Supra-50M
A Llama-style reference model developed by QLNI for comparative experiments, with 51.8M parameters trained using standard floating-point arithmetic. It serves as a baseline for evaluating the compression effectiveness of SHADOW-50M.
Timeline (last 90 days)
The reference model Supra-50M has 51.8M parameters, is 103.6 MB in bf16, 56.2 MB after 8-bit quantization, perplexity degrades from 165 to 193 when reduced to int4, and collapses entirely when compressed to ternary.
Supra-50M leads SHADOW on standard benchmarks: ARC-Easy 0.435 vs. 0.307, PIQA 0.600 vs. 0.570, and WikiText-2 perplexity 165 vs. 186.
All Facts (2)
The reference model Supra-50M has 51.8M parameters, is 103.6 MB in bf16, 56.2 MB after 8-bit quantization, perplexity degrades from 165 to 193 when reduced to int4, and collapses entirely when compressed to ternary.
50%UnverifiedSupra-50M leads SHADOW on standard benchmarks: ARC-Easy 0.435 vs. 0.307, PIQA 0.600 vs. 0.570, and WikiText-2 perplexity 165 vs. 186.
50%