Steerling-8B is an interpretable causal diffusion language model that combines masked diffusion language modeling with concept decomposition, enabling generation, attribution, steering, and extraction of hidden representations. It offers features like block-causal attention and decomposition of hidden states into known and unknown concepts.
