Dontopedia

128K run

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

128K run has 21 facts recorded in Dontopedia across 2 references.

21 facts·19 predicates·2 sources

Mostly:has fewer optimizer steps than(2), presupposes underperformance(1), warmup as fraction of epoch(1)

Maturity scale raw canonical shape-checked rule-derived certified

Inbound mentions (15)

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.

isRecommendedFixForIs Recommended Fix for(4)

hasSameTokenBudgetAsHas Same Token Budget As(2)

providesBetterGradientDiversityThanProvides Better Gradient Diversity Than(2)

causesPoorPerformanceInCauses Poor Performance in(1)

hasMoreLocalContextsPerStepThanHas More Local Contexts Per Step Than(1)

potentiallyInappropriateForPotentially Inappropriate for(1)

providesMoreIterativeRefinementThanProvides More Iterative Refinement Than(1)

reducesIterativeRefinementInReduces Iterative Refinement in(1)

statedConditionalActionStated Conditional Action(1)

unsuitableForUnsuitable for(1)

Other facts (20)

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.

20 facts
PredicateValueRef
Has Fewer Optimizer Steps Than2k Run[1]
Has Fewer Optimizer Steps Than2k Run[2]
Presupposes UnderperformanceRelative to 2k Run[1]
Warmup As Fraction of Epochabout 32%[1]
Uses Sequence Length131072[1]
Uses Batch Size1[1]
Has Steps Per Epoch1554[1]
Underperforms Compared to2k Run[1]
Involves One Giant Window Per UpdateTrue[1]
Spent Massive Fraction Under RampedTraining[1]
Steps Per Epoch1554[2]
Warmup Ratio Approximate0.32[2]
Trained Under Rampedtrue[2]
Context Length131072[2]
Pays Optimization Cost Without Proportionate Signaltrue[2]
Has Worse Batch Diversity Than2k Run[2]
Batch Size1[2]
Sequence Length131072[2]
Has One Giant Window Per Updatetrue[2]
Position Handling Difficultyhigh[2]

Timeline

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.

presupposesUnderperformanceblah/watt-activation/part-265
ex:relative-to-2k-run
warmupAsFractionOfEpochblah/watt-activation/part-265
about 32%
hasFewerOptimizerStepsThanblah/watt-activation/part-265
ex:2k-run
usesSequenceLengthblah/watt-activation/part-265
131072
usesBatchSizeblah/watt-activation/part-265
1
hasStepsPerEpochblah/watt-activation/part-265
1554
underperformsComparedToblah/watt-activation/part-265
ex:2k-run
involvesOneGiantWindowPerUpdateblah/watt-activation/part-265
ex:true
spentMassiveFractionUnderRampedblah/watt-activation/part-265
ex:training
labelblah/watt-activation/263
128K run
stepsPerEpochblah/watt-activation/263
1554
warmupRatioApproximateblah/watt-activation/263
0.32
trainedUnderRampedblah/watt-activation/263
true
hasFewerOptimizerStepsThanblah/watt-activation/263
ex:2k-run
contextLengthblah/watt-activation/263
131072
paysOptimizationCostWithoutProportionateSignalblah/watt-activation/263
true
hasWorseBatchDiversityThanblah/watt-activation/263
ex:2k-run
batchSizeblah/watt-activation/263
1
sequenceLengthblah/watt-activation/263
131072
hasOneGiantWindowPerUpdateblah/watt-activation/263
true
positionHandlingDifficultyblah/watt-activation/263
high

References (2)

2 references
  1. [1]Part 2659 facts
    ctx:discord/blah/watt-activation/part-265
  2. [2]26312 facts
    ctx:discord/blah/watt-activation/263
    • full textwatt-activation-263
      text/plain2 KBdoc:agent/watt-activation-263/18f755d3-7fe3-4c11-aca5-c67fdc4eb174
      Show excerpt
      [2026-03-13 04:59] xenonfun: ``` • Yes. Several plausible reasons, even without blaming the architecture. Most important ones: - Warmup/schedule mismatch - the 2K run had 12,432 steps/epoch - the 128K run had only 1,554

See also

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.