cross-architecture
cross-architecture-triangulation
Cross-Architecture Triangulation: Transformer vs. Diffusion vs. RWKV
31 prompts, 5 categories, run through three genuinely different inference paradigms on the same hardware (Apple Silicon, MPS):
Transformer (autoregressive): Qwen2.5-14B-Instruct-4bit, MLX, greedy decoding
Diffusion: LLaDA-8B-Instruct, masked denoising, 64 steps / 2 blocks of 32, low-confidence remasking
RWKV (recurrent/state-space): RWKV-Raven-1.5B, no attention
Built and run by Wayfinder6, Heuremen… See the full description on the dataset page: https://huggingface.co/datasets/Wayfinder6/cross-architecture-triangulation.cross-architecture-consciousness-validation-v1