AGORA
·
Finished
·
Pluralis-8B
Pluralis-8B
·
531.8B Tokens processed
Docs and Joining Instructions
Docs
Nodes
0
/ 592
Contributors
0
/ 98
Peak
112
Loss
2.438
TPS
5.9 kTPS
Protocol overhead
38
%
Pluralis
0
%
Showing single run
·
Pluralis-8B
Finished
Overview
Telemetry
Evals
Analysis
Hardware
Geography
Nodes
Leaderboard
Pluralis-8B
Finished
Model
Pluralis-8B
Dataset
FineWeb-Edu (1.3T)
Smooth
0.00
Y: Lin
X
Elapsed
Step
Training loss
Cross-entropy
Clamp Y axis
Throughput
Tokens / sec
Throughput per TFLOP
Tok/s divided by active dense BF16 TFLOPs · higher is better
4.20
~
Baselines
Baselines: optimal central setting
6.74
and Megatron-LM at 200 Mbps
0.37
tok/s/TFLOP. See Analysis tab for full breakdown + sources.
Total swarm TFLOPS
Active dense BF16 compute · Pluralis-operated vs Contributor split
39.88k
Pluralis 56%
Contributor 46%
Pluralis / Contributor split
Mean MFU
Model FLOP utilisation · achieved ÷ peak · higher is better (Chinchilla 6N, 8B, denom: BF16)
20.2%
Evaluations
No data yet for Pluralis-8B
Benchmark results will appear here as the training run progresses.