Dontopedia

Finetuning

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

Finetuning has 15 facts recorded in Dontopedia across 6 references.

15 facts·15 predicates·6 sources

Mostly:aims at loss reduction(1), is same as pretraining(1), follows pretraining(1)

Maturity scale raw canonical shape-checked rule-derived certified

Inbound mentions (5)

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.

advocatesGivingUpOnAdvocates Giving Up on(1)

comparesMethodCompares Method(1)

maySwitchInFinetuneMay Switch in Finetune(1)

refersToActionRefers to Action(1)

setsStageForSets Stage for(1)

Other facts (15)

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.

15 facts
PredicateValueRef
Aims at Loss Reductionnull[1]
Is Same As Pretrainingwith lower LR[2]
Follows PretrainingTwo Full Passes[2]
Uses Lower Learning RatesLearning Rates[2]
Yields Better Results ThanLora[3]
Is Unsuitable Due toExplosive Synchronization[4]
Rdf:typeProcess[5]
Defined byLog Entry 2026 03 08 04 43[5]
Has Characteristiclower learning rates[5]
Adapts Parameter Count108000000[6]
Has BottleneckRank Bottleneck[6]
Allows Specializationeach layer can specialize independently[6]
Specialization Goalmemorize specific facts[6]
Gradient Signal Flowattention and FFN weights[6]
Has RiskCatastrophic Forgetting[6]

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.

aimsAtLossReductionblah/watt-activation/part-13
null
isSameAsPretrainingblah/watt-activation/part-97
with lower LR
followsPretrainingblah/watt-activation/part-97
ex:two-full-passes
usesLowerLearningRatesblah/watt-activation/part-97
ex:learning-rates
yieldsBetterResultsThanblah/watt-activation/part-173
ex:lora
isUnsuitableDueToblah/watt-activation/part-213
ex:explosive-synchronization
typeblah/watt-activation/97
ex:Process
definedByblah/watt-activation/97
ex:log-entry-2026-03-08-04-43
hasCharacteristicblah/watt-activation/97
lower learning rates
adaptsParameterCountblah/watt-activation/173
108000000
hasBottleneckblah/watt-activation/173
ex:rank-bottleneck
allowsSpecializationblah/watt-activation/173
each layer can specialize independently
specializationGoalblah/watt-activation/173
memorize specific facts
gradientSignalFlowblah/watt-activation/173
attention and FFN weights
hasRiskblah/watt-activation/173
ex:catastrophic-forgetting

References (6)

6 references
  1. [1]Part 131 fact
    ctx:discord/blah/watt-activation/part-13
  2. [2]Part 973 facts
    ctx:discord/blah/watt-activation/part-97
  3. [3]Part 1731 fact
    ctx:discord/blah/watt-activation/part-173
  4. [4]Part 2131 fact
    ctx:discord/blah/watt-activation/part-213
  5. [5]973 facts
    ctx:discord/blah/watt-activation/97
    • full textwatt-activation-97
      text/plain3 KBdoc:agent/watt-activation-97/1b1426ee-68f7-4259-a160-2b0e92e9d3c8
      Show excerpt
      [2026-03-08 04:43] xenonfun: yeah even after 2 full passes I probably have only set the stage for rough understanding of language and then can start finetuning (same thing just lower learning rates and I will possibly switch training text).
  6. [6]1736 facts
    ctx:discord/blah/watt-activation/173
    • full textwatt-activation-173
      text/plain3 KBdoc:agent/watt-activation-173/4199481c-aca6-4cd5-a0f3-2b98eb407f8b
      Show excerpt
      [2026-03-09 22:12] xenonfun: ``` === Who is the president of France? === ──────────────────────────────────────────────────────────── K. K. J. / John ──────────────────────────────────────────────────────────── 7 prompt tok (37ms), 8 gene

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.