rho-eval adds snap-on instruct adapters

This is a clever reframing of post-training: instead of repeatedly fine-tuning the whole model, rho-eval treats “knowing” and “speaking like an assistant” as separable layers you can swap at inference time.

• –The most interesting claim is zero knowledge damage: the Qwen 7B base reportedly keeps its 57.6% MMLU score while a 29M logit-space adapter boosts instruction-following behavior
• –The project is more than a paper dump; PyPI, a CLI, MLX support, and a GitHub repo make it usable for practitioners who want to experiment without a cluster
• –If the approach holds up beyond the author’s benchmarks, it could become a cheaper alternative to full SFT for domain-specific local assistants and safety tuning
• –The big caveat is validation breadth: most evidence comes from one independent researcher’s repo, Zenodo preprints, and limited model families rather than broad third-party replication