simpo-loss

name	simpo-loss
description	Guide on understanding and implementing the SimPO (Simple Preference Optimization) loss function for language models.

SimPO Loss

SimPO (Simple Preference Optimization) is an alternative to DPO for preference learning. It optimizes the reward directly on the sequence level without requiring a reference model, using the average log probability of the generated sequence.

Core Formula

The SimPO loss involves:

1. Calculating the average log probability per token for the chosen and rejected sequences.
2. Adding a target reward margin $\gamma$ (gamma).
3. Passing the difference through a sigmoid function (specifically, -log(sigmoid(diff))).

Formula: $$L_{\text{SimPO}} = - \log \sigma \left( \beta \left( \frac{1}{|y_w|} \sum_{i} \log p(y_{w,i}|x) - \frac{1}{|y_l|} \sum_{i} \log p(y_{l,i}|x) \right) - \gamma \right)$$

where:

• $\beta$ is a scaling factor.
• $\gamma$ is the reward margin.
• $y_w$ is the chosen/winning sequence.
• $y_l$ is the rejected/losing sequence.

Usage

In a trainer like DPOTrainer, this can be implemented by computing the average log probability for both policy and reference (though SimPO only needs the policy logprobs, they might be formatted similar to DPO).