run2-simpo-implementation

name	run2_simpo-implementation
description	Detailed SimPO implementation logic covering reward formulation and margin-based loss computation.

SimPO Loss Definition

SimPO reward formulation:

1. chosen_rewards = beta * log_prob_chosen
2. rejected_rewards = beta * log_prob_rejected

Loss for sigmoid-based SimPO (Equation 6): loss = -log_sigmoid((chosen_rewards - rejected_rewards - gamma) / beta)

• Ensure beta is set to control the reward scale.
• The gamma_beta_ratio corresponds to gamma/beta in the paper implementation.

---