lohe_spherical

as donto believed on

From Dontopedia, the open, paraconsistent wiki. (Last updated 2026-06-06.)

lohe_spherical has 102 facts recorded in Dontopedia across 22 references, with 5 live disagreements.

102 facts·78 predicates·22 sources·5 in dispute

Mostly:has ppl at step(7), rdf:type(7), has perplexity(2)

Inbound mentions (30)

Other subjects in dontopedia point AT this entity as a value. These are inverse relationships — e.g. "X motherOf this subject" — and answer questions the forward facts can't. Grouped by predicate.

`usesAttention`Uses Attention(2)

Harmonic Gptex:harmonic-gpt
Lohe Decoder D512 L6ex:lohe-decoder-d512-l6

`adaptsToAttentionType`Adapts to Attention Type(1)

Lohe Optimizerex:lohe-optimizer

`advocatesSpherical`Advocates Spherical(1)

Xenonfunex:xenonfun

`assumesSharedKnowledge`Assumes Shared Knowledge(1)

Chat Contextex:chat-context

`behind`Behind(1)

Current Bpbex:current-bpb

`claimsComponentHasLowerPerplexity`Claims Component Has Lower Perplexity(1)

Message 0418ex:message-0418

`claimsComponentIsFaster`Claims Component Is Faster(1)

Message 0418ex:message-0418

`claimsTradeoffIsOffset`Claims Tradeoff Is Offset(1)

Message 0418ex:message-0418

`comparedTo`Compared to(1)

Topo Anchor Kan Quality Advantageex:topo-anchor-kan-quality-advantage

`comparesUnfavorablyTo`Compares Unfavorably to(1)

Model Training Processex:model-training-process

`convergesIdenticallyTo`Converges Identically to(1)

Spectralex:spectral

`dependsOn`Depends on(1)

Phase Metrics Sidecarex:phase-metrics-sidecar

`describes`Describes(1)

Lohe Spherical Speed Advantageex:lohe-spherical-speed-advantage

`evaluatedSuperior`Evaluated Superior(1)

Complex Attentionex:complex-attention

`has4_3xBetterPPLThan`Has4 3x Better Ppl Than(1)

Topo Anchor Kanex:topo-anchor-kan

`hasAttentionType`Has Attention Type(1)

Lohe Spherical Plus Jacobiex:lohe-spherical-plus-jacobi

`hasLowerLossThan`Has Lower Loss Than(1)

Topo Star Plus Kanex:topo-star-plus-kan

`hasSameModelQualityAs`Has Same Model Quality As(1)

Spectralex:spectral

`hasSyncHubRoleForConfig`Has Sync Hub Role for Config(1)

Block 10ex:block-10

`involvesComponent`Involves Component(1)

D512 Configurationex:d512-configuration

`mentionsModel`Mentions Model(1)

Message 2026 03 12 01 34ex:message-2026-03-12-01-34

`needsDesignAfterResults`Needs Design After Results(1)

Feedback Mechanismex:feedback-mechanism

`refersToSpecificAttention`Refers to Specific Attention(1)

Loheex:lohe

`subject`Subject(1)

Lohe Spherical Strictly Betterex:lohe-spherical-strictly-better

`subject2`Subject2(1)

Spectral Lohe Identical Convergenceex:spectral-lohe-identical-convergence

`suggests`Suggests(1)

Recommendation Fastex:recommendation-fast

`usedWith`Used With(1)

D 512ex:d-512

`usesConfig`Uses Config(1)

Run 2ex:run-2

`usesMethod`Uses Method(1)

Experiment Lohe Spherical Lohe V2ex:experiment-lohe-spherical-lohe-v2

Other facts (95)

The long tail: predicates that appear too rarely to warrant their own section. Filter or scroll to find a specific one. Each row links to its source.

95 facts

Predicate	Value	Ref
Has Ppl at Step	1567	[13]
Has Ppl at Step	517	[13]
Has Ppl at Step	449	[13]
Has Ppl at Step	335	[13]
Has Ppl at Step	216	[13]
Has Ppl at Step	190	[13]
Has Ppl at Step	210	[13]
Rdf:type	Experiment	[14]
Rdf:type	Model Component	[15]
Rdf:type	Attention Configuration	[17]
Rdf:type	Model	[18]
Rdf:type	Model Architecture	[19]
Rdf:type	Attention Type	[20]
Rdf:type	Component	[21]
Has Perplexity	97	[6]
Has Perplexity	8225	[20]
Combined With	Lohe V3	[7]
Combined With	Lohe V3	[10]
Has Param Count	93.5M	[9]
Has Param Count	93500000	[20]
Has Tokens Per Second	16043	[9]
Has Tokens Per Second	16043	[20]
Has Loss	9.01	[9]
Has Loss	9.01	[20]
Results Pending	true	[1]
Is Attention Variant	Lohe V2	[2]
Reduces Qk Params Vs	Spectralattn	[2]
Has Simpler Attention Mechanism Than	Spectralattn	[2]
Has Fewer Params in Qk	32x768	[2]
Has Gh	32	[3]
Has Dmodel	768	[3]
Keeps V at	full d_model=768 across GH=32 spectral channels	[3]
Has D	768	[3]
Optimizes Qk But Not V	true	[3]
Has Kv Inp Shape	(B, T, GH=32, D=768)	[3]
Achieves Only	~60%	[4]
Has Weaker Sync	Spectral	[4]
Underperforms	Spectral	[4]
Has DC Mode	0.148	[4]
Has Sync Hub	Block 10	[4]
Has Tok S	14000	[4]
Has R Hub	0.385	[4]
Uses Ring Topology	null	[5]
Gets Partially There	null	[5]
Outperforms Others	Competitors	[6]
Superior in	Perplexity Metric	[6]
Wins by Margin	Big Margin	[6]
Is Model Architecture	Training Setup	[8]
Uses Fewest Parameters	null	[9]
Might Catch Up With More Steps	null	[9]
Has Ppl Per Token	0.51	[9]
Is Strictly Better Than	Spectral	[9]
Is Speed King	true	[9]
References Lohe Method	null	[9]
Is25 Percent Faster Than	Spectral	[9]
Is2 25x Faster Than	Topo Anchor Kan	[9]
Has Ppl	8225	[9]
Excels in Inference Speed	null	[9]
Recommended for	fast iteration/long training	[9]
Has13 Percent Fewer Params Than	Spectral	[9]
Ontologically Efficient	D 512	[10]
Causes Fewer Params	Attention Projections	[10]
Parameter Efficient at	512	[10]
Is Prior Model	true	[11]
Achieves Bpb	2.52	[11]
Was Faster	Previous Config	[12]
Has Less Ppl Per Step	Previous Config	[12]
Wins in Ppl	null	[13]
Benefits More From	Strong Rotational Dynamics	[13]
Has27 Percent Fewer Params Than	Spectral	[13]
Has Fewest Params	null	[13]
Has Final Ppl	210	[13]
Has Lower Ppl Than Spectral At1000	null	[13]
Has Params	12.3M	[13]
Has Tok Per S	83777	[13]
Is3 2 Times Faster Than	Star Kan	[13]
Is47 Percent Faster Than	Spectral	[13]
Is Fastest	null	[13]
Maintains Lead Throughout	null	[13]
Overtakes	Spectral	[13]
Overtakes at Step	200	[13]
Preferred by Author	null	[13]
Superior in All Metrics	null	[13]
Uses Lohe Sync	null	[13]
Uses O T Cumsum Recurrence	null	[13]
Wins Everything	null	[13]
Status	pending	[14]
Reaches State Partially	DC Kuramoto State	[16]
Has Sync Strength Magnitude	0.6	[17]
Has Baseline Throughput	14000	[17]
Has Throughput Unit	Tok S	[17]
Has Ppl Per Token	0.51	[20]
Property	parameter-efficient	[21]
Component Name	lohe_spherical+lohe_v3	[21]
Has Bpb	2.52	[22]

Timeline axis is valid_time — when each source says the fact was true in the world, not when Dontopedia learned about it. Retracted rows are kept for provenance; coloured stripes indicate the context kind.

102 undated

—

resultsPendingblah/watt-activation/part-206

true

—

isAttentionVariantblah/watt-activation/part-204

ex:lohe-v2

—

reducesQKParamsVsblah/watt-activation/part-204

ex:spectralattn

—

hasSimplerAttentionMechanismThanblah/watt-activation/part-204

ex:spectralattn

—

hasFewerParamsInQKblah/watt-activation/part-204

32x768

—

hasGhblah/watt-activation/part-202

—

hasDmodelblah/watt-activation/part-202

768

—

keepsVAtblah/watt-activation/part-202

full d_model=768 across GH=32 spectral channels

—

hasDblah/watt-activation/part-202

768

—

optimizesQKButNotVblah/watt-activation/part-202

true

—

hasKvInpShapeblah/watt-activation/part-202

(B, T, GH=32, D=768)

—

achievesOnlyblah/watt-activation/part-221

~60%

—

hasWeakerSyncblah/watt-activation/part-221

ex:spectral

—

underperformsblah/watt-activation/part-221

ex:spectral

—

hasDcModeblah/watt-activation/part-221

0.148

—

hasSyncHubblah/watt-activation/part-221

ex:block-10

—

hasTokSblah/watt-activation/part-221

14000

—

hasRHubblah/watt-activation/part-221

0.385

—

usesRingTopologyblah/watt-activation/part-219

null

—

getsPartiallyThereblah/watt-activation/part-219

null

—

outperformsOthersblah/watt-activation/part-240

ex:competitors

—

hasPerplexityblah/watt-activation/part-240

—

superiorInblah/watt-activation/part-240

ex:perplexity-metric

—

winsByMarginblah/watt-activation/part-240

ex:big-margin

—

combinedWithblah/watt-activation/part-266

ex:lohe-v3

—

isModelArchitectureblah/watt-activation/part-262

ex:training-setup

—

usesFewestParametersblah/watt-activation/part-317

null

—

hasParamCountblah/watt-activation/part-317

93.5M

—

mightCatchUpWithMoreStepsblah/watt-activation/part-317

null

—

hasPPLPerTokenblah/watt-activation/part-317

0.51

—

isStrictlyBetterThanblah/watt-activation/part-317

ex:spectral

—

isSpeedKingblah/watt-activation/part-317

true

—

referencesLoheMethodblah/watt-activation/part-317

null

—

hasTokensPerSecondblah/watt-activation/part-317

16043

—

is25PercentFasterThanblah/watt-activation/part-317

ex:spectral

—

is2_25xFasterThanblah/watt-activation/part-317

ex:topo-anchor-kan

—

hasLossblah/watt-activation/part-317

9.01

—

hasPPLblah/watt-activation/part-317

8225

—

excelsInInferenceSpeedblah/watt-activation/part-317

null

—

recommendedForblah/watt-activation/part-317

fast iteration/long training

—

has13PercentFewerParamsThanblah/watt-activation/part-317

ex:spectral

—

combinedWithblah/watt-activation/part-322

ex:lohe-v3

—

ontologicallyEfficientblah/watt-activation/part-322

ex:d-512

—

causesFewerParamsblah/watt-activation/part-322

ex:attention-projections

—

parameterEfficientAtblah/watt-activation/part-322

512

—

isPriorModelblah/watt-activation/part-338

true

—

achievesBpbblah/watt-activation/part-338

2.52

—

wasFasterblah/watt-activation/part-215

ex:previous-config

—

hasLessPplPerStepblah/watt-activation/part-215

ex:previous-config

—

winsInPplblah/watt-activation/part-319

null

—

benefitsMoreFromblah/watt-activation/part-319

ex:strong-rotational-dynamics

—

has27PercentFewerParamsThanblah/watt-activation/part-319

ex:spectral

—

hasFewestParamsblah/watt-activation/part-319

null

—

hasFinalPplblah/watt-activation/part-319

210

—

hasLowerPplThanSpectralAt1000blah/watt-activation/part-319

null

—

hasParamsblah/watt-activation/part-319

12.3M

—