Stage2R-A4A-R2 native IAP oracle 20260905T033649Z — NATIVE_IAP_SEGMENT_HEADROOM_CONFIRMED_MOVE_TO_A4A_R3_READER
Browse files- experiments/stage2r/a4a_r2/20260905T033649Z/README_STAGE2R_A4A_R2.md +28 -0
- experiments/stage2r/a4a_r2/20260905T033649Z/hf_write_probe.txt +1 -0
- experiments/stage2r/a4a_r2/20260905T033649Z/native_iap_calibration_summary.csv +9 -0
- experiments/stage2r/a4a_r2/20260905T033649Z/native_iap_confirm.csv +9 -0
- experiments/stage2r/a4a_r2/20260905T033649Z/native_iap_oracle_beta_sweep.csv +73 -0
- experiments/stage2r/a4a_r2/20260905T033649Z/native_iap_scale_diagnostic.csv +65 -0
- experiments/stage2r/a4a_r2/20260905T033649Z/stage2r_a4a_r2_summary.json +292 -0
experiments/stage2r/a4a_r2/20260905T033649Z/README_STAGE2R_A4A_R2.md
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# MANN-GPT Stage2R-A4A-R2 — Native-IAP Remote Value Geometry Oracle
|
| 2 |
+
|
| 3 |
+
Run: 20260905T033649Z
|
| 4 |
+
|
| 5 |
+
Parent:
|
| 6 |
+
- checkpoint/stage2r/a2d1/20260903T102303Z/winner/model.safetensors
|
| 7 |
+
- SHA256: 96e786420db004ea541bdf58270eba3e62b6264cee5d742c10ae3fc8cb217092
|
| 8 |
+
|
| 9 |
+
A4A-R1 automatic verdict:
|
| 10 |
+
SEGMENT_COMPRESSION_VALUE_PATH_NOT_QUALIFIED
|
| 11 |
+
|
| 12 |
+
R2 correction:
|
| 13 |
+
- no ConnectionNetwork.v_proj remote values
|
| 14 |
+
- remote values live in the parent's native IAP activation space
|
| 15 |
+
- remote propagation is combined with local propagated field before the frozen parent IAP gate/transform/norm/output projection
|
| 16 |
+
- beta-zero exact parent identity required
|
| 17 |
+
- Layer23 and Layer24 compared
|
| 18 |
+
- VALUE_TOKEN / LINE_MEAN / SEGMENT_UNIFORM / SEGMENT_SOURCE oracle granularity compared
|
| 19 |
+
- signed beta diagnostic
|
| 20 |
+
- disjoint DEV and sealed OOD confirmation
|
| 21 |
+
- no reader training
|
| 22 |
+
- no model promotion
|
| 23 |
+
|
| 24 |
+
Winner:
|
| 25 |
+
{'layer': 24, 'variant': 'SEGMENT_SOURCE', 'beta': 0.4, 'dev_baseline_nll': 10.985894471406937, 'dev_nll': 10.365149319171906, 'dev_improvement': 0.05650383351584606, 'ood_baseline_nll': 9.837896992762884, 'ood_nll': 9.326098799705505, 'ood_improvement': 0.052023129885774996, 'remote_512plus_improvement': 0.053399075920911865}
|
| 26 |
+
|
| 27 |
+
Verdict:
|
| 28 |
+
NATIVE_IAP_SEGMENT_HEADROOM_CONFIRMED_MOVE_TO_A4A_R3_READER
|
experiments/stage2r/a4a_r2/20260905T033649Z/hf_write_probe.txt
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
stage2r-a4a-r2 write probe 20260905T033649Z
|
experiments/stage2r/a4a_r2/20260905T033649Z/native_iap_calibration_summary.csv
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
layer,variant,baseline_nll,best_positive_beta,best_positive_nll,best_positive_improvement,best_signed_beta,best_signed_nll,best_signed_improvement
|
| 2 |
+
23,LINE_MEAN,10.978641659021378,0.4,10.980940222740173,-0.00020936685886882908,-0.2,10.922842025756836,0.005082562579013581
|
| 3 |
+
23,SEGMENT_SOURCE,10.978641659021378,0.2,10.995024532079697,-0.001492249548454562,-0.2,10.905917823314667,0.006624119628401582
|
| 4 |
+
23,SEGMENT_UNIFORM,10.978641659021378,0.2,10.991773277521133,-0.0011961059398423242,-0.2,10.91215893626213,0.006055641929492852
|
| 5 |
+
23,VALUE_TOKEN,10.978641659021378,0.4,10.940850377082825,0.0034422547991170634,-0.2,10.928884595632553,0.0045321693643164
|
| 6 |
+
24,LINE_MEAN,10.978641659021378,0.4,10.18743970990181,0.07206738080110492,0.4,10.18743970990181,0.07206738080110492
|
| 7 |
+
24,SEGMENT_SOURCE,10.978641659021378,0.4,10.394158840179443,0.053238172534910325,0.4,10.394158840179443,0.053238172534910325
|
| 8 |
+
24,SEGMENT_UNIFORM,10.978641659021378,0.4,10.472400963306427,0.046111414457084686,0.4,10.472400963306427,0.046111414457084686
|
| 9 |
+
24,VALUE_TOKEN,10.978641659021378,0.4,5.281682297587395,0.5189129528380838,0.4,5.281682297587395,0.5189129528380838
|
experiments/stage2r/a4a_r2/20260905T033649Z/native_iap_confirm.csv
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
layer,variant,beta,dev_baseline_nll,dev_nll,dev_improvement,ood_baseline_nll,ood_nll,ood_improvement,remote_512plus_improvement
|
| 2 |
+
23,LINE_MEAN,0.4,10.985894471406937,10.994243770837784,-0.0007600017870714076,9.837896992762884,9.830050299564997,0.0007975986334945006,-0.00038288624687230643
|
| 3 |
+
23,SEGMENT_SOURCE,0.2,10.985894471406937,10.984524816274643,0.0001246739749647616,9.837896992762884,9.825668136278788,0.001243035629778644,0.00013565003319037587
|
| 4 |
+
23,SEGMENT_UNIFORM,0.2,10.985894471406937,10.98234212398529,0.0003233553199415703,9.837896992762884,9.830602049827576,0.0007415144660159421,0.00029142319137812696
|
| 5 |
+
23,VALUE_TOKEN,0.4,10.985894471406937,11.00829620162646,-0.00203913575520206,9.837896992762884,9.715553959210714,0.01243589291920517,-0.0012006826324355328
|
| 6 |
+
24,LINE_MEAN,0.4,10.985894471406937,10.120602319637934,0.07876392350400822,9.837896992762884,9.263883978128433,0.05834712592098858,0.07703442173003387
|
| 7 |
+
24,SEGMENT_SOURCE,0.4,10.985894471406937,10.365149319171906,0.05650383351584606,9.837896992762884,9.326098799705505,0.052023129885774996,0.053399075920911865
|
| 8 |
+
24,SEGMENT_UNIFORM,0.4,10.985894471406937,10.455435504515966,0.04828545989328402,9.837896992762884,9.454481571912766,0.0389733111794292,0.04541453365658499
|
| 9 |
+
24,VALUE_TOKEN,0.4,10.985894471406937,5.357221667965253,0.5123545304473358,9.837896992762884,4.1935978308320045,0.5737302561800588,0.5001135406958085
|
experiments/stage2r/a4a_r2/20260905T033649Z/native_iap_oracle_beta_sweep.csv
ADDED
|
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
layer,variant,beta,nll,improvement,nll_320_512,nll_512_768,nll_768_plus
|
| 2 |
+
23,VALUE_TOKEN,-0.4,11.017887085676193,-0.0035747069513437386,11.039673233032227,10.3667653799057,11.374395098005023
|
| 3 |
+
23,VALUE_TOKEN,-0.2,10.928884595632553,0.0045321693643164,10.945668411254882,10.153899788856506,11.359744616917201
|
| 4 |
+
23,VALUE_TOKEN,-0.1,10.93619480729103,0.0038663117941796763,10.978764915466309,10.071024537086487,11.400170598711286
|
| 5 |
+
23,VALUE_TOKEN,-0.05,10.937726855278015,0.0037267637485682834,11.00197467803955,10.049156188964844,11.39959021977016
|
| 6 |
+
23,VALUE_TOKEN,0.0,10.978641659021378,0.0,11.041075325012207,10.010846257209778,11.487072127205986
|
| 7 |
+
23,VALUE_TOKEN,0.05,10.990325182676315,-0.0010642048459006902,11.028413200378418,9.999081015586853,11.529544694083077
|
| 8 |
+
23,VALUE_TOKEN,0.1,10.985579878091812,-0.0006319742720387719,11.022087860107423,9.976730465888977,11.535988126482282
|
| 9 |
+
23,VALUE_TOKEN,0.2,10.97530597448349,0.0003038339934473198,11.005486297607423,9.93709945678711,11.547009468078613
|
| 10 |
+
23,VALUE_TOKEN,0.4,10.940850377082825,0.0034422547991170634,10.970702171325684,9.837347030639648,11.550101007734026
|
| 11 |
+
23,LINE_MEAN,-0.4,10.965295135974884,0.001215680724539034,11.112001609802245,10.010301113128662,11.40621566772461
|
| 12 |
+
23,LINE_MEAN,-0.2,10.922842025756836,0.005082562579013581,10.961572456359864,10.024611711502075,11.408451897757393
|
| 13 |
+
23,LINE_MEAN,-0.1,10.925179302692413,0.004869669490035054,10.960775184631348,10.028554439544678,11.412110737391881
|
| 14 |
+
23,LINE_MEAN,-0.05,10.952687084674835,0.0023640970488562167,11.04982681274414,10.020433902740479,11.416017668587822
|
| 15 |
+
23,LINE_MEAN,0.0,10.978641659021378,0.0,11.041075325012207,10.010846257209778,11.487072127205986
|
| 16 |
+
23,LINE_MEAN,0.05,10.998839408159256,-0.00183973115847911,11.03923740386963,10.01203191280365,11.533873694283622
|
| 17 |
+
23,LINE_MEAN,0.1,10.996576339006424,-0.0016335973558540449,11.037078475952148,10.018680930137634,11.526443617684501
|
| 18 |
+
23,LINE_MEAN,0.2,10.982673943042755,-0.00036728441883920605,11.0263916015625,9.960100412368774,11.535774775913783
|
| 19 |
+
23,LINE_MEAN,0.4,10.980940222740173,-0.00020936685886882908,11.01388339996338,9.971659898757935,11.534140995570592
|
| 20 |
+
23,SEGMENT_UNIFORM,-0.4,10.91734105348587,0.005583623861621827,11.022799491882324,9.980653524398804,11.377263614109584
|
| 21 |
+
23,SEGMENT_UNIFORM,-0.2,10.91215893626213,0.006055641929492852,10.940474891662598,10.02183997631073,11.400686945234026
|
| 22 |
+
23,SEGMENT_UNIFORM,-0.1,10.923412889242172,0.00503056493640292,10.978022193908691,10.017869591712952,11.401859555925641
|
| 23 |
+
23,SEGMENT_UNIFORM,-0.05,10.946186274290085,0.002956229535429227,11.035396575927734,10.021774649620056,11.410699844360352
|
| 24 |
+
23,SEGMENT_UNIFORM,0.0,10.978641659021378,0.0,11.041075325012207,10.010846257209778,11.487072127205986
|
| 25 |
+
23,SEGMENT_UNIFORM,0.05,10.996900260448456,-0.001663102047972827,11.034091567993164,10.021667718887329,11.527610778808594
|
| 26 |
+
23,SEGMENT_UNIFORM,0.1,10.997924566268921,-0.00175640191623325,11.033050537109375,10.019168138504028,11.532123974391393
|
| 27 |
+
23,SEGMENT_UNIFORM,0.2,10.991773277521133,-0.0011961059398423242,11.042951965332032,9.964562058448792,11.54219491141183
|
| 28 |
+
23,SEGMENT_UNIFORM,0.4,11.00258731842041,-0.0021811131233485472,11.042685890197754,9.965011596679688,11.566845893859863
|
| 29 |
+
23,SEGMENT_SOURCE,-0.4,10.931465595960617,0.004297076498711928,11.0803466796875,9.977930188179016,11.36999934060233
|
| 30 |
+
23,SEGMENT_SOURCE,-0.2,10.905917823314667,0.006624119628401582,10.939155960083008,10.017264366149902,11.389978272574288
|
| 31 |
+
23,SEGMENT_SOURCE,-0.1,10.919816881418228,0.005358110723544007,10.970736503601074,10.01013958454132,11.403261320931572
|
| 32 |
+
23,SEGMENT_SOURCE,-0.05,10.946822881698608,0.0028982435451495965,11.041253471374512,10.031398296356201,11.402472223554339
|
| 33 |
+
23,SEGMENT_SOURCE,0.0,10.978641659021378,0.0,11.041075325012207,10.010846257209778,11.487072127205986
|
| 34 |
+
23,SEGMENT_SOURCE,0.05,11.001796185970306,-0.0021090520729312883,11.034967613220214,10.026754379272461,11.535269056047712
|
| 35 |
+
23,SEGMENT_SOURCE,0.1,11.000587046146393,-0.001998916423962365,11.037054061889648,10.018771648406982,11.53557654789516
|
| 36 |
+
23,SEGMENT_SOURCE,0.2,10.995024532079697,-0.001492249548454562,11.044454956054688,9.966683983802795,11.547340256827217
|
| 37 |
+
23,SEGMENT_SOURCE,0.4,11.009960353374481,-0.0028526930130166345,11.039267539978027,9.979060411453247,11.57811232975551
|
| 38 |
+
24,VALUE_TOKEN,-0.4,18.53719472885132,-0.6884779833959623,17.640575408935547,18.75506353378296,19.05314064025879
|
| 39 |
+
24,VALUE_TOKEN,-0.2,14.711318612098694,-0.33999442453885886,14.539565658569336,13.9931480884552,15.244382449558803
|
| 40 |
+
24,VALUE_TOKEN,-0.1,12.798395812511444,-0.1657540349715975,12.782742881774903,11.881459474563599,13.333540099007744
|
| 41 |
+
24,VALUE_TOKEN,-0.05,11.874789208173752,-0.08162645042850007,11.896850395202637,10.91076672077179,12.409901210239955
|
| 42 |
+
24,VALUE_TOKEN,0.0,10.978641659021378,0.0,11.041075325012207,10.010846257209778,11.487072127205986
|
| 43 |
+
24,VALUE_TOKEN,0.05,10.114535361528397,0.07870794259715426,10.186910820007324,9.176892399787903,10.598634583609444
|
| 44 |
+
24,VALUE_TOKEN,0.1,9.269077628850937,0.1557172629608196,9.328698539733887,8.395200848579407,9.725849424089704
|
| 45 |
+
24,VALUE_TOKEN,0.2,7.7276283502578735,0.29612163414515663,7.775427341461182,6.983428359031677,8.118743351527623
|
| 46 |
+
24,VALUE_TOKEN,0.4,5.281682297587395,0.5189129528380838,5.218171501159668,4.823667585849762,5.588769844600132
|
| 47 |
+
24,LINE_MEAN,-0.4,12.542697668075562,-0.1424635266940302,12.58638858795166,11.316420078277588,13.21221991947719
|
| 48 |
+
24,LINE_MEAN,-0.2,11.59233444929123,-0.05589879051800253,11.723330688476562,10.486396074295044,12.130730492728096
|
| 49 |
+
24,LINE_MEAN,-0.1,11.258146107196808,-0.02545892805835007,11.353463935852051,10.232623815536499,11.77607468196324
|
| 50 |
+
24,LINE_MEAN,-0.05,11.110961318016052,-0.012052461780273598,11.1885196685791,10.112710237503052,11.625991685049874
|
| 51 |
+
24,LINE_MEAN,0.0,10.978641659021378,0.0,11.041075325012207,10.010846257209778,11.487072127205986
|
| 52 |
+
24,LINE_MEAN,0.05,10.857821196317673,0.011005046567343251,10.897086906433106,9.93898332118988,11.354824474879674
|
| 53 |
+
24,LINE_MEAN,0.1,10.744109272956848,0.021362605078908765,10.761430168151856,9.85317873954773,11.240840366908483
|
| 54 |
+
24,LINE_MEAN,0.2,10.536054134368896,0.04031350493062115,10.500845146179199,9.739683389663696,11.016272408621651
|
| 55 |
+
24,LINE_MEAN,0.4,10.18743970990181,0.07206738080110492,10.0389404296875,9.591671347618103,10.633949688502721
|
| 56 |
+
24,SEGMENT_UNIFORM,-0.4,12.163267642259598,-0.10790278251451886,12.038189315795899,11.136754393577576,12.839188303266253
|
| 57 |
+
24,SEGMENT_UNIFORM,-0.2,11.420922040939331,-0.04028552854300719,11.455481147766113,10.432356119155884,11.961131777082171
|
| 58 |
+
24,SEGMENT_UNIFORM,-0.1,11.17023777961731,-0.017451714569670238,11.22244930267334,10.185556411743164,11.695618901933942
|
| 59 |
+
24,SEGMENT_UNIFORM,-0.05,11.067880034446716,-0.00812836215963108,11.125558662414551,10.095938682556152,11.582076072692871
|
| 60 |
+
24,SEGMENT_UNIFORM,0.0,10.978641659021378,0.0,11.041075325012207,10.010846257209778,11.487072127205986
|
| 61 |
+
24,SEGMENT_UNIFORM,0.05,10.898026078939438,0.007342946658223078,10.961811447143555,9.944170117378235,11.397525651114327
|
| 62 |
+
24,SEGMENT_UNIFORM,0.1,10.826695412397385,0.013840168150413721,10.8861665725708,9.890886425971985,11.318964004516602
|
| 63 |
+
24,SEGMENT_UNIFORM,0.2,10.68470025062561,0.02677393228826532,10.72810344696045,9.788888931274414,11.165590150015694
|
| 64 |
+
24,SEGMENT_UNIFORM,0.4,10.472400963306427,0.046111414457084686,10.501699638366699,9.660747051239014,10.915275573730469
|
| 65 |
+
24,SEGMENT_SOURCE,-0.4,12.304673194885254,-0.12078284154344802,12.103841018676757,11.404544115066528,12.962484223502022
|
| 66 |
+
24,SEGMENT_SOURCE,-0.2,11.470941096544266,-0.04484156171709599,11.480557250976563,10.52170741558075,12.006491661071777
|
| 67 |
+
24,SEGMENT_SOURCE,-0.1,11.199419409036636,-0.02010975099399853,11.230317306518554,10.25226891040802,11.718578338623047
|
| 68 |
+
24,SEGMENT_SOURCE,-0.05,11.082075119018555,-0.009421334916436016,11.132221984863282,10.120735883712769,11.595592635018486
|
| 69 |
+
24,SEGMENT_SOURCE,0.0,10.978641659021378,0.0,11.041075325012207,10.010846257209778,11.487072127205986
|
| 70 |
+
24,SEGMENT_SOURCE,0.05,10.886018246412277,0.008436691485689376,10.955853462219238,9.914545178413391,11.39126341683524
|
| 71 |
+
24,SEGMENT_SOURCE,0.1,10.801152646541595,0.016166755231867478,10.871356391906739,9.845130681991577,11.297305379595075
|
| 72 |
+
24,SEGMENT_SOURCE,0.2,10.645188927650452,0.030372858658422545,10.719643783569335,9.712151765823364,11.125170980181013
|
| 73 |
+
24,SEGMENT_SOURCE,0.4,10.394158840179443,0.053238172534910325,10.456561088562012,9.4965078830719,10.86252920968192
|
experiments/stage2r/a4a_r2/20260905T033649Z/native_iap_scale_diagnostic.csv
ADDED
|
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
layer,variant,local_prop_rms,remote_prop_rms,ratio_remote_to_local,query_gate
|
| 2 |
+
23,VALUE_TOKEN,0.357602059841156,0.43167003989219666,1.2071239948272705,0.4965539574623108
|
| 3 |
+
23,LINE_MEAN,0.357602059841156,0.4010293483734131,1.1214402914047241,0.4965539574623108
|
| 4 |
+
23,SEGMENT_UNIFORM,0.357602059841156,0.39370954036712646,1.1009711027145386,0.4965539574623108
|
| 5 |
+
23,SEGMENT_SOURCE,0.357602059841156,0.39371258020401,1.1009795665740967,0.4965539574623108
|
| 6 |
+
24,VALUE_TOKEN,0.3278133273124695,0.5201942920684814,1.5868613719940186,0.5638696551322937
|
| 7 |
+
24,LINE_MEAN,0.3278133273124695,0.32002949714660645,0.9762552976608276,0.5638696551322937
|
| 8 |
+
24,SEGMENT_UNIFORM,0.3278133273124695,0.3005026578903198,0.9166883230209351,0.5638696551322937
|
| 9 |
+
24,SEGMENT_SOURCE,0.3278133273124695,0.3004978597164154,0.9166737198829651,0.5638696551322937
|
| 10 |
+
23,VALUE_TOKEN,0.3047577440738678,0.44503065943717957,1.4602768421173096,0.49629509449005127
|
| 11 |
+
23,LINE_MEAN,0.3047577440738678,0.40272703766822815,1.321466088294983,0.49629509449005127
|
| 12 |
+
23,SEGMENT_UNIFORM,0.3047577440738678,0.39738455414772034,1.3039358854293823,0.49629509449005127
|
| 13 |
+
23,SEGMENT_SOURCE,0.3047577440738678,0.3975256383419037,1.304398775100708,0.49629509449005127
|
| 14 |
+
24,VALUE_TOKEN,0.14888304471969604,0.5678533315658569,3.81408953666687,0.5858930945396423
|
| 15 |
+
24,LINE_MEAN,0.14888304471969604,0.369051069021225,2.4787981510162354,0.5858930945396423
|
| 16 |
+
24,SEGMENT_UNIFORM,0.14888304471969604,0.344941645860672,2.3168630599975586,0.5858930945396423
|
| 17 |
+
24,SEGMENT_SOURCE,0.14888304471969604,0.34838011860847473,2.3399581909179688,0.5858930945396423
|
| 18 |
+
23,VALUE_TOKEN,0.35742616653442383,0.4457889497280121,1.2472196817398071,0.4957565367221832
|
| 19 |
+
23,LINE_MEAN,0.35742616653442383,0.40614598989486694,1.1363073587417603,0.4957565367221832
|
| 20 |
+
23,SEGMENT_UNIFORM,0.35742616653442383,0.4029514193534851,1.1273696422576904,0.4957565367221832
|
| 21 |
+
23,SEGMENT_SOURCE,0.35742616653442383,0.4029514193534851,1.1273696422576904,0.4957565367221832
|
| 22 |
+
24,VALUE_TOKEN,0.21974028646945953,0.5766969323158264,2.624447822570801,0.5793724060058594
|
| 23 |
+
24,LINE_MEAN,0.21974028646945953,0.3649982213973999,1.661043643951416,0.5793724060058594
|
| 24 |
+
24,SEGMENT_UNIFORM,0.21974028646945953,0.33140453696250916,1.5081645250320435,0.5793724060058594
|
| 25 |
+
24,SEGMENT_SOURCE,0.21974028646945953,0.33140453696250916,1.5081645250320435,0.5793724060058594
|
| 26 |
+
23,VALUE_TOKEN,0.36201202869415283,0.4344441294670105,1.2000820636749268,0.4972192049026489
|
| 27 |
+
23,LINE_MEAN,0.36201202869415283,0.4080333709716797,1.1271265745162964,0.4972192049026489
|
| 28 |
+
23,SEGMENT_UNIFORM,0.36201202869415283,0.39829200506210327,1.100217580795288,0.4972192049026489
|
| 29 |
+
23,SEGMENT_SOURCE,0.36201202869415283,0.40039390325546265,1.1060237884521484,0.4972192049026489
|
| 30 |
+
24,VALUE_TOKEN,0.22535328567028046,0.6066068410873413,2.6918036937713623,0.5893503427505493
|
| 31 |
+
24,LINE_MEAN,0.22535328567028046,0.3820682466030121,1.6954189538955688,0.5893503427505493
|
| 32 |
+
24,SEGMENT_UNIFORM,0.22535328567028046,0.33360910415649414,1.4803825616836548,0.5893503427505493
|
| 33 |
+
24,SEGMENT_SOURCE,0.22535328567028046,0.35120800137519836,1.5584772825241089,0.5893503427505493
|
| 34 |
+
23,VALUE_TOKEN,0.3565719723701477,0.4288981258869171,1.2028374671936035,0.496520459651947
|
| 35 |
+
23,LINE_MEAN,0.3565719723701477,0.4024380147457123,1.128630518913269,0.496520459651947
|
| 36 |
+
23,SEGMENT_UNIFORM,0.3565719723701477,0.39933526515960693,1.1199289560317993,0.496520459651947
|
| 37 |
+
23,SEGMENT_SOURCE,0.3565719723701477,0.3996120095252991,1.120705008506775,0.496520459651947
|
| 38 |
+
24,VALUE_TOKEN,0.1608145385980606,0.5614151358604431,3.4910717010498047,0.5904121994972229
|
| 39 |
+
24,LINE_MEAN,0.1608145385980606,0.3505589962005615,2.1798958778381348,0.5904121994972229
|
| 40 |
+
24,SEGMENT_UNIFORM,0.1608145385980606,0.3400644361972809,2.1146371364593506,0.5904121994972229
|
| 41 |
+
24,SEGMENT_SOURCE,0.1608145385980606,0.34285810589790344,2.132009267807007,0.5904121994972229
|
| 42 |
+
23,VALUE_TOKEN,0.3565036356449127,0.4050653278827667,1.1362165212631226,0.4957903325557709
|
| 43 |
+
23,LINE_MEAN,0.3565036356449127,0.39326611161231995,1.1031194925308228,0.4957903325557709
|
| 44 |
+
23,SEGMENT_UNIFORM,0.3565036356449127,0.39478257298469543,1.1073732376098633,0.4957903325557709
|
| 45 |
+
23,SEGMENT_SOURCE,0.3565036356449127,0.39479002356529236,1.1073940992355347,0.4957903325557709
|
| 46 |
+
24,VALUE_TOKEN,0.19219240546226501,0.6083715558052063,3.1654295921325684,0.5877295732498169
|
| 47 |
+
24,LINE_MEAN,0.19219240546226501,0.32465657591819763,1.689226746559143,0.5877295732498169
|
| 48 |
+
24,SEGMENT_UNIFORM,0.19219240546226501,0.33176109194755554,1.7261923551559448,0.5877295732498169
|
| 49 |
+
24,SEGMENT_SOURCE,0.19219240546226501,0.3322206437587738,1.7285834550857544,0.5877295732498169
|
| 50 |
+
23,VALUE_TOKEN,0.356195867061615,0.43436071276664734,1.219443440437317,0.4956546425819397
|
| 51 |
+
23,LINE_MEAN,0.356195867061615,0.4046727120876312,1.1360960006713867,0.4956546425819397
|
| 52 |
+
23,SEGMENT_UNIFORM,0.356195867061615,0.4021194577217102,1.1289279460906982,0.4956546425819397
|
| 53 |
+
23,SEGMENT_SOURCE,0.356195867061615,0.4021187722682953,1.1289260387420654,0.4956546425819397
|
| 54 |
+
24,VALUE_TOKEN,0.2597600221633911,0.5757870078086853,2.216611385345459,0.5853672027587891
|
| 55 |
+
24,LINE_MEAN,0.2597600221633911,0.326217383146286,1.2558413743972778,0.5853672027587891
|
| 56 |
+
24,SEGMENT_UNIFORM,0.2597600221633911,0.3230356276035309,1.2435925006866455,0.5853672027587891
|
| 57 |
+
24,SEGMENT_SOURCE,0.2597600221633911,0.3230355978012085,1.243592381477356,0.5853672027587891
|
| 58 |
+
23,VALUE_TOKEN,0.30488714575767517,0.4380365312099457,1.436716914176941,0.49714216589927673
|
| 59 |
+
23,LINE_MEAN,0.30488714575767517,0.40896400809288025,1.3413618803024292,0.49714216589927673
|
| 60 |
+
23,SEGMENT_UNIFORM,0.30488714575767517,0.4026072025299072,1.3205121755599976,0.49714216589927673
|
| 61 |
+
23,SEGMENT_SOURCE,0.30488714575767517,0.4025217294692993,1.3202319145202637,0.49714216589927673
|
| 62 |
+
24,VALUE_TOKEN,0.16106350719928741,0.5976373553276062,3.710569381713867,0.599200427532196
|
| 63 |
+
24,LINE_MEAN,0.16106350719928741,0.41173073649406433,2.5563251972198486,0.599200427532196
|
| 64 |
+
24,SEGMENT_UNIFORM,0.16106350719928741,0.387154757976532,2.4037394523620605,0.599200427532196
|
| 65 |
+
24,SEGMENT_SOURCE,0.16106350719928741,0.39184531569480896,2.4328620433807373,0.599200427532196
|
experiments/stage2r/a4a_r2/20260905T033649Z/stage2r_a4a_r2_summary.json
ADDED
|
@@ -0,0 +1,292 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_id": "20260905T033649Z",
|
| 3 |
+
"parent": {
|
| 4 |
+
"path": "checkpoint/stage2r/a2d1/20260903T102303Z/winner/model.safetensors",
|
| 5 |
+
"sha256": "96e786420db004ea541bdf58270eba3e62b6264cee5d742c10ae3fc8cb217092"
|
| 6 |
+
},
|
| 7 |
+
"a4a_r1_reference": {
|
| 8 |
+
"run_id": "20260904T165341Z",
|
| 9 |
+
"automatic_verdict": "SEGMENT_COMPRESSION_VALUE_PATH_NOT_QUALIFIED",
|
| 10 |
+
"r2_reassessment": [
|
| 11 |
+
"R1 used ConnectionNetwork.v_proj as value although parent IAP propagates activation space",
|
| 12 |
+
"R1 query-only GQA value concatenation transposed head/head_dim before reshape"
|
| 13 |
+
]
|
| 14 |
+
},
|
| 15 |
+
"identity": {
|
| 16 |
+
"L23_VALUE_TOKEN": 0.0,
|
| 17 |
+
"L23_LINE_MEAN": 0.0,
|
| 18 |
+
"L23_SEGMENT_UNIFORM": 0.0,
|
| 19 |
+
"L23_SEGMENT_SOURCE": 0.0,
|
| 20 |
+
"L24_VALUE_TOKEN": 0.0,
|
| 21 |
+
"L24_LINE_MEAN": 0.0,
|
| 22 |
+
"L24_SEGMENT_UNIFORM": 0.0,
|
| 23 |
+
"L24_SEGMENT_SOURCE": 0.0
|
| 24 |
+
},
|
| 25 |
+
"scale_diagnostic_mean": [
|
| 26 |
+
{
|
| 27 |
+
"layer": 23,
|
| 28 |
+
"variant": "LINE_MEAN",
|
| 29 |
+
"local_prop_rms": 0.3444945774972439,
|
| 30 |
+
"remote_prop_rms": 0.40340957418084145,
|
| 31 |
+
"ratio_remote_to_local": 1.176943525671959,
|
| 32 |
+
"query_gate": 0.49636654928326607
|
| 33 |
+
},
|
| 34 |
+
{
|
| 35 |
+
"layer": 23,
|
| 36 |
+
"variant": "SEGMENT_SOURCE",
|
| 37 |
+
"local_prop_rms": 0.3444945774972439,
|
| 38 |
+
"remote_prop_rms": 0.39920325949788094,
|
| 39 |
+
"ratio_remote_to_local": 1.1645036041736603,
|
| 40 |
+
"query_gate": 0.49636654928326607
|
| 41 |
+
},
|
| 42 |
+
{
|
| 43 |
+
"layer": 23,
|
| 44 |
+
"variant": "SEGMENT_UNIFORM",
|
| 45 |
+
"local_prop_rms": 0.3444945774972439,
|
| 46 |
+
"remote_prop_rms": 0.3988977521657944,
|
| 47 |
+
"ratio_remote_to_local": 1.1636545658111572,
|
| 48 |
+
"query_gate": 0.49636654928326607
|
| 49 |
+
},
|
| 50 |
+
{
|
| 51 |
+
"layer": 23,
|
| 52 |
+
"variant": "VALUE_TOKEN",
|
| 53 |
+
"local_prop_rms": 0.3444945774972439,
|
| 54 |
+
"remote_prop_rms": 0.43291180953383446,
|
| 55 |
+
"ratio_remote_to_local": 1.2637396156787872,
|
| 56 |
+
"query_gate": 0.49636654928326607
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"layer": 24,
|
| 60 |
+
"variant": "LINE_MEAN",
|
| 61 |
+
"local_prop_rms": 0.2119525521993637,
|
| 62 |
+
"remote_prop_rms": 0.3561638407409191,
|
| 63 |
+
"ratio_remote_to_local": 1.8116006553173065,
|
| 64 |
+
"query_gate": 0.5851493626832962
|
| 65 |
+
},
|
| 66 |
+
{
|
| 67 |
+
"layer": 24,
|
| 68 |
+
"variant": "SEGMENT_SOURCE",
|
| 69 |
+
"local_prop_rms": 0.2119525521993637,
|
| 70 |
+
"remote_prop_rms": 0.34018127247691154,
|
| 71 |
+
"ratio_remote_to_local": 1.7325401082634926,
|
| 72 |
+
"query_gate": 0.5851493626832962
|
| 73 |
+
},
|
| 74 |
+
{
|
| 75 |
+
"layer": 24,
|
| 76 |
+
"variant": "SEGMENT_UNIFORM",
|
| 77 |
+
"local_prop_rms": 0.2119525521993637,
|
| 78 |
+
"remote_prop_rms": 0.3365592323243618,
|
| 79 |
+
"ratio_remote_to_local": 1.7137824892997742,
|
| 80 |
+
"query_gate": 0.5851493626832962
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"layer": 24,
|
| 84 |
+
"variant": "VALUE_TOKEN",
|
| 85 |
+
"local_prop_rms": 0.2119525521993637,
|
| 86 |
+
"remote_prop_rms": 0.5768203064799309,
|
| 87 |
+
"ratio_remote_to_local": 2.912610560655594,
|
| 88 |
+
"query_gate": 0.5851493626832962
|
| 89 |
+
}
|
| 90 |
+
],
|
| 91 |
+
"calibration": [
|
| 92 |
+
{
|
| 93 |
+
"layer": 23,
|
| 94 |
+
"variant": "LINE_MEAN",
|
| 95 |
+
"baseline_nll": 10.978641659021378,
|
| 96 |
+
"best_positive_beta": 0.4,
|
| 97 |
+
"best_positive_nll": 10.980940222740173,
|
| 98 |
+
"best_positive_improvement": -0.00020936685886882908,
|
| 99 |
+
"best_signed_beta": -0.2,
|
| 100 |
+
"best_signed_nll": 10.922842025756836,
|
| 101 |
+
"best_signed_improvement": 0.005082562579013581
|
| 102 |
+
},
|
| 103 |
+
{
|
| 104 |
+
"layer": 23,
|
| 105 |
+
"variant": "SEGMENT_SOURCE",
|
| 106 |
+
"baseline_nll": 10.978641659021378,
|
| 107 |
+
"best_positive_beta": 0.2,
|
| 108 |
+
"best_positive_nll": 10.995024532079697,
|
| 109 |
+
"best_positive_improvement": -0.001492249548454562,
|
| 110 |
+
"best_signed_beta": -0.2,
|
| 111 |
+
"best_signed_nll": 10.905917823314667,
|
| 112 |
+
"best_signed_improvement": 0.006624119628401582
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"layer": 23,
|
| 116 |
+
"variant": "SEGMENT_UNIFORM",
|
| 117 |
+
"baseline_nll": 10.978641659021378,
|
| 118 |
+
"best_positive_beta": 0.2,
|
| 119 |
+
"best_positive_nll": 10.991773277521133,
|
| 120 |
+
"best_positive_improvement": -0.0011961059398423242,
|
| 121 |
+
"best_signed_beta": -0.2,
|
| 122 |
+
"best_signed_nll": 10.91215893626213,
|
| 123 |
+
"best_signed_improvement": 0.006055641929492852
|
| 124 |
+
},
|
| 125 |
+
{
|
| 126 |
+
"layer": 23,
|
| 127 |
+
"variant": "VALUE_TOKEN",
|
| 128 |
+
"baseline_nll": 10.978641659021378,
|
| 129 |
+
"best_positive_beta": 0.4,
|
| 130 |
+
"best_positive_nll": 10.940850377082825,
|
| 131 |
+
"best_positive_improvement": 0.0034422547991170634,
|
| 132 |
+
"best_signed_beta": -0.2,
|
| 133 |
+
"best_signed_nll": 10.928884595632553,
|
| 134 |
+
"best_signed_improvement": 0.0045321693643164
|
| 135 |
+
},
|
| 136 |
+
{
|
| 137 |
+
"layer": 24,
|
| 138 |
+
"variant": "LINE_MEAN",
|
| 139 |
+
"baseline_nll": 10.978641659021378,
|
| 140 |
+
"best_positive_beta": 0.4,
|
| 141 |
+
"best_positive_nll": 10.18743970990181,
|
| 142 |
+
"best_positive_improvement": 0.07206738080110492,
|
| 143 |
+
"best_signed_beta": 0.4,
|
| 144 |
+
"best_signed_nll": 10.18743970990181,
|
| 145 |
+
"best_signed_improvement": 0.07206738080110492
|
| 146 |
+
},
|
| 147 |
+
{
|
| 148 |
+
"layer": 24,
|
| 149 |
+
"variant": "SEGMENT_SOURCE",
|
| 150 |
+
"baseline_nll": 10.978641659021378,
|
| 151 |
+
"best_positive_beta": 0.4,
|
| 152 |
+
"best_positive_nll": 10.394158840179443,
|
| 153 |
+
"best_positive_improvement": 0.053238172534910325,
|
| 154 |
+
"best_signed_beta": 0.4,
|
| 155 |
+
"best_signed_nll": 10.394158840179443,
|
| 156 |
+
"best_signed_improvement": 0.053238172534910325
|
| 157 |
+
},
|
| 158 |
+
{
|
| 159 |
+
"layer": 24,
|
| 160 |
+
"variant": "SEGMENT_UNIFORM",
|
| 161 |
+
"baseline_nll": 10.978641659021378,
|
| 162 |
+
"best_positive_beta": 0.4,
|
| 163 |
+
"best_positive_nll": 10.472400963306427,
|
| 164 |
+
"best_positive_improvement": 0.046111414457084686,
|
| 165 |
+
"best_signed_beta": 0.4,
|
| 166 |
+
"best_signed_nll": 10.472400963306427,
|
| 167 |
+
"best_signed_improvement": 0.046111414457084686
|
| 168 |
+
},
|
| 169 |
+
{
|
| 170 |
+
"layer": 24,
|
| 171 |
+
"variant": "VALUE_TOKEN",
|
| 172 |
+
"baseline_nll": 10.978641659021378,
|
| 173 |
+
"best_positive_beta": 0.4,
|
| 174 |
+
"best_positive_nll": 5.281682297587395,
|
| 175 |
+
"best_positive_improvement": 0.5189129528380838,
|
| 176 |
+
"best_signed_beta": 0.4,
|
| 177 |
+
"best_signed_nll": 5.281682297587395,
|
| 178 |
+
"best_signed_improvement": 0.5189129528380838
|
| 179 |
+
}
|
| 180 |
+
],
|
| 181 |
+
"confirmation": [
|
| 182 |
+
{
|
| 183 |
+
"layer": 23,
|
| 184 |
+
"variant": "LINE_MEAN",
|
| 185 |
+
"beta": 0.4,
|
| 186 |
+
"dev_baseline_nll": 10.985894471406937,
|
| 187 |
+
"dev_nll": 10.994243770837784,
|
| 188 |
+
"dev_improvement": -0.0007600017870714076,
|
| 189 |
+
"ood_baseline_nll": 9.837896992762884,
|
| 190 |
+
"ood_nll": 9.830050299564997,
|
| 191 |
+
"ood_improvement": 0.0007975986334945006,
|
| 192 |
+
"remote_512plus_improvement": -0.00038288624687230643
|
| 193 |
+
},
|
| 194 |
+
{
|
| 195 |
+
"layer": 23,
|
| 196 |
+
"variant": "SEGMENT_SOURCE",
|
| 197 |
+
"beta": 0.2,
|
| 198 |
+
"dev_baseline_nll": 10.985894471406937,
|
| 199 |
+
"dev_nll": 10.984524816274643,
|
| 200 |
+
"dev_improvement": 0.0001246739749647616,
|
| 201 |
+
"ood_baseline_nll": 9.837896992762884,
|
| 202 |
+
"ood_nll": 9.825668136278788,
|
| 203 |
+
"ood_improvement": 0.001243035629778644,
|
| 204 |
+
"remote_512plus_improvement": 0.00013565003319037587
|
| 205 |
+
},
|
| 206 |
+
{
|
| 207 |
+
"layer": 23,
|
| 208 |
+
"variant": "SEGMENT_UNIFORM",
|
| 209 |
+
"beta": 0.2,
|
| 210 |
+
"dev_baseline_nll": 10.985894471406937,
|
| 211 |
+
"dev_nll": 10.98234212398529,
|
| 212 |
+
"dev_improvement": 0.0003233553199415703,
|
| 213 |
+
"ood_baseline_nll": 9.837896992762884,
|
| 214 |
+
"ood_nll": 9.830602049827576,
|
| 215 |
+
"ood_improvement": 0.0007415144660159421,
|
| 216 |
+
"remote_512plus_improvement": 0.00029142319137812696
|
| 217 |
+
},
|
| 218 |
+
{
|
| 219 |
+
"layer": 23,
|
| 220 |
+
"variant": "VALUE_TOKEN",
|
| 221 |
+
"beta": 0.4,
|
| 222 |
+
"dev_baseline_nll": 10.985894471406937,
|
| 223 |
+
"dev_nll": 11.00829620162646,
|
| 224 |
+
"dev_improvement": -0.00203913575520206,
|
| 225 |
+
"ood_baseline_nll": 9.837896992762884,
|
| 226 |
+
"ood_nll": 9.715553959210714,
|
| 227 |
+
"ood_improvement": 0.01243589291920517,
|
| 228 |
+
"remote_512plus_improvement": -0.0012006826324355328
|
| 229 |
+
},
|
| 230 |
+
{
|
| 231 |
+
"layer": 24,
|
| 232 |
+
"variant": "LINE_MEAN",
|
| 233 |
+
"beta": 0.4,
|
| 234 |
+
"dev_baseline_nll": 10.985894471406937,
|
| 235 |
+
"dev_nll": 10.120602319637934,
|
| 236 |
+
"dev_improvement": 0.07876392350400822,
|
| 237 |
+
"ood_baseline_nll": 9.837896992762884,
|
| 238 |
+
"ood_nll": 9.263883978128433,
|
| 239 |
+
"ood_improvement": 0.05834712592098858,
|
| 240 |
+
"remote_512plus_improvement": 0.07703442173003387
|
| 241 |
+
},
|
| 242 |
+
{
|
| 243 |
+
"layer": 24,
|
| 244 |
+
"variant": "SEGMENT_SOURCE",
|
| 245 |
+
"beta": 0.4,
|
| 246 |
+
"dev_baseline_nll": 10.985894471406937,
|
| 247 |
+
"dev_nll": 10.365149319171906,
|
| 248 |
+
"dev_improvement": 0.05650383351584606,
|
| 249 |
+
"ood_baseline_nll": 9.837896992762884,
|
| 250 |
+
"ood_nll": 9.326098799705505,
|
| 251 |
+
"ood_improvement": 0.052023129885774996,
|
| 252 |
+
"remote_512plus_improvement": 0.053399075920911865
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"layer": 24,
|
| 256 |
+
"variant": "SEGMENT_UNIFORM",
|
| 257 |
+
"beta": 0.4,
|
| 258 |
+
"dev_baseline_nll": 10.985894471406937,
|
| 259 |
+
"dev_nll": 10.455435504515966,
|
| 260 |
+
"dev_improvement": 0.04828545989328402,
|
| 261 |
+
"ood_baseline_nll": 9.837896992762884,
|
| 262 |
+
"ood_nll": 9.454481571912766,
|
| 263 |
+
"ood_improvement": 0.0389733111794292,
|
| 264 |
+
"remote_512plus_improvement": 0.04541453365658499
|
| 265 |
+
},
|
| 266 |
+
{
|
| 267 |
+
"layer": 24,
|
| 268 |
+
"variant": "VALUE_TOKEN",
|
| 269 |
+
"beta": 0.4,
|
| 270 |
+
"dev_baseline_nll": 10.985894471406937,
|
| 271 |
+
"dev_nll": 5.357221667965253,
|
| 272 |
+
"dev_improvement": 0.5123545304473358,
|
| 273 |
+
"ood_baseline_nll": 9.837896992762884,
|
| 274 |
+
"ood_nll": 4.1935978308320045,
|
| 275 |
+
"ood_improvement": 0.5737302561800588,
|
| 276 |
+
"remote_512plus_improvement": 0.5001135406958085
|
| 277 |
+
}
|
| 278 |
+
],
|
| 279 |
+
"winner": {
|
| 280 |
+
"layer": 24,
|
| 281 |
+
"variant": "SEGMENT_SOURCE",
|
| 282 |
+
"beta": 0.4,
|
| 283 |
+
"dev_baseline_nll": 10.985894471406937,
|
| 284 |
+
"dev_nll": 10.365149319171906,
|
| 285 |
+
"dev_improvement": 0.05650383351584606,
|
| 286 |
+
"ood_baseline_nll": 9.837896992762884,
|
| 287 |
+
"ood_nll": 9.326098799705505,
|
| 288 |
+
"ood_improvement": 0.052023129885774996,
|
| 289 |
+
"remote_512plus_improvement": 0.053399075920911865
|
| 290 |
+
},
|
| 291 |
+
"VERDICT": "NATIVE_IAP_SEGMENT_HEADROOM_CONFIRMED_MOVE_TO_A4A_R3_READER"
|
| 292 |
+
}
|