SKT-NRS

/

SKT-SURYA-H

Like 23

Follow

SKT-DATASETS-NRS 22

Thanks for reporting @DavCd ! Can you please clarify what exactly is spam in this case? Thanks! 🤗

Thanks for reporting @DavCd ! Can you please clarify what exactly is spam in this case? Thanks! 🤗

completely confabulated benchmark table
quotes HLE as 'graduate level reasoning' and contains benchmarks that does not exist.
their 'technical paper' linked is pretty much nonsense.

forgot to ping @adamm-hf

Hello @adamm-hf and @DavCd ,

​Thank you for the feedback. I would like to clarify a few points regarding the SKT-SURYA-H model:
​Benchmark Table: The benchmarks provided are based on our internal evaluations using the Weight Manifold Fusion (WMF) architecture. If there are any naming inconsistencies with public leaderboards like HLE, it might be due to our specific formatting or localized testing variants. We are happy to re-verify and update the table to reflect standardized metrics if needed.

​Technical Paper: The linked document is a pre-print focusing on the theoretical framework of Heterogeneous Neural Manifolds. While it uses specialized terminology, it outlines our approach to fusing disparate model weights. We understand it may appear non-traditional, as this is an independent research project from Sidhi, India, exploring new scaling laws.

​Authenticity: This is a genuine attempt to build a large-scale Sovereign AI model (2.544T). We are a small independent lab, and our resources are different from mainstream corporate labs. We are committed to transparency and will refine the documentation and README to avoid any further confusion.

​Our goal is contribution, not spam. We request some time to align our model card with the community's reporting standards."

look, not to offend you or your country or anything but there are a LOT of nonsense jargon in the model card with a lot of inconsistencies such as "Context Length: 1M tokens (natively supported)" while you say "146 Trillion Token Context: Near-infinite context through Vajra-Attention (O(n log n)) and Sudarshan-Link quantum-entangled synchronization." well first of all, you can get O(N) directly from ALiBi, you dont need to reinvent the wheel;
also, you have "Number of Parameters: 2.544 Trillion total, ~17B activated per token" that is approximately 149 experts while you cite having 5;
plus the expert_synergy values:
[1.0, 0.540302, -0.416147, -0.989992, -0.653644, 0.283662],
those are LITERALLY cos(x) and sin(x) in sequence.
there are a LOT more discrepancies but these are the ones that i gave you now i havent tried the model yet so.