06cf098ea2b839fa3657f9f36af7fa41c3878b5ce7bbfaa310fd38c72bd7
8
+
9
+ B1 semantic bridge:
10
+ - run: 20260906T113825Z
11
+ - extension SHA256: ffea0e0ba10fbc1d5e3bbc661aefe7bddaa602d0681a41117ef6011edfe51761
12
+ - verdict: B1_NEW_ROW_SEMANTIC_BRIDGE_QUALIFIED_MOVE_TO_B2_BILINGUAL_REQUALIFICATION
13
+
14
+ Identity:
15
+ - hidden maxdiff: 0.0
16
+ - legacy logit-slice maxdiff: 0.0
17
+
18
+ Causality:
19
+ - qualified: True
20
+
21
+ Remote summary:
22
+ {"EN_EN": {"n": 18, "gain": 0.001761051981981259, "target_penalty": 0.04289200570848253, "specificity": 0.08781719207763672, "full_nll": 9.313207387924194, "beta0_nll": 9.326774915059408, "by_len": {"1024": {"gain": 0.0183153584880076, "target_penalty": -0.011800765991210938, "specificity": 0.0925300121307373}, "1536": {"gain": 0.00017831405575348606, "target_penalty": 0.03487245241800944, "specificity": 0.05888803799947103}, "2048": {"gain": -0.013210516597817315, "target_penalty": 0.1056043306986491, "specificity": 0.11203352610270183}}}, "ZH_ZH": {"n": 18, "gain": 0.018074180170252786, "target_penalty": 0.03611124886406793, "specificity": 0.015738593207465276, "full_nll": 11.472676912943522, "beta0_nll": 11.675063133239746, "by_len": {"1024": {"gain": 0.019140431870394967, "target_penalty": 0.059359073638916016, "specificity": 0.03464857737223307}, "1536": {"gain": 0.028962117676235553, "target_penalty": 0.031238238016764324, "specificity": -0.013074239095052084}, "2048": {"gain": 0.006119990964127833, "target_penalty": 0.017736434936523438, "specificity": 0.025641441345214844}}}, "ZH_EN": {"n": 18, "gain": 0.0038233656745104306, "target_penalty": 0.1322231822543674, "specificity": 0.11281657218933105, "full_nll": 9.066032462649876, "beta0_nll": 9.110931369993422, "by_len": {"1024": {"gain": 0.026020476014022045, "target_penalty": 0.09061082204182942, "specificity": 0.06905301411946614}, "1536": {"gain": -0.01407167885487304, "target_penalty": 0.23560492197672525, "specificity": 0.24075047175089517}, "2048": {"gain": -0.0004787001356177147, "target_penalty": 0.07045380274454753, "specificity": 0.028646230697631836}}}, "EN_ZH": {"n": 18, "gain": 0.010407720666442612, "target_penalty": -0.0018983417087131077, "specificity": 0.0001796616448296441, "full_nll": 11.652510378095839, "beta0_nll": 11.772504170735678, "by_len": {"1024": {"gain": 0.009588295536660428, "target_penalty": 0.0014437039693196614, "specificity": 0.01219336191813151}, "1536": {"gain": 0.010254239325024093, "target_penalty": -0.0020531018575032554, "specificity": -0.011034965515136719}, "2048": {"gain": 0.01138062713764331, "target_penalty": -0.0050856272379557295, "specificity": -0.0006194114685058594}}}}
23
+
24
+ Writer diagnostic:
25
+ {"EN_EN": {"n": 6, "source_advantage": -0.019238902100235048}, "ZH_ZH": {"n": 6, "source_advantage": 0.01275022079200055}, "ZH_EN": {"n": 6, "source_advantage": 0.023772776524905018}, "EN_ZH": {"n": 6, "source_advantage": 0.0059973296267139264}}
26
+
27
+ Semantic cross-language diagnostic:
28
+ {"EN_ZH_SEM": {"gain": -0.0031344768144604194, "beta0_nll": 13.297510504722595, "global_nll": 13.342041015625}, "ZH_EN_SEM": {"gain": 0.12379684830955245, "beta0_nll": 9.510125041007996, "global_nll": 8.339370965957642}}
29
+
30
+ Gates:
31
+ {"identity_ok": true, "causality_ok": true, "safety_ok": true, "english_control_ok": false, "zh_zh_ok": true, "zh_en_ok": false, "en_zh_ok": false, "bilingual_transfer_ok": false, "qualified": false}
32
+
33
+ Verdict:
34
+ B2_REMOTE_RULER_INVALID_ENGLISH_CONTROL_FAILED
35
+
36
+ B2 does not train any parameter.
experiments/stage2r/b2/20260906T122804Z/b2_bilingual_remote_counterfactual.csv ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ mode,seq_len,beta0_nll,full_nll,target_ko_nll,control_ko_nll,gain,target_penalty,specificity
2
+ EN_EN,1024,10.749608993530273,10.542425155639648,10.541848182678223,10.457904815673828,0.019273616185976627,-0.0005769729614257812,0.08394336700439453
3
+ EN_EN,1024,9.27511215209961,9.796390533447266,9.157304763793945,9.932697296142578,-0.05620184131462543,-0.6390857696533203,-0.7753925323486328
4
+ EN_EN,1024,10.968110084533691,10.549010276794434,10.564669609069824,10.534157752990723,0.03821075869125686,0.015659332275390625,0.030511856079101562
5
+ EN_EN,1024,10.919503211975098,10.743886947631836,10.83105754852295,10.592069625854492,0.01608280715103124,0.08717060089111328,0.23898792266845703
6
+ EN_EN,1024,9.607673645019531,9.628615379333496,9.633077621459961,9.19324779510498,-0.0021796883499285713,0.004462242126464844,0.43982982635498047
7
+ EN_EN,1024,8.772396087646484,7.941593170166016,8.403159141540527,7.865859508514404,0.09470649856433488,0.4615659713745117,0.537299633026123
8
+ EN_EN,1536,8.839607238769531,8.806118965148926,8.898093223571777,8.791816711425781,0.003788434566835689,0.09197425842285156,0.1062765121459961
9
+ EN_EN,1536,7.719666004180908,7.4812912940979,7.52534294128418,7.458940029144287,0.030878889054773356,0.0440516471862793,0.06640291213989258
10
+ EN_EN,1536,9.721976280212402,9.533473014831543,9.542749404907227,9.611867904663086,0.019389397787827254,0.009276390075683594,-0.06911849975585938
11
+ EN_EN,1536,10.442026138305664,10.512548446655273,10.536829948425293,10.484184265136719,-0.006753699657090919,0.02428150177001953,0.05264568328857422
12
+ EN_EN,1536,11.325526237487793,11.793264389038086,11.804014205932617,11.780308723449707,-0.041299462977893886,0.01074981689453125,0.023705482482910156
13
+ EN_EN,1536,7.575194358825684,7.612567901611328,7.6414690017700195,7.468052864074707,-0.004933674439930572,0.028901100158691406,0.1734161376953125
14
+ EN_EN,2048,7.536928176879883,7.601231098175049,7.63449764251709,7.615623950958252,-0.008531714749839365,0.033266544342041016,0.01887369155883789
15
+ EN_EN,2048,7.65134859085083,7.303419589996338,7.678154945373535,7.282966136932373,0.04547289889138386,0.37473535537719727,0.3951888084411621
16
+ EN_EN,2048,8.180591583251953,8.72011947631836,8.72606086730957,8.727821350097656,-0.06595218543497228,0.0059413909912109375,-0.0017604827880859375
17
+ EN_EN,2048,8.310962677001953,8.620610237121582,8.732303619384766,8.593276023864746,-0.03725772478517848,0.1116933822631836,0.13902759552001953
18
+ EN_EN,2048,10.798026084899902,11.14548110961914,11.191671371459961,11.129679679870605,-0.032177642653144156,0.04619026184082031,0.06199169158935547
19
+ EN_EN,2048,9.487690925598145,9.305685997009277,9.367485046386719,9.308605194091797,0.01918326914484652,0.061799049377441406,0.058879852294921875
20
+ ZH_ZH,1024,11.7988862991333,11.62775993347168,11.624785423278809,11.618570327758789,0.014503603248909295,-0.0029745101928710938,0.006215095520019531
21
+ ZH_ZH,1024,11.909625053405762,11.55871868133545,11.726059913635254,11.599889755249023,0.029464099037271103,0.1673412322998047,0.12617015838623047
22
+ ZH_ZH,1024,11.944084167480469,11.765787124633789,11.777177810668945,11.773270606994629,0.014927644543239213,0.01139068603515625,0.003907203674316406
23
+ ZH_ZH,1024,11.881986618041992,11.555014610290527,11.7373628616333,11.553999900817871,0.027518294563215546,0.18234825134277344,0.1833629608154297
24
+ ZH_ZH,1024,13.137389183044434,12.956341743469238,12.951815605163574,13.028634071350098,0.01378108215054338,-0.0045261383056640625,-0.07681846618652344
25
+ ZH_ZH,1024,13.04588508605957,12.854790687561035,12.857365608215332,12.892311096191406,0.014647867679191252,0.002574920654296875,-0.03494548797607422
26
+ ZH_ZH,1536,10.034578323364258,9.681558609008789,9.694668769836426,9.640292167663574,0.03518032377439385,0.013110160827636719,0.05437660217285156
27
+ ZH_ZH,1536,11.426116943359375,11.141769409179688,11.15112018585205,11.145821571350098,0.02488575389077778,0.009350776672363281,0.005298614501953125
28
+ ZH_ZH,1536,13.01644515991211,12.826310157775879,12.924354553222656,12.832344055175781,0.014607290992306099,0.09804439544677734,0.092010498046875
29
+ ZH_ZH,1536,9.820341110229492,9.341851234436035,9.403997421264648,9.404776573181152,0.04872436409515669,0.06214618682861328,-0.0007791519165039062
30
+ ZH_ZH,1536,11.76656436920166,11.625513076782227,11.630002975463867,11.624889373779297,0.01198746617905118,0.004489898681640625,0.0051136016845703125
31
+ ZH_ZH,1536,9.7482328414917,9.374022483825684,9.374310493469238,9.608776092529297,0.03838750712572772,0.0002880096435546875,-0.2344655990600586
32
+ ZH_ZH,2048,9.94054126739502,9.84298324584961,9.84619140625,9.842761993408203,0.009814155881571612,0.003208160400390625,0.003429412841796875
33
+ ZH_ZH,2048,12.685158729553223,12.710062026977539,12.767237663269043,12.70713996887207,-0.001963183745292678,0.057175636291503906,0.060097694396972656
34
+ ZH_ZH,2048,12.928396224975586,12.81466293334961,12.81389045715332,12.817079544067383,0.008797169397257651,-0.0007724761962890625,-0.0031890869140625
35
+ ZH_ZH,2048,12.017095565795898,11.876530647277832,11.878580093383789,11.843847274780273,0.011697079194256763,0.0020494461059570312,0.034732818603515625
36
+ ZH_ZH,2048,11.90754508972168,11.876608848571777,11.911111831665039,11.876213073730469,0.0025980368679523877,0.03450298309326172,0.03489875793457031
37
+ ZH_ZH,2048,11.142264366149902,11.077898979187012,11.088153839111328,11.064274787902832,0.0057766881890212625,0.010254859924316406,0.023879051208496094
38
+ ZH_EN,1024,7.8295183181762695,7.513352394104004,7.788667678833008,7.410763740539551,0.04038127394609764,0.2753152847290039,0.37790393829345703
39
+ ZH_EN,1024,10.40095043182373,10.093021392822266,10.205775260925293,10.249237060546875,0.029605855832107043,0.11275386810302734,-0.04346179962158203
40
+ ZH_EN,1024,9.567639350891113,9.29701042175293,9.35739517211914,9.247150421142578,0.028285862291932824,0.06038475036621094,0.1102447509765625
41
+ ZH_EN,1024,10.586263656616211,10.232487678527832,10.24944019317627,10.166850090026855,0.03341839855530858,0.0169525146484375,0.08259010314941406
42
+ ZH_EN,1024,7.57728385925293,7.583877086639404,7.5628437995910645,7.785764217376709,-0.0008701307102839167,-0.021033287048339844,-0.22292041778564453
43
+ ZH_EN,1024,9.180891990661621,8.948600769042969,9.047892570495605,8.937931060791016,0.0253015961689701,0.09929180145263672,0.10996150970458984
44
+ ZH_EN,1536,9.517528533935547,9.385146141052246,9.575587272644043,9.473428726196289,0.013909324507016737,0.19044113159179688,0.1021585464477539
45
+ ZH_EN,1536,8.702055931091309,8.44572925567627,8.87137508392334,8.427855491638184,0.02945587542125733,0.4256458282470703,0.44351959228515625
46
+ ZH_EN,1536,11.168956756591797,11.456048011779785,11.676727294921875,11.261507034301758,-0.025704393117875594,0.22067928314208984,0.4152202606201172
47
+ ZH_EN,1536,9.489741325378418,9.362833023071289,9.624134063720703,9.376943588256836,0.013373209864818759,0.26130104064941406,0.2471904754638672
48
+ ZH_EN,1536,8.351749420166016,9.191099166870117,9.311279296875,9.252742767333984,-0.10049987188042539,0.12018013000488281,0.058536529541015625
49
+ ZH_EN,1536,9.70090103149414,9.846067428588867,10.041449546813965,9.863572120666504,-0.014964217924030085,0.19538211822509766,0.17787742614746094
50
+ ZH_EN,2048,7.864268779754639,7.794396877288818,7.857064723968506,7.793521404266357,0.00888472970884399,0.0626678466796875,0.06354331970214844
51
+ ZH_EN,2048,9.951677322387695,9.690244674682617,9.895607948303223,9.70632553100586,0.02627020945674642,0.20536327362060547,0.18928241729736328
52
+ ZH_EN,2048,8.963698387145996,9.227341651916504,9.205081939697266,9.241612434387207,-0.029412331091882123,-0.02225971221923828,-0.036530494689941406
53
+ ZH_EN,2048,7.694599151611328,7.807319164276123,7.81449556350708,7.793223857879639,-0.014649237789234309,0.007176399230957031,0.021271705627441406
54
+ ZH_EN,2048,7.6500654220581055,7.920267581939697,7.801901340484619,7.937970161437988,-0.03532024172008963,-0.11836624145507812,-0.13606882095336914
55
+ ZH_EN,2048,9.798974990844727,9.393741607666016,9.681882858276367,9.611503601074219,0.04135467062190936,0.28814125061035156,0.07037925720214844
56
+ EN_ZH,1024,11.90514850616455,11.837657928466797,11.847641944885254,11.82481861114502,0.005669024427776514,0.009984016418457031,0.022823333740234375
57
+ EN_ZH,1024,11.99621868133545,11.946331977844238,11.981805801391602,11.949453353881836,0.004158535686651673,0.03547382354736328,0.032352447509765625
58
+ EN_ZH,1024,11.197628021240234,11.067214012145996,11.050483703613281,11.06206226348877,0.011646574510857327,-0.016730308532714844,-0.011578559875488281
59
+ EN_ZH,1024,12.009136199951172,11.845398902893066,11.852136611938477,11.860312461853027,0.01363439420886668,0.006737709045410156,-0.008175849914550781
60
+ EN_ZH,1024,12.058638572692871,11.896965980529785,11.891292572021484,11.857430458068848,0.013407201085634833,-0.005673408508300781,0.03386211395263672
61
+ EN_ZH,1024,12.060737609863281,11.952021598815918,11.930891990661621,11.92701530456543,0.009014043300175542,-0.021129608154296875,0.0038766860961914062
62
+ EN_ZH,1536,11.005392074584961,10.97523021697998,10.971362113952637,10.983627319335938,0.002740643622741441,-0.00386810302734375,-0.012265205383300781
63
+ EN_ZH,1536,11.964938163757324,11.86760139465332,11.866427421569824,11.845449447631836,0.008135166916185503,-0.0011739730834960938,0.02097797393798828
64
+ EN_ZH,1536,11.993650436401367,11.917298316955566,11.879707336425781,11.917051315307617,0.00636604508782981,-0.037590980529785156,-0.03734397888183594
65
+ EN_ZH,1536,9.928557395935059,9.693144798278809,9.71164608001709,9.708553314208984,0.023710654858341496,0.01850128173828125,0.0030927658081054688
66
+ EN_ZH,1536,12.922861099243164,12.789019584655762,12.794496536254883,12.797893524169922,0.010356956834832873,0.005476951599121094,-0.0033969879150390625
67
+ EN_ZH,1536,13.386674880981445,13.249917030334473,13.256253242492676,13.293527603149414,0.010215968630213438,0.006336212158203125,-0.03727436065673828
68
+ EN_ZH,2048,13.067452430725098,13.01659870147705,12.993956565856934,13.017291069030762,0.0038916330109207875,-0.022642135620117188,-0.023334503173828125
69
+ EN_ZH,2048,9.82612133026123,9.648903846740723,9.65003776550293,9.629632949829102,0.018035344523452616,0.0011339187622070312,0.020404815673828125
70
+ EN_ZH,2048,12.124834060668945,11.845436096191406,11.854303359985352,11.850503921508789,0.023043446457041594,0.008867263793945312,0.0037994384765625
71
+ EN_ZH,2048,11.157578468322754,11.024057388305664,11.009932518005371,11.02387523651123,0.011966851086565667,-0.014124870300292969,-0.013942718505859375
72
+ EN_ZH,2048,12.209148406982422,12.19516372680664,12.195716857910156,12.189757347106934,0.0011454263401191355,0.000553131103515625,0.005959510803222656
73
+ EN_ZH,2048,11.09035873413086,10.977225303649902,10.97292423248291,10.969527244567871,0.010201061407760061,-0.0043010711669921875,0.0033969879150390625
experiments/stage2r/b2/20260906T122804Z/b2_same_shape_causality.csv ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ kind,seq_len,max_abs_prefix_logit_diff,mean_abs_prefix_logit_diff,new_ids_in_prefix
2
+ OLD,512,0.0,0.0,0
3
+ OLD,1024,0.0,0.0,0
4
+ OLD,2048,0.0,0.0,0
5
+ CHINESE,512,0.0,0.0,212
6
+ CHINESE,1024,0.0,0.0,424
7
+ CHINESE,2048,0.0,0.0,848
8
+ MIXED,512,0.0,0.0,72
9
+ MIXED,1024,0.0,0.0,152
10
+ MIXED,2048,0.0,0.0,297
experiments/stage2r/b2/20260906T122804Z/b2_semantic_cross_language_diagnostic.csv ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ direction,beta0_nll,global_nll,gain
2
+ ZH_EN_SEM,9.275351524353027,8.429616928100586,0.0911808672729989
3
+ ZH_EN_SEM,10.015918731689453,8.868841171264648,0.11452544605773966
4
+ ZH_EN_SEM,10.092586517333984,9.076164245605469,0.10070979029833568
5
+ ZH_EN_SEM,8.74323844909668,7.404326438903809,0.15313685174984593
6
+ ZH_EN_SEM,9.41280746459961,7.557804107666016,0.19707227242350692
7
+ ZH_EN_SEM,8.432592391967773,7.433209419250488,0.11851432231791721
8
+ ZH_EN_SEM,10.015918731689453,8.868841171264648,0.11452544605773966
9
+ ZH_EN_SEM,10.092586517333984,9.076164245605469,0.10070979029833568
10
+ EN_ZH_SEM,12.556635856628418,12.586727142333984,-0.0023964448797550953
11
+ EN_ZH_SEM,12.6200532913208,12.664470672607422,-0.0035195874582533097
12
+ EN_ZH_SEM,12.553155899047852,12.522358894348145,0.002453327668944426
13
+ EN_ZH_SEM,14.30704116821289,14.502897262573242,-0.01368948981537153
14
+ EN_ZH_SEM,13.10323715209961,13.039129257202148,0.004892523439308167
15
+ EN_ZH_SEM,13.516977310180664,13.525676727294922,-0.0006435919003655958
16
+ EN_ZH_SEM,14.053777694702148,14.261920928955078,-0.014810482901788992
17
+ EN_ZH_SEM,13.669205665588379,13.633147239685059,0.0026379313315985733
experiments/stage2r/b2/20260906T122804Z/hf_write_probe.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ stage2r-b2 write probe 20260906T122804Z
experiments/stage2r/b2/20260906T122804Z/stage2r_b2_summary.json ADDED
@@ -0,0 +1,246 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "run_id": "20260906T122804Z",
3
+ "formal_parent": {
4
+ "path": "checkpoint/stage2r/a4b_q2/20260905T090918Z/winner/model.safetensors",
5
+ "sha256": "bd0d06cf098ea2b839fa3657f9f36af7fa41c3878b5ce7bbfaa310fd38c72bd7"
6
+ },
7
+ "b1_reference": {
8
+ "run_id": "20260906T113825Z",
9
+ "verdict": "B1_NEW_ROW_SEMANTIC_BRIDGE_QUALIFIED_MOVE_TO_B2_BILINGUAL_REQUALIFICATION",
10
+ "extension_sha256": "ffea0e0ba10fbc1d5e3bbc661aefe7bddaa602d0681a41117ef6011edfe51761",
11
+ "b1_metrics": {
12
+ "zh_heldout_improvement": 0.11485626104849815,
13
+ "en_regression": -0.004029283458668341,
14
+ "code_regression": -0.0033570897675795305,
15
+ "train_stream_improvement": 0.06114005260756121,
16
+ "old_parent_zh_nll_per_char": 6.105846285820007,
17
+ "pre_b0r1_zh_nll_per_char": 6.836660647392273,
18
+ "post_b1_zh_nll_per_char": 5.99948263168335,
19
+ "new_rows_maxdiff": 0.019636046141386032,
20
+ "new_rows_rms_delta": 0.006740771699696779,
21
+ "legacy_embedding_maxdiff": 0.0,
22
+ "legacy_hidden_maxdiff": 0.0,
23
+ "legacy_logit_slice_maxdiff": 0.0,
24
+ "frozen_sample_maxdiff": {
25
+ "semantic_adapter.0.weight": 0.0,
26
+ "layers.24.connection_network.q_proj.weight": 0.0,
27
+ "layers.24.propagation.gate.weight": 0.0,
28
+ "thinking_net.router.weight": 0.0,
29
+ "norm.weight": 0.0
30
+ }
31
+ }
32
+ },
33
+ "legacy_identity": {
34
+ "hidden_maxdiff": 0.0,
35
+ "legacy_logit_slice_maxdiff": 0.0
36
+ },
37
+ "causality": [
38
+ {
39
+ "kind": "OLD",
40
+ "seq_len": 512,
41
+ "max_abs_prefix_logit_diff": 0.0,
42
+ "mean_abs_prefix_logit_diff": 0.0,
43
+ "new_ids_in_prefix": 0
44
+ },
45
+ {
46
+ "kind": "OLD",
47
+ "seq_len": 1024,
48
+ "max_abs_prefix_logit_diff": 0.0,
49
+ "mean_abs_prefix_logit_diff": 0.0,
50
+ "new_ids_in_prefix": 0
51
+ },
52
+ {
53
+ "kind": "OLD",
54
+ "seq_len": 2048,
55
+ "max_abs_prefix_logit_diff": 0.0,
56
+ "mean_abs_prefix_logit_diff": 0.0,
57
+ "new_ids_in_prefix": 0
58
+ },
59
+ {
60
+ "kind": "CHINESE",
61
+ "seq_len": 512,
62
+ "max_abs_prefix_logit_diff": 0.0,
63
+ "mean_abs_prefix_logit_diff": 0.0,
64
+ "new_ids_in_prefix": 212
65
+ },
66
+ {
67
+ "kind": "CHINESE",
68
+ "seq_len": 1024,
69
+ "max_abs_prefix_logit_diff": 0.0,
70
+ "mean_abs_prefix_logit_diff": 0.0,
71
+ "new_ids_in_prefix": 424
72
+ },
73
+ {
74
+ "kind": "CHINESE",
75
+ "seq_len": 2048,
76
+ "max_abs_prefix_logit_diff": 0.0,
77
+ "mean_abs_prefix_logit_diff": 0.0,
78
+ "new_ids_in_prefix": 848
79
+ },
80
+ {
81
+ "kind": "MIXED",
82
+ "seq_len": 512,
83
+ "max_abs_prefix_logit_diff": 0.0,
84
+ "mean_abs_prefix_logit_diff": 0.0,
85
+ "new_ids_in_prefix": 72
86
+ },
87
+ {
88
+ "kind": "MIXED",
89
+ "seq_len": 1024,
90
+ "max_abs_prefix_logit_diff": 0.0,
91
+ "mean_abs_prefix_logit_diff": 0.0,
92
+ "new_ids_in_prefix": 152
93
+ },
94
+ {
95
+ "kind": "MIXED",
96
+ "seq_len": 2048,
97
+ "max_abs_prefix_logit_diff": 0.0,
98
+ "mean_abs_prefix_logit_diff": 0.0,
99
+ "new_ids_in_prefix": 297
100
+ }
101
+ ],
102
+ "remote_summary": {
103
+ "EN_EN": {
104
+ "n": 18,
105
+ "gain": 0.001761051981981259,
106
+ "target_penalty": 0.04289200570848253,
107
+ "specificity": 0.08781719207763672,
108
+ "full_nll": 9.313207387924194,
109
+ "beta0_nll": 9.326774915059408,
110
+ "by_len": {
111
+ "1024": {
112
+ "gain": 0.0183153584880076,
113
+ "target_penalty": -0.011800765991210938,
114
+ "specificity": 0.0925300121307373
115
+ },
116
+ "1536": {
117
+ "gain": 0.00017831405575348606,
118
+ "target_penalty": 0.03487245241800944,
119
+ "specificity": 0.05888803799947103
120
+ },
121
+ "2048": {
122
+ "gain": -0.013210516597817315,
123
+ "target_penalty": 0.1056043306986491,
124
+ "specificity": 0.11203352610270183
125
+ }
126
+ }
127
+ },
128
+ "ZH_ZH": {
129
+ "n": 18,
130
+ "gain": 0.018074180170252786,
131
+ "target_penalty": 0.03611124886406793,
132
+ "specificity": 0.015738593207465276,
133
+ "full_nll": 11.472676912943522,
134
+ "beta0_nll": 11.675063133239746,
135
+ "by_len": {
136
+ "1024": {
137
+ "gain": 0.019140431870394967,
138
+ "target_penalty": 0.059359073638916016,
139
+ "specificity": 0.03464857737223307
140
+ },
141
+ "1536": {
142
+ "gain": 0.028962117676235553,
143
+ "target_penalty": 0.031238238016764324,
144
+ "specificity": -0.013074239095052084
145
+ },
146
+ "2048": {
147
+ "gain": 0.006119990964127833,
148
+ "target_penalty": 0.017736434936523438,
149
+ "specificity": 0.025641441345214844
150
+ }
151
+ }
152
+ },
153
+ "ZH_EN": {
154
+ "n": 18,
155
+ "gain": 0.0038233656745104306,
156
+ "target_penalty": 0.1322231822543674,
157
+ "specificity": 0.11281657218933105,
158
+ "full_nll": 9.066032462649876,
159
+ "beta0_nll": 9.110931369993422,
160
+ "by_len": {
161
+ "1024": {
162
+ "gain": 0.026020476014022045,
163
+ "target_penalty": 0.09061082204182942,
164
+ "specificity": 0.06905301411946614
165
+ },
166
+ "1536": {
167
+ "gain": -0.01407167885487304,
168
+ "target_penalty": 0.23560492197672525,
169
+ "specificity": 0.24075047175089517
170
+ },
171
+ "2048": {
172
+ "gain": -0.0004787001356177147,
173
+ "target_penalty": 0.07045380274454753,
174
+ "specificity": 0.028646230697631836
175
+ }
176
+ }
177
+ },
178
+ "EN_ZH": {
179
+ "n": 18,
180
+ "gain": 0.010407720666442612,
181
+ "target_penalty": -0.0018983417087131077,
182
+ "specificity": 0.0001796616448296441,
183
+ "full_nll": 11.652510378095839,
184
+ "beta0_nll": 11.772504170735678,
185
+ "by_len": {
186
+ "1024": {
187
+ "gain": 0.009588295536660428,
188
+ "target_penalty": 0.0014437039693196614,
189
+ "specificity": 0.01219336191813151
190
+ },
191
+ "1536": {
192
+ "gain": 0.010254239325024093,
193
+ "target_penalty": -0.0020531018575032554,
194
+ "specificity": -0.011034965515136719
195
+ },
196
+ "2048": {
197
+ "gain": 0.01138062713764331,
198
+ "target_penalty": -0.0050856272379557295,
199
+ "specificity": -0.0006194114685058594
200
+ }
201
+ }
202
+ }
203
+ },
204
+ "writer_diagnostic": {
205
+ "EN_EN": {
206
+ "n": 6,
207
+ "source_advantage": -0.019238902100235048
208
+ },
209
+ "ZH_ZH": {
210
+ "n": 6,
211
+ "source_advantage": 0.01275022079200055
212
+ },
213
+ "ZH_EN": {
214
+ "n": 6,
215
+ "source_advantage": 0.023772776524905018
216
+ },
217
+ "EN_ZH": {
218
+ "n": 6,
219
+ "source_advantage": 0.0059973296267139264
220
+ }
221
+ },
222
+ "semantic_cross_language_diagnostic": {
223
+ "EN_ZH_SEM": {
224
+ "gain": -0.0031344768144604194,
225
+ "beta0_nll": 13.297510504722595,
226
+ "global_nll": 13.342041015625
227
+ },
228
+ "ZH_EN_SEM": {
229
+ "gain": 0.12379684830955245,
230
+ "beta0_nll": 9.510125041007996,
231
+ "global_nll": 8.339370965957642
232
+ }
233
+ },
234
+ "gates": {
235
+ "identity_ok": true,
236
+ "causality_ok": true,
237
+ "safety_ok": true,
238
+ "english_control_ok": false,
239
+ "zh_zh_ok": true,
240
+ "zh_en_ok": false,
241
+ "en_zh_ok": false,
242
+ "bilingual_transfer_ok": false,
243
+ "qualified": false
244
+ },
245
+ "VERDICT": "B2_REMOTE_RULER_INVALID_ENGLISH_CONTROL_FAILED"
246
+ }