Dontopedia

Evaluation 1

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

Evaluation 1 has 25 facts recorded in Dontopedia across 3 references, with 3 live disagreements.

25 facts·22 predicates·3 sources·3 in dispute

Mostly:compares to(2), possible cause of gap(2), requires reproduction of(2)

Maturity scale raw canonical shape-checked rule-derived certified

Inbound mentions (1)

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.

evaluatedPerformanceEvaluated Performance(1)

Other facts (25)

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.

25 facts
PredicateValueRef
Compares toGpt 2 Training[1]
Compares toRun Config 5[2]
Possible Cause of Gapforward logit parity compounding[3]
Possible Cause of Gapdataset version drift[3]
Requires Reproduction ofper-batch averaging[3]
Requires Reproduction ofbf16/tf32 quirks[3]
Assesses Duration13.5 days[1]
Hardware Configsingle consumer GPU[1]
Parameter Count145000000[1]
Training Token Count9000000000[1]
Opinionnot bad at all[1]
Entity EvaluatedLearned Gate[2]
Plateau Value4.58[2]
Worse ThanRun Config 5[2]
Reported inXenonfun[3]
Has Metric1.4192[3]
Metric Nameeval BPB[3]
Metric Conditionpython-stride[3]
Gap to Reference0.0041[3]
Previous Gap0.0072[3]
Compounding Over Tokens114000[3]
Reference Timestamp Context1.4151 was logged[3]
Requires Reproduction for Zero Gaptrue[3]
Reproduction Is Worth Chasingfalse[3]
Parity Band~0.01[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.

assessesDurationblah/watt-activation/94
13.5 days
hardwareConfigblah/watt-activation/94
single consumer GPU
parameterCountblah/watt-activation/94
145000000
trainingTokenCountblah/watt-activation/94
9000000000
opinionblah/watt-activation/94
not bad at all
comparesToblah/watt-activation/94
ex:GPT-2-training
entityEvaluatedblah/watt-activation/409
ex:learned-gate
plateauValueblah/watt-activation/409
4.58
worseThanblah/watt-activation/409
ex:run-config-5
comparesToblah/watt-activation/409
ex:run-config-5
reportedInblah/watt-activation/633
ex:xenonfun
hasMetricblah/watt-activation/633
1.4192
metricNameblah/watt-activation/633
eval BPB
metricConditionblah/watt-activation/633
python-stride
gapToReferenceblah/watt-activation/633
0.0041
previousGapblah/watt-activation/633
0.0072
possibleCauseOfGapblah/watt-activation/633
forward logit parity compounding
compoundingOverTokensblah/watt-activation/633
114000
possibleCauseOfGapblah/watt-activation/633
dataset version drift
referenceTimestampContextblah/watt-activation/633
1.4151 was logged
requiresReproductionForZeroGapblah/watt-activation/633
true
requiresReproductionOfblah/watt-activation/633
per-batch averaging
requiresReproductionOfblah/watt-activation/633
bf16/tf32 quirks
reproductionIsWorthChasingblah/watt-activation/633
false
parityBandblah/watt-activation/633
~0.01

References (3)

3 references
  1. [1]946 facts
    ctx:discord/blah/watt-activation/94
    • full textwatt-activation-94
      text/plain3 KBdoc:agent/watt-activation-94/3ac27cdb-b74d-48fa-9f59-8e2eb7c9c9ad
      Show excerpt
      [2026-03-08 02:30] xenonfun: okay so quick estimate how long would it take for us to train a full 9B tokens if we had them? I'm getting something like 13.5 days at 7.7Ktok/s ? ⏺ Your math checks out. 9B tokens / 7.7K tok/s = 1,168,831 seco
  2. [2]4094 facts
    ctx:discord/blah/watt-activation/409
    • full textwatt-activation-409
      text/plain2 KBdoc:agent/watt-activation-409/27cf5bf6-23d3-4bb2-8da9-9ee785b7e6a7
      Show excerpt
      [2026-03-19 13:28] xenonfun: ``` ⏺ Interesting contrast: ┌─────┬──────┬──────┬──────┬─────────┬──────┐ │ LR │ Step │ BPB │ r │ SNR │ DC │ ├─────┼──────┼──────┼──────┼─────────┼──────┤ │ 0.1 │ 350 │ 4.58 │ 0.41 │ -6.9dB
  3. [3]63315 facts
    ctx:discord/blah/watt-activation/633
    • full textwatt-activation-633
      text/plain2 KBdoc:agent/watt-activation-633/98cb2cdc-0a8e-4624-a978-d2d562057e94
      Show excerpt
      [2026-04-15 23:58] lisamegawatts: yours is way faster to train, the data sets are all in hf now, should do a 50m chinchilla optimal, i think we have enough data, use cleaned corpus and the synth ones but skip text pipeline probably as it di

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.