This model is identical in training to HuggingFaceH4/zephyr-7b-gemma-v0.1, except instead of using Direct Preference Optimization (DPO), it uses DiscoPOP.
DiscoPOP is our Discovered Preference Optimization algorithm, which is defined as follows:
def log_ratio_modulated_loss(
self,
policy_chosen_logps: torch.FloatTensor,
policy_rejected_logps: torch.FloatTensor,
reference_chosen_logps: torch.FloatTensor,
reference_rejected_logps: torch.FloatTensor,
) -> torch.FloatTensor:
pi_logratios = policy_chosen_logps - policy_rejected_logps
ref_logratios = reference_chosen_logps - reference_rejected_logps
logits = pi_logratios - ref_logratios
# Modulate the mixing coefficient based on the log ratio magnitudes
log_ratio_modulation = torch.sigmoid(logits)
logistic_component = -F.logsigmoid(self.beta * logits)
exp_component = torch.exp(-self.beta * logits)
# Blend between logistic and exponential component based on log ratio modulation
losses = logistic_component * (1 - log_ratio_modulation) + exp_component * log_ratio_modulation
return losses
Training hyperparameters
The following hyperparameters were used during training:
learning_rate: 5e-07
train_batch_size: 2
eval_batch_size: 4
seed: 42
distributed_type: multi-GPU
num_devices: 8
gradient_accumulation_steps: 8
total_train_batch_size: 128
total_eval_batch_size: 32
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
