DiscoPOP-zephyr-7b-gemma

This model is a fine-tuned version of HuggingFaceH4/zephyr-7b-gemma-sft-v0.1 on the argilla/dpo-mix-7k dataset.

This model is from the paper "Discovering Preference Optimization Algorithms with and for Large Language Models"

Read the blog post on it here!

See the codebase to generate it here: https://github.com/SakanaAI/DiscoPOP

Model description

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
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 2

Framework versions

• Transformers 4.40.1
• Pytorch 2.1.2+cu121
• Datasets 2.19.0
• Tokenizers 0.19.1