| name | simpo-loss-function |
| description | Implement SimPO loss with length-normalized rewards and target margin. |
SimPO Loss Function Implementation
Overview
SimPO (Simple Preference Optimization) implements a preference optimization objective that uses length-normalized average log probability as an implicit reward, with a target reward margin component.
Key Formula
L_SimPO(πθ) = -E_(x,yw,yl)~D log σ(β/|yw| log πθ(yw|x) - β/|yl| log πθ(yl|x) - γ)
Components
1. Length-Normalized Reward
- Formula:
r_SimPO(x, y) = β/|y| * log πθ(y|x)
- Purpose: Average log probability per token, prevents length bias
- Why: Aligns training with generation metric (which uses average log likelihood for beam search)
2. Bradley-Terry Objective
- Formula:
p(yw ≻ yl | x) = σ(r(x, yw) - r(x, yl) - γ)
- Purpose: Probabilistic ranking between winning and losing responses
- Function: σ is sigmoid function
3. Target Reward Margin (γ)
- Purpose: Ensure reward difference exceeds a target threshold
- Effect: Improves generalization by enforcing margin between classes
- Typical range: 0.5 to 1.5
Implementation Details
Computing Log Probabilities
log_prob_sum = log_probs.sum(dim=1)
seq_lengths = (input_ids != pad_token_id).sum(dim=1)
avg_log_prob = log_prob_sum / seq_lengths.float()
Computing Reward Differences
batch_size = avg_log_probs.shape[0]
winning_rewards = avg_log_probs[:batch_size//2]
losing_rewards = avg_log_probs[batch_size//2:]
reward_diff = beta * winning_rewards - beta * losing_rewards - gamma
Computing Loss
import torch.nn.functional as F
sigmoid_term = torch.sigmoid(reward_diff)
loss = -torch.log(sigmoid_term).mean()
Common Pitfalls
- Not using length normalization: Creates bias toward longer sequences
- Wrong batch structure: Ensure paired winning/losing responses
- Missing average in log probabilities: Use sum/length, not just sum
- Gradient flow: Ensure no detach() breaks gradients to model
Hyperparameters
- β: Temperature parameter, typically 2.0-2.5
- γ: Target margin, typically 0.3-1.6, depends on setting
- learning_rate: Usually small, 1e-6 to 5e-7
References
- SimPO Paper: Section 2.3 "The SimPO Objective"
- Length normalization: Equation (3) in paper
- Gradient analysis: Appendix F