Unified Rotor Gpu 25m
From Dontopedia, the open, paraconsistent wiki. (Last updated 2026-06-06.)
Unified Rotor Gpu 25m has 75 facts recorded in Dontopedia across 2 references, with 4 live disagreements.
Mostly:has best bpb at step(6), has bpb at step(6), has parameter count(2)
Maturity scale
raw canonical shape-checked rule-derived certifiedHas Parameter Countin disputehasParameterCount
Crossed Bpb Threshold at Approx Stepin disputecrossedBpbThresholdAtApproxStep
Has Best Bpb at Stepin disputehasBestBpbAtStep
Has Bpb at Stepin disputehasBpbAtStep
Approaches Bpb ofapproachesBpbOf
- Kura Pp Warm Restart Wavenative[2]all time · Part 682
Assumes Trajectory ThroughassumesTrajectoryThrough
- Cosine Decay[2]all time · Part 682
Crossed Bpb ThresholdcrossedBpbThreshold
- 1.89[2]all time · Part 682
Crossed Bpb Threshold of Lohe Delta E BindingcrossedBpbThresholdOfLoheDeltaEBinding
- 2.1[2]all time · Part 682
Current Training StepcurrentTrainingStep
- 7220[2]all time · Part 682
Has Best Bpb So FarhasBestBpbSoFar
- 1.56[2]all time · Part 682
Has Best Bpb Value at Step1000hasBestBpbValueAtStep1000
- 2.2[2]all time · Part 682
Has Best Bpb Value at Step2000hasBestBpbValueAtStep2000
- 1.86[2]all time · Part 682
Inbound mentions (9)
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.
isEasyDoReMiCategoryIs Easy Do Re Mi Category(2)
- Narrative
ex:narrative - Tinystories
ex:tinystories
isHardestDoReMiCategoryIs Hardest Do Re Mi Category(2)
- Dense Science
ex:dense-science - Edu
ex:edu
comparesRunsCompares Runs(1)
- Xenonfun Message 2026 04 22 14 42
ex:xenonfun-message-2026-04-22-14-42
givesOutlookGives Outlook(1)
- Xenonfun Message Outlook
ex:xenonfun-message-outlook
improvesOverTimeSlowerImproves Over Time Slower(1)
- Lohe Delta 25m Replication
ex:lohe-delta-25m-replication
storesDataForStores Data for(1)
- Metrics Jsonl
ex:metrics-jsonl
subjectOfForecastSubject of Forecast(1)
- Forecast 20k Step
ex:forecast-20k-step
Other facts (51)
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.
| Predicate | Value | Ref |
|---|---|---|
| Has Best Bpb Value at Step3000 | 1.71 | [2] |
| Has Best Bpb Value at Step5000 | 1.65 | [2] |
| Has Best Bpb Value at Step7000 | 1.61 | [2] |
| Has Best Bpb Value at Step7200 | 1.56 | [2] |
| Has Best Bpb Value at Step7220 | 1.56 | [2] |
| Has Bpb Value at Step1000 | 2.28 | [2] |
| Has Bpb Value at Step2000 | 1.94 | [2] |
| Has Bpb Value at Step3000 | 1.91 | [2] |
| Has Bpb Value at Step5000 | 1.72 | [2] |
| Has Bpb Value at Step7000 | 1.8 | [2] |
| Has Bpb Value at Step7220 | 1.74 | [2] |
| Has Current Best Bpb Distance to Wavenative | 0.22 | [2] |
| Has Do Re Mi Weight Dense Science | 0.21 | [2] |
| Has Do Re Mi Weight Edu | 0.19 | [2] |
| Has Do Re Mi Weight Narrative | 0.06 | [2] |
| Has Do Re Mi Weight Tinystories | 0.06 | [2] |
| Has Remaining Steps to Total | 12800 | [2] |
| Has Throughput Tok Per Sec | 23400 | [2] |
| Is Ahead by Bpb | 1.2 | [2] |
| Is Ahead of Prior Run at Same Step | Best Comparable Prior Run | [2] |
| Is Slower Per Token by Factor | 2.7 | [2] |
| Predicted Best Bpb at Step20000 | 1.35-1.45 | [2] |
| Predicted Better Than | Lohe Delta Variants | [2] |
| Predicted Competitive With | Wavenative 25m Production | [2] |
| Represents Net Win | Wall Clock Time | [2] |
| Uses Method | Doremi | [2] |
| Crosses Threshold Earlier Than | Lohe Delta a | [2] |
| Expected to Be Competitive | Wavenative 25m Production | [2] |
| Implies Superior Learning Rate | Prior Runs | [2] |
| Improves Over Time | Bpb Decrease | [2] |
| Predicted to Be Much Better Than | Lohe Delta Variants | [2] |
| Presupposes Prior Runs Exist | Comparable Prior Runs | [2] |
| Has Schedule | cosine decay | [1] |
| Advantageous in | Wall Clock Time | [1] |
| Slowness Factor | 2.7 | [1] |
| Slower Than | Lohe Delta 25m Replication | [1] |
| Has Throughput | 23400 | [1] |
| Approaches Value | 1.392 | [1] |
| Has Component | Unified Block | [1] |
| Target Is | Wavenative 25m Production | [1] |
| Steps Remaining | 12800 | [1] |
| Distance From Target Bpb | 0.22 | [1] |
| Crossed Bpb of | 2.10 at ~1500 | [1] |
| Crossed Bpb at Step | 1.89 at 2000 | [1] |
| Outperforms | Lohe Delta 25m Replication | [1] |
| Outperforms by Margin | ~1.2 BPB | [1] |
| Has Metric | Bpb | [1] |
| Current Step | 7220 | [1] |
| Uses Data | Doremi Dataset | [1] |
| Rdfs:label | unified_rotor_gpu_25m | [1] |
| Teleologically Aims for Low Bpb | Training Objective | [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.
References (2)
- custom
ctx:discord/blah/watt-activation/679- full textwatt-activation-679text/plain2 KB
doc:agent/watt-activation-679/e7b96ce5-bed6-4b84-baf1-a773f080c0dbShow excerpt
[2026-04-22 14:42] xenonfun: ``` ⏺ Pulling the per-step bpb trajectory from metrics.jsonl of the two most-comparable prior runs: step | unified_rotor_gpu_25m | lohe_delta_25m_replication | lohe_delta_16L_G16_H4_R16_doremi …
- custom
ctx:discord/blah/watt-activation/part-682
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.