LoheDelta A

as donto believed on

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

LoheDelta A has 24 facts recorded in Dontopedia across 2 references, with 1 live disagreement.

24 facts·22 predicates·2 sources·1 in dispute

Mostly:has parameter count(1), has role(1), has bpb at step(1)

Other facts (22)

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.

22 facts

Predicate	Value	Ref
Has Parameter Count	28000	[1]
Has Role	Small Model Baseline	[1]
Has Bpb at Step	1.89	[1]
Trained on	Fineweb	[1]
Rdf:type	Neural Network Architecture	[2]
Described As	good fit — moderate scope, concrete target	[2]
Total Parameter Count	28000	[2]
Bits Per Byte	1.89	[2]
Training Steps	20000	[2]
Throughput	71000	[2]
Variant Count	5	[2]
Variant Labels	A-E	[2]
Has Variant	Lohedelta F	[2]
Includes Eval Suite	true	[2]
Eval Suite Content	KV recall, copy, overwrite	[2]
Has Full Bptt Implementation	true	[2]
Implementation Language	Rust	[2]
Supports Hardware	CPU + Metal	[2]
Comparison Target Status	true	[2]
Architecture Relation	modern linear attention / state-space	[2]
Memory Mechanism	delta-rule memory	[2]
Estimated Port Effort	~3-5 days	[2]

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.

24 undated

—

labelblah/watt-activation/679

LoheDelta A

—

hasParameterCountblah/watt-activation/679

28000

—

hasRoleblah/watt-activation/679

ex:small-model-baseline

—

hasBpbAtStepblah/watt-activation/679

1.89

—

trainedOnblah/watt-activation/679

ex:fineweb

—

labelblah/watt-activation/693

LoheDelta-A

—

typeblah/watt-activation/693

ex:NeuralNetworkArchitecture

—

describedAsblah/watt-activation/693

good fit — moderate scope, concrete target

—

totalParameterCountblah/watt-activation/693

28000

—

bitsPerByteblah/watt-activation/693

1.89

—

trainingStepsblah/watt-activation/693

20000

—

throughputblah/watt-activation/693

71000

—

variantCountblah/watt-activation/693

—

variantLabelsblah/watt-activation/693

A-E

—

hasVariantblah/watt-activation/693

ex:lohedelta-f

—

includesEvalSuiteblah/watt-activation/693

true

—

evalSuiteContentblah/watt-activation/693

KV recall, copy, overwrite

—

hasFullBpttImplementationblah/watt-activation/693

true

—

implementationLanguageblah/watt-activation/693

ex:rust

—

supportsHardwareblah/watt-activation/693

CPU + Metal

—

comparisonTargetStatusblah/watt-activation/693

true

—

architectureRelationblah/watt-activation/693

modern linear attention / state-space

—

memoryMechanismblah/watt-activation/693

delta-rule memory

—

estimatedPortEffortblah/watt-activation/693

~3-5 days

References (2)

2 references

[1]6795 facts
ctx:discord/blah/watt-activation/679
- full textwatt-activation-679
  text/plain2 KBdoc:agent/watt-activation-679/e7b96ce5-bed6-4b84-baf1-a773f080c0db
  Show excerpt
  [2026-04-22 14:42] xenonfun: ``` ⏺ Pulling the per-step bpb trajectory from metrics.jsonl of the two most-comparable prior runs: step | unified_rotor_gpu_25m | lohe_delta_25m_replication | lohe_delta_16L_G16_H4_R16_doremi …
[2]69319 facts
ctx:discord/blah/watt-activation/693
- full textwatt-activation-693
  text/plain3 KBdoc:agent/watt-activation-693/8372b985-fcfe-4f68-a730-4cd9d14a0fb9
  Show excerpt
  [2026-04-28 23:22] xenonfun: KickModel (sparse extreme — top pick for fast iteration) 25 params total at g=8 (yes, twenty-five). BPB 2.04 on fineweb. Architecture: ∂ₜΨ = [Ω, Ψ] − γ⟨Ψ⟩ + kick, G independent S³ oscillators (3G+1 params), exp…

Keep researching

Missing something or suspicious of what's here? Kick off a research session — a Claude agent will investigate, cite its sources, and file new facts into a dedicated context you can review before accepting into the shared view.

LoheDelta A

Other facts (22)

Timeline

References (2)

See also

Keep researching