Dontopedia

Training Run Config

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

Training Run Config has 21 facts recorded in Dontopedia across 3 references, with 3 live disagreements.

21 facts·18 predicates·3 sources·3 in dispute

Mostly:batch size(2), sequence length(2), learning rate(2)

Maturity scale raw canonical shape-checked rule-derived certified

Inbound mentions (2)

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.

requestedRunRequested Run(1)

usesConfigurationUses Configuration(1)

Other facts (21)

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.

21 facts
PredicateValueRef
Batch Size4[1]
Batch Size64[3]
Sequence Length2048[1]
Sequence Length8000[3]
Learning Rate0.0001[1]
Learning Rate0.000008[2]
Warmup Steps500[1]
Save Frequency2000[1]
Validation Frequency2000[1]
Log Frequency100[1]
Checkpoint Directorycheckpoints/epoch_cl100k[1]
Rdf:typeTraining Configuration[2]
Step Count6000[2]
Throughput Rate12000[2]
Calculated Tokens Seen82000000[2]
Data TypeInstruction Data[2]
Learning Rate Decay Schedulecosine decay[2]
Initial Learning Rate0.00002[2]
Current Epoch2[2]
Iterations100[3]
Applies toBoth Configs[3]

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.

batchSizeblah/watt-activation/139
4
sequenceLengthblah/watt-activation/139
2048
learningRateblah/watt-activation/139
0.0001
warmupStepsblah/watt-activation/139
500
saveFrequencyblah/watt-activation/139
2000
validationFrequencyblah/watt-activation/139
2000
logFrequencyblah/watt-activation/139
100
checkpointDirectoryblah/watt-activation/139
checkpoints/epoch_cl100k
typeblah/watt-activation/169
ex:TrainingConfiguration
stepCountblah/watt-activation/169
6000
throughputRateblah/watt-activation/169
12000
calculatedTokensSeenblah/watt-activation/169
82000000
dataTypeblah/watt-activation/169
ex:instruction-data
learningRateblah/watt-activation/169
0.000008
learningRateDecayScheduleblah/watt-activation/169
cosine decay
initialLearningRateblah/watt-activation/169
0.00002
currentEpochblah/watt-activation/169
2
iterationsblah/watt-activation/408
100
batchSizeblah/watt-activation/408
64
sequenceLengthblah/watt-activation/408
8000
appliesToblah/watt-activation/408
ex:both-configs

References (3)

3 references
  1. [1]1398 facts
    ctx:discord/blah/watt-activation/139
    • full textwatt-activation-139
      text/plain2 KBdoc:agent/watt-activation-139/84b665a8-03e4-433b-97dc-f40068180b02
      Show excerpt
      [2026-03-09 07:23] xenonfun: ⏺ Training is resumed from step 10000. The NaN is fixed — all three patches held: 1. r_global = mx.sqrt(sum_sq + 1e-8) — eps inside sqrt for order parameter 2. mean_spec / mx.sqrt(sum_sq + 1e-8) — eps insid
  2. [2]1699 facts
    ctx:discord/blah/watt-activation/169
    • full textwatt-activation-169
      text/plain3 KBdoc:agent/watt-activation-169/41f410cf-a551-436e-a024-061c08e0e3ef
      Show excerpt
      [2026-03-09 20:11] xenonfun: ⏺ Still confused — it sees "Write a short poem" and almost gets there ("He was a large time... It was a dream of a song") but then it mixes in code and FineWeb patterns. The model knows the instruction format
  3. [3]4084 facts
    ctx:discord/blah/watt-activation/408
    • full textwatt-activation-408
      text/plain2 KBdoc:agent/watt-activation-408/5084227c-09bf-490d-affa-7c62f4061de3
      Show excerpt
      [2026-03-19 06:38] xenonfun: ⏺ Yes — the same way we did decay gates. The coupling strength should be derived from the oscillator state, not fixed: Principled approach: coupling = f(coherence between the two groups being coupled). Groups

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.