SteerLM & HelpSteer Multi-Attribute Conditioned Post-Training

Replaces single-scalar reward modeling with a 5-dimensional regression Reward Model (HelpSteer / HelpSteer2 / HelpSteer3) predicting explicit human-interpretable attributes (Helpfulness, Correctness, Coherence, Complexity, Verbosity) $\in [0, 4]^5$. During SteerLM SFT, responses are conditioned on their ground-truth attribute vector tag in the prompt, and at inference time or during online RL, attributes are set to maximum quality (Helpfulness=4, Correctness=4, Coherence=4) while explicitly controlling Verbosity to prevent length hacking.