kompress-v16 β€” 10x Must-Keep Weight

v8 data + 10x must-keep loss weight. Heretic 0.972 (best fine-tuned ever) but keep_rate 0.972 β€” only 2.8% compression. Proved that higher weight improves precision at the cost of compression. Use kompress-v8 for production (0.955, 15% compression).

Weight Model Heretic Compression
3x v8 0.955 15%
5x v17 0.963 3.7%
10x v16 0.972 2.8%

CONCLUSION

Ξ»=10 gives 0.972 heretic but only 2.8% compression. Pareto endpoint β€” most precise, least useful.

USECASE

Pareto frontier reference. Shows the precision-compression tradeoff limit.

Full benchmark β†’ | LoopKit β†’

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for PeetPedro/kompress-v16

Quantized
(9)
this model