Simpo loss

SimPO Loss Implementation

Core Formula (Eq. 6 from the paper)

The SimPO objective:

L_SimPO(πθ) = -E log σ( (β/|yw|) log πθ(yw|x) - (β/|yl|) log πθ(yl|x) - γ )

Where:

• β (beta): scaling factor for reward difference (default: 2.0)
• γ (gamma): target reward margin, computed as gamma_beta_ratio * beta (default ratio: 0.25)
• |yw|, |yl|: lengths of winning/losing responses
• σ: sigmoid function

Key Design Decisions

1. Length-normalized reward: r(x,y) = (β/|y|) * log πθ(y|x) — the average log probability scaled by β
2. Target reward margin γ: ensures winning response reward exceeds losing by at least γ
3. No reference model needed — unlike DPO

Implementation Pattern

Given policy_chosen_logps and policy_rejected_logps (already length-normalized average log probs):

pi_logratios = policy_chosen_logps - policy_rejected_logps
gamma = self.gamma_beta_ratio * self.beta
logits = pi_logratios * self.beta - gamma

# Sigmoid loss (with optional label smoothing)
losses = (
    -F.logsigmoid(logits) * (1 - self.label_smoothing)
    - F.logsigmoid(-logits) * self.label_smoothing
)

# Hinge loss variant
# losses = torch.relu(1 - logits)

chosen_rewards = self.beta * policy_chosen_logps.detach()
rejected_rewards = self.beta * policy_rejected_logps.detach()

Config Parameters (SimPOConfig)

Parameter	Default	Description
beta	2.0	Reward scaling factor
gamma_beta_ratio	0.25	γ/β ratio
label_smoothing	0.0	Label smoothing factor
loss_type	"sigmoid"	"sigmoid" or "hinge"
sft_weight	0.0	Optional SFT loss weight

Returns

Tuple of (losses, chosen_rewards, rejected_rewards) — all tensors of shape (batch_size,).