SALOMI Reveals Hidden Trade-offs in Extreme Transformer Quantization

SALOMI challenges overstated claims about binary quantization for large language models, providing rigorous evidence that naive 1-bit compression fails for GPT-2-scale models under realistic evaluation. The repository's critical contribution is its honesty about practical trade-offs, showing that viable extreme-compression techniques cluster around 1.2-1.35 bits per parameter using methods like Hessian-guided VQ. For ML practitioners and researchers, this work is essential for setting realistic expectations about model compression limits and avoiding misleading performance claims.