The RM training pairwise loss function uses
AMean squared error on absolute reward scores
BCross-entropy hinge over pairwise margins
CSigmoid of the reward difference
DTriplet loss on response embeddings
Answer & Solution
Correct answer: C. Sigmoid of the reward difference
Loss = -log(sigmoid(r(x, y_w) - r(x, y_l))) where y_w is preferred over y_l. This makes the RM's reward gap reflect the human preference probability.
Related questions
The FLAN and T0 datasets are mentioned as baselines because theyWhy is honesty measured via 'truthfulness' rather than directly?Prioritising helpfulness during training while prioritising harmlessness at evaluation is RM models in InstructGPT are sized atInter-annotator agreement among training labelers is roughlyRM training uses all C(K,2) comparisons from a single prompt as'PPO-ptx' differs from plain PPO byThe PPO objective includes a KL penalty against the SFT model to