Two phases run

as donto believed on

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

Two phases run has 107 facts recorded in Dontopedia across 33 references, with 5 live disagreements.

107 facts·94 predicates·33 sources·5 in dispute

Mostly:rdf:type(5), at step(3), has loss(3)

Inbound mentions (20)

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.

`believesInImprovementTrajectory`Believes in Improvement Trajectory(1)

Xenonfunex:xenonfun

`claimsUnfairComparisonDueToMoreDataSeen`Claims Unfair Comparison Due to More Data Seen(1)

Xenonfunex:xenonfun

`diagnosesProperly`Diagnoses Properly(1)

Xenonfunex:xenonfun

`documentProgress`Document Progress(1)

Screenshotsex:screenshots

`evaluatedAsStrong`Evaluated As Strong(1)

Trajectoryex:trajectory

`evaluatedPerformanceAsBetter`Evaluated Performance As Better(1)

Xenonfunex:xenonfun

`evaluatesPositively`Evaluates Positively(1)

Foxhopex:foxhop

`evaluatesTrainingNegatively`Evaluates Training Negatively(1)

Xenonfunex:xenonfun

`fixedMetrics`Fixed Metrics(1)

Xenonfunex:xenonfun

`historicallyPrecedes`Historically Precedes(1)

Previous Runex:previous-run

`improving`Improving(1)

Bpbex:bpb

`isDone`Is Done(1)

Trainingex:training

`isFinishing`Is Finishing(1)

Runex:run

`isLowerInSomeRuns`Is Lower in Some Runs(1)

Per Iteration Lossex:per-iteration-loss

`isStrong`Is Strong(1)

Trajectoryex:trajectory

`observesPerIterationLowerLoss`Observes Per Iteration Lower Loss(1)

Xenonfunex:xenonfun

`ongoing`Ongoing(1)

Training Runex:training-run

`plansToKillProcess`Plans to Kill Process(1)

Xenonfunex:xenonfun

`worseThan`Worse Than(1)

Previous Runex:previous-run

`yieldsPerformanceJump`Yields Performance Jump(1)

Gate Matmulex:gate-matmul

Other facts (106)

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.

106 facts

Predicate	Value	Ref
Rdf:type	Training Run	[26]
Rdf:type	Training Run	[28]
Rdf:type	Experiment Run	[29]
Rdf:type	Run	[31]
Rdf:type	Process	[32]
At Step	650	[3]
At Step	11000	[4]
At Step	3400	[19]
Has Loss	7.079	[3]
Has Loss	2.85	[4]
Has Loss	0.8339	[20]
Measures Metric	DC 16	[31]
Measures Metric	DC 64	[31]
Measures Metric	DC 128	[31]
Progressing Well	true	[3]
Progressing Well	true	[19]
Has Checkpoint	Step 000500	[28]
Has Checkpoint	Step 001000	[28]
Total Iters	12K	[1]
Is Named	super_chat_20260225143238_sa7h	[2]
Contextualizes Specific Experiment	super_chat_20260225143238_sa7h	[2]
Loss Decreased From	8	[3]
Follows	Initial Loss 8 0	[3]
Has Architecture Diversity	1	[3]
Total Steps	50000	[3]
Descended From Loss	12.0 to 2.85	[4]
Improves Over	Old Run	[4]
Confirmed by	Break Through 2 65	[4]
Better Trajectory Than	Previous Run	[5]
Used Warmup	0	[5]
Used Batch Size	64	[5]
Reached in Epochs	16 auto-stopped	[5]
Used Accum	1	[5]
Is in Language	Python	[6]
Runs at It Per S	25	[7]
Achieves4x More Tokens Per Step Than Before	null	[7]
Has Effective Throughput Multiplier	~2.5x	[7]
Has Eta To50k Iterations	~32 min	[7]
Lacks Adaptive Reduction	null	[8]
Uses Purely Cosine Decay	null	[8]
Is Baseline or Llrd	null	[8]
Pairwise L256	Ongoing	[9]
Uses Tiny256 Seq Size	True	[10]
Improved Severity	Previous Run	[11]
Still Going Wrong Direction	Loss Trend	[11]
Shows Improvement	Slight	[11]
Precedes	Next Run	[12]
Lacks Global R	true	[12]
Iterates	100,500,1000	[13]
Switches to	Lohe Run	[13]
Pretrains With	Rotadamw	[13]
Improved Over	Earlier Stuff	[14]
Has Tags and Things	true	[14]
Still Going and Will Improve	null	[15]
Expected to Improve Text Performance	null	[15]
Uses Fixed Data	null	[15]
Has Phases	Two Phases	[16]
Takes Total Time	~15 min	[16]
Has Expected Base Lr During Warmup	0.0015	[17]
Exhibits Same Issue As	R Adaptive Lr Issue	[17]
Has Lr at Step50	0.000733	[17]
Has R Value	0.79	[17]
References Screenshots	true	[18]
Has No Rotation	true	[18]
Remaining Time	75 min	[19]
Nearing Completion	true	[19]
Matches Final Bpb of	10k Run	[19]
Bpb Is Dropping	true	[19]
Reached Step	3650	[20]
Has Tok Per S	10306	[20]
Has Bpb	1.203	[20]
Has Gnorm	0.2225	[20]
Has Lr	0.00195	[20]
Prioritized Style Over Semantics	Reverse	[21]
May Have Done Reverse Too Strongly	Reverse	[21]
Is Coasting Downhill on	Spent Lr Budget	[22]
Retains Entropy Better	Baseline Run	[23]
Moves Toward	Baseline Attractor	[23]
Exhibits	Early Warmup Dynamics	[23]
Has Higher	Scg Range Than Baseline	[23]
Has Elapsed	19 minutes	[24]
Is Still Running	null	[24]
Identifier	super_chat_20260225143238_sa7h	[25]
Has Loss at Step	2.85	[26]
Loss at Step Number	11000	[26]
Has Learning Rate Status	Live Lr	[26]
Steps Remaining	389000	[26]
Has Condition	Breaks Through 2.65	[26]
Warmup Value	0	[27]
Bs Value	64	[27]
Accum Value	1	[27]
Epoch Count	16	[27]
Ppl Value	83	[27]
Trajectory	strong	[27]
Includes Feature	Grouped V Fix	[29]
Has Throughput Ratio to Baseline	1.28	[29]
Has Fix	Fixed Data	[30]
Status	going	[30]
Expected Outcome	improve	[30]
Has Phase Count	2	[31]

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.

107 undated

—

totalItersblah/random/part-27

12K

—

isNamedblah/training-and-evals/part-24

super_chat_20260225143238_sa7h

—

contextualizesSpecificExperimentblah/training-and-evals/part-24

super_chat_20260225143238_sa7h

—

progressingWellblah/training-and-evals/part-29

true

—

atStepblah/training-and-evals/part-29

650

—

lossDecreasedFromblah/training-and-evals/part-29

—

followsblah/training-and-evals/part-29

ex:initial-loss-8-0

—

hasArchitectureDiversityblah/training-and-evals/part-29

—

totalStepsblah/training-and-evals/part-29

50000

—

hasLossblah/training-and-evals/part-29

7.079

—

descendedFromLossblah/unturf/part-70

12.0 to 2.85

—

improvesOverblah/unturf/part-70

ex:old-run

—

atStepblah/unturf/part-70

11000

—

hasLossblah/unturf/part-70

2.85

—

confirmedByblah/unturf/part-70

ex:break-through-2-65

—

betterTrajectoryThanblah/watt-activation/part-9

ex:previous-run

—

usedWarmupblah/watt-activation/part-9

—

usedBatchSizeblah/watt-activation/part-9

—

reachedInEpochsblah/watt-activation/part-9

16 auto-stopped

—

usedAccumblah/watt-activation/part-9

—

isInLanguageblah/watt-activation/part-10

ex:python

—

runsAtItPerSblah/watt-activation/part-19

—

achieves4xMoreTokensPerStepThanBeforeblah/watt-activation/part-19

null

—

hasEffectiveThroughputMultiplierblah/watt-activation/part-19

~2.5x

—

hasEtaTo50kIterationsblah/watt-activation/part-19

~32 min

—

lacksAdaptiveReductionblah/watt-activation/part-38

null

—

usesPurelyCosineDecayblah/watt-activation/part-38

null

—

isBaselineOrLlrdblah/watt-activation/part-38

null

—

pairwiseL256blah/watt-activation/part-57

ex:ongoing

—

usesTiny256SeqSizeblah/watt-activation/part-110

ex:true

—

improvedSeverityblah/watt-activation/part-191

ex:previous-run

—

stillGoingWrongDirectionblah/watt-activation/part-191

ex:loss-trend

—

showsImprovementblah/watt-activation/part-191

ex:slight

—

precedesblah/watt-activation/part-187

ex:next-run

—

lacksGlobalRblah/watt-activation/part-187

true

—

iteratesblah/watt-activation/part-205

100,500,1000

—

switchesToblah/watt-activation/part-205

ex:lohe-run

—

pretrainsWithblah/watt-activation/part-205

ex:rotadamw

—

improvedOverblah/watt-activation/part-230

ex:earlier-stuff

—

hasTagsAndThingsblah/watt-activation/part-230

true

—

stillGoingAndWillImproveblah/watt-activation/part-244

null

—

expectedToImproveTextPerformanceblah/watt-activation/part-244

null

—

usesFixedDatablah/watt-activation/part-244

null

—

hasPhasesblah/watt-activation/part-360

ex:two-phases

—

takesTotalTimeblah/watt-activation/part-360

~15 min

—

hasExpectedBaseLrDuringWarmupblah/watt-activation/part-393

0.0015

—

exhibitsSameIssueAsblah/watt-activation/part-393

ex:r-adaptive-lr-issue

—

hasLrAtStep50blah/watt-activation/part-393

0.000733

—

hasRValueblah/watt-activation/part-393

0.79

—

referencesScreenshotsblah/watt-activation/part-390

true

—

hasNoRotationblah/watt-activation/part-390

true

—

remainingTimeblah/watt-activation/part-413

75 min

—

nearingCompletionblah/watt-activation/part-413

true

—

matchesFinalBpbOfblah/watt-activation/part-413

ex:10k-run

—

atStepblah/watt-activation/part-413

3400

—

progressingWellblah/watt-activation/part-413

true

—

bpbIsDroppingblah/watt-activation/part-413

true

—

hasLossblah/watt-activation/part-659

0.8339

—

reachedStepblah/watt-activation/part-659

3650

—

hasTokPerSblah/watt-activation/part-659

10306

—

hasBpbblah/watt-activation/part-659

1.203

—

hasGnormblah/watt-activation/part-659

0.2225

—

hasLrblah/watt-activation/part-659

0.00195

—

prioritizedStyleOverSemanticsblah/watt-activation/part-703

ex:reverse

—

mayHaveDoneReverseTooStronglyblah/watt-activation/part-703

ex:reverse

—

isCoastingDownhillOnblah/random/part-25

ex:spent-lr-budget

—

retainsEntropyBetterblah/watt-activation/part-353

ex:baseline-run

—

movesTowardblah/watt-activation/part-353

ex:baseline-attractor

—

exhibitsblah/watt-activation/part-353

ex:early-warmup-dynamics

—

hasHigherblah/watt-activation/part-353

ex:scg-range-than-baseline

—

hasElapsedblah/watt-activation/part-475

19 minutes

—

isStillRunningblah/watt-activation/part-475

null

—

identifierblah/training-and-evals/24

super_chat_20260225143238_sa7h

—

typeblah/unturf/70

ex:TrainingRun

—

hasLossAtStepblah/unturf/70

2.85

—

lossAtStepNumberblah/unturf/70

11000

—

hasLearningRateStatusblah/unturf/70

ex:live-lr

—

stepsRemainingblah/unturf/70

389000

—

hasConditionblah/unturf/70

ex:breaks-through-2.65

—

warmupValueblah/watt-activation/9

—

bsValueblah/watt-activation/9

—

accumValueblah/watt-activation/9

—

epochCountblah/watt-activation/9

—

pplValueblah/watt-activation/9

—

trajectoryblah/watt-activation/9

strong

—

typeblah/watt-activation/175

ex:TrainingRun

—

hasCheckpointblah/watt-activation/175

ex:step-000500

—

hasCheckpointblah/watt-activation/175

ex:step-001000

—

typeblah/watt-activation/203

ex:ExperimentRun

—

includesFeatureblah/watt-activation/203

ex:grouped-v-fix

—

hasThroughputRatioToBaselineblah/watt-activation/203

1.28

—

hasFixblah/watt-activation/243

ex:fixed-data

—

statusblah/watt-activation/243

going

—

expectedOutcomeblah/watt-activation/243

improve

—

typeblah/watt-activation/358

ex:Run

—

labelblah/watt-activation/358

Two phases run

—

hasPhaseCountblah/watt-activation/358

—

measuresMetricblah/watt-activation/358

ex:dc-16

—

measuresMetricblah/watt-activation/358

ex:dc-64

—

measuresMetricblah/watt-activation/358

ex:dc-128

—

hasMeasurementFrequencyblah/watt-activation/358

1000

—

hasEstimatedDurationblah/watt-activation/358

~15 min

—

typeblah/watt-activation/656

ex:Process

—

hasProcessIdentifierblah/watt-activation/656

93094

—

writesCheckpointFormatblah/watt-activation/656

model.json

—

usesVersionblah/watt-activation/656

ex:old-binary

—

performanceRatingblah/watt-activation/692

not doing great

References (33)

33 references

[1]Part 271 fact
ctx:discord/blah/random/part-27
[2]Part 242 facts
ctx:discord/blah/training-and-evals/part-24
[3]Part 297 facts
ctx:discord/blah/training-and-evals/part-29
[4]Part 705 facts
ctx:discord/blah/unturf/part-70
[5]Part 95 facts
ctx:discord/blah/watt-activation/part-9
[6]Part 101 fact
ctx:discord/blah/watt-activation/part-10
[7]Part 194 facts
ctx:discord/blah/watt-activation/part-19
[8]Part 383 facts
ctx:discord/blah/watt-activation/part-38
[9]Part 571 fact
ctx:discord/blah/watt-activation/part-57
[10]Part 1101 fact
ctx:discord/blah/watt-activation/part-110
[11]Part 1913 facts
ctx:discord/blah/watt-activation/part-191
[12]Part 1872 facts
ctx:discord/blah/watt-activation/part-187
[13]Part 2053 facts
ctx:discord/blah/watt-activation/part-205
[14]Part 2302 facts
ctx:discord/blah/watt-activation/part-230
[15]Part 2443 facts
ctx:discord/blah/watt-activation/part-244
[16]Part 3602 facts
ctx:discord/blah/watt-activation/part-360
[17]Part 3934 facts
ctx:discord/blah/watt-activation/part-393
[18]Part 3902 facts
ctx:discord/blah/watt-activation/part-390
[19]Part 4136 facts
ctx:discord/blah/watt-activation/part-413
[20]Part 6596 facts
ctx:discord/blah/watt-activation/part-659
[21]Part 7032 facts
ctx:discord/blah/watt-activation/part-703
[22]Part 251 fact
ctx:discord/blah/random/part-25
[23]Part 3534 facts
ctx:discord/blah/watt-activation/part-353
[24]Part 4752 facts
ctx:discord/blah/watt-activation/part-475
[25]241 fact
ctx:discord/blah/training-and-evals/24
- full texttraining-and-evals-24
  text/plain2 KBdoc:agent/training-and-evals-24/b280a4e7-48a7-4bc2-9593-e4261e806744
  Show excerpt
  [2026-02-25 14:46] omega [bot]: **Selective Pressure — Three Mechanisms** **1. Evaluation Pressure** (per-candidate) Each candidate gets exactly 30 steps to prove itself. Records bestLoss, bestValLoss, fitnessScore. At 30 steps during warm…
[26]706 facts
ctx:discord/blah/unturf/70
- full textunturf-70
  text/plain3 KBdoc:agent/unturf-70/069aca86-d010-45b0-961c-c5d6a8358036
  Show excerpt
  [2026-03-12 21:29] foxhop.: ● The spikes are fine. Here's what's happening: - 3 spikes out of 46 points (~6.5%) — all exactly ~18-20 loss, all at firehose round boundaries - This is one bad batch when the CUDA engine finishes a round a…
[27]96 facts
ctx:discord/blah/watt-activation/9
- full textwatt-activation-9
  text/plain3 KBdoc:agent/watt-activation-9/b3eaeadb-4d30-449a-926d-ac2cf3ed101d
  Show excerpt
  [2026-02-27 01:34] xenonfun: (files: Screenshot_2026-02-26_at_8.34.13_PM.png) [2026-02-27 02:10] xenonfun: ``` +**Conclusion**: The phase-gated no-FFN architecture is the correct design. The mixer should be the sole sequence+channel mixing…
[28]1753 facts
ctx:discord/blah/watt-activation/175
- full textwatt-activation-175
  text/plain3 KBdoc:agent/watt-activation-175/d5d93509-931b-40d2-96b6-c9c6844cefac
  Show excerpt
  [2026-03-09 22:46] xenonfun: ``` === What is the capital of the United States? === The first country ──────────────────────────────────────────────────────────── 9 prompt tok (64ms), 3 generated tok 21.4 tok/s gen=140ms === What year di…
[29]2033 facts
ctx:discord/blah/watt-activation/203
- full textwatt-activation-203
  text/plain2 KBdoc:agent/watt-activation-203/1245ede9-7e37-4bbe-a212-202bb86ee81d
  Show excerpt
  [2026-03-10 07:17] xenonfun: ## Summary ``` Config Params Val PPL sys_r global_r attn_r Tok/s Wall(m) ----------------------------------------------------------------------------------------------- spec…
[30]2433 facts
ctx:discord/blah/watt-activation/243
- full textwatt-activation-243
  text/plain3 KBdoc:agent/watt-activation-243/14f8ddd1-c20c-4aa1-99ee-73dc849eba12
  Show excerpt
  [2026-03-12 05:04] xenonfun: ⏺ While we wait for the image data to re-prep, let me summarize the issues found and fixed: Problems found: 1. tok/s inflated — was averaging all modality step times but computing tokens as bs*seq which onl…
[31]3588 facts
ctx:discord/blah/watt-activation/358
- full textwatt-activation-358
  text/plain2 KBdoc:agent/watt-activation-358/5e6372ad-404d-41bf-87ea-6453f5db1dc5
  Show excerpt
  [2026-03-17 17:23] xenonfun: This is a decisive result for the antenna branch. Updated conclusion: - For the antenna architecture in the tested regime, RotAdamW / strict sphere ∩ zero-mean manifold constraint is actively harmful. - Adam wi…
[32]6564 facts
ctx:discord/blah/watt-activation/656
- full textwatt-activation-656
  text/plain2 KBdoc:agent/watt-activation-656/5433fed2-51af-40c7-a493-d7a7bde6f077
  Show excerpt
  [2026-04-19 09:55] xenonfun: ``` ⏺ Massive quality jump over LoheDelta at same step. Real English: ▎ The scientific method requires. She says he is a lot of float. He was a harmble. She flies helpless and turned the store and put his hug…
[33]6921 fact
ctx:discord/blah/watt-activation/692
- full textwatt-activation-692
  text/plain1 KBdoc:agent/watt-activation-692/b0e0cd01-937c-4f55-bc38-a0f551fae13f
  Show excerpt
  [2026-04-28 14:51] xenonfun: Step 1 BPB = 4.54 — vs 8.18 without unigram init. 3.6-bit head start banked. Run going. Step 100: BPB 4.52. Already below where the constant-LR run was at step 100 (4.54), and far below where v1-cosine was at s…

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.

Two phases run

Inbound mentions (20)

believesInImprovementTrajectoryBelieves in Improvement Trajectory(1)

claimsUnfairComparisonDueToMoreDataSeenClaims Unfair Comparison Due to More Data Seen(1)

diagnosesProperlyDiagnoses Properly(1)

documentProgressDocument Progress(1)

evaluatedAsStrongEvaluated As Strong(1)

evaluatedPerformanceAsBetterEvaluated Performance As Better(1)

evaluatesPositivelyEvaluates Positively(1)

evaluatesTrainingNegativelyEvaluates Training Negatively(1)

fixedMetricsFixed Metrics(1)

historicallyPrecedesHistorically Precedes(1)

improvingImproving(1)

isDoneIs Done(1)

isFinishingIs Finishing(1)

isLowerInSomeRunsIs Lower in Some Runs(1)

isStrongIs Strong(1)

observesPerIterationLowerLossObserves Per Iteration Lower Loss(1)

ongoingOngoing(1)

plansToKillProcessPlans to Kill Process(1)

worseThanWorse Than(1)

yieldsPerformanceJumpYields Performance Jump(1)

Other facts (106)

Timeline

References (33)

See also

Keep researching

`believesInImprovementTrajectory`Believes in Improvement Trajectory(1)

`claimsUnfairComparisonDueToMoreDataSeen`Claims Unfair Comparison Due to More Data Seen(1)

`diagnosesProperly`Diagnoses Properly(1)

`documentProgress`Document Progress(1)

`evaluatedAsStrong`Evaluated As Strong(1)

`evaluatedPerformanceAsBetter`Evaluated Performance As Better(1)

`evaluatesPositively`Evaluates Positively(1)

`evaluatesTrainingNegatively`Evaluates Training Negatively(1)

`fixedMetrics`Fixed Metrics(1)

`historicallyPrecedes`Historically Precedes(1)

`improving`Improving(1)

`isDone`Is Done(1)

`isFinishing`Is Finishing(1)

`isLowerInSomeRuns`Is Lower in Some Runs(1)

`isStrong`Is Strong(1)

`observesPerIterationLowerLoss`Observes Per Iteration Lower Loss(1)

`ongoing`Ongoing(1)

`plansToKillProcess`Plans to Kill Process(1)

`worseThan`Worse Than(1)

`yieldsPerformanceJump`Yields Performance Jump(1)