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

Model Card for Zephyr 7B Gemma

Zephyr is a series of language models that are trained to act as helpful assistants. Zephyr 7B Gemma is the third model in the series, and is a fine-tuned version of google/gemma-7b that was trained on on a mix of publicly available, synthetic datasets using Direct Preference Optimization (DPO). You can reproduce the training of this model via the recipe provided in the Alignment Handbook.

Model description

• Model type: A 7B parameter GPT-like model fine-tuned on a mix of publicly available, synthetic datasets.
• Language(s) (NLP): Primarily English
• License: Gemma Terms of Use
• Finetuned from model: google/gemma-7b

Model Sources

• Repository: https://github.com/huggingface/alignment-handbook
• Demo: https://huggingface.co/spaces/HuggingFaceH4/zephyr-7b-gemma-chat

Performance

Intended uses & limitations

The model was initially fine-tuned on the DEITA 10K dataset, which contains a diverse range of synthetic dialogues generated by ChatGPT. We then further aligned the model with 🤗 TRL’s DPOTrainer on the argilla/dpo-mix-7k dataset, which contains 7k prompts and model completions that are ranked by GPT-4. As a result, the model can be used for chat and you can check out our demo to test its capabilities.

Here’s how you can run the model using the pipeline() function from 🤗 Transformers:

# pip install transformers>=4.38.2
# pip install accelerate

import torch
from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="HuggingFaceH4/zephyr-7b-gemma-v0.1",
    device_map="auto",
    torch_dtype=torch.bfloat16,
)
messages = [
    {
        "role": "system",
        "content": "",  # Model not yet trained for follow this
    },
    {"role": "user", "content": "How many helicopters can a human eat in one sitting?"},
]
outputs = pipe(
    messages,
    max_new_tokens=128,
    do_sample=True,
    temperature=0.7,
    top_k=50,
    top_p=0.95,
    stop_sequence="<|im_end|>",
)
print(outputs[0]["generated_text"][-1]["content"])
# It is not possible for a human to eat a helicopter in one sitting, as a
# helicopter is a large and inedible machine. Helicopters are made of metal,
# plastic, and other materials that are not meant to be consumed by humans.
# Eating a helicopter would be extremely dangerous and would likely cause
# serious health problems, including choking, suffocation, and poisoning. It is
# important to only eat food that is safe and intended for human consumption.

Bias, Risks, and Limitations

Zephyr 7B Gemma has not been aligned to human preferences for safety within the RLHF phase or deployed with in-the-loop filtering of responses like ChatGPT, so the model can produce problematic outputs (especially when prompted to do so). It is also unknown what the size and composition of the corpus was used to train the base model (google/gemma-7b), however it is likely to have included a mix of Web data and technical sources like books and code. See the StarCoder2 model card for an example of this.

Training and evaluation data

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

It achieves the following results on the evaluation set: - Loss: 0.4695 - Rewards/chosen: -3.3746 - Rewards/rejected: -4.9715 - Rewards/accuracies: 0.7188 - Rewards/margins: 1.5970 - Logps/rejected: -459.4853 - Logps/chosen: -429.9115 - Logits/rejected: 86.4684 - Logits/chosen: 92.8200

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

Training results

Framework versions

• Transformers 4.39.0.dev0
• Pytorch 2.1.2+cu121
• Datasets 2.14.6
• Tokenizers 0.15.1

Citation Information

If you find this model useful in your work, please consider citing the Zephyr technical report:

@misc{tunstall2023zephyr,
      title={Zephyr: Direct Distillation of LM Alignment}, 
      author={Lewis Tunstall and Edward Beeching and Nathan Lambert and Nazneen Rajani and Kashif Rasul and Younes Belkada and Shengyi Huang and Leandro von Werra and Clémentine Fourrier and Nathan Habib and Nathan Sarrazin and Omar Sanseviero and Alexander M. Rush and Thomas Wolf},
      year={2023},
      eprint={2310.16944},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

You may also wish to cite the creators of this model as well:

@misc{zephyr_7b_gemma,
  author = {Lewis Tunstall and Philipp Schmid},
  title = {Zephyr 7B Gemma},
  year = {2024},
  publisher = {Hugging Face},
  journal = {Hugging Face repository},
  howpublished = {\url{https://huggingface.co/HuggingFaceH4/zephyr-7b-gemma-v0.1}}
}

Open LLM Leaderboard Evaluation Results

Detailed results can be found here