Zephyr is a series of language models that are trained to act as helpful assistants. Zephyr 141B-A35B is the latest model in the series, and is a fine-tuned version of mistral-community/Mixtral-8x22B-v0.1 that was trained using a novel alignment algorithm called Odds Ratio Preference Optimization (ORPO) with 7k instances for 1.3 hours on 4 nodes of 8 x H100s. ORPO does not require an SFT step to achieve high performance and is thus much more computationally efficient than methods like DPO and PPO. To train Zephyr-141B-A35B, we used the argilla/distilabel-capybara-dpo-7k-binarized preference dataset, which consists of synthetic, high-quality, multi-turn preferences that have been scored via LLMs.
[!NOTE]
This model was trained collaboratively between Argilla, KAIST, and Hugging Face
Model Details
Model Description
Model type: A Mixture of Experts (MoE) model with 141B total parameters and 35B active parameters. Fine-tuned on a mix of publicly available, synthetic datasets.
Zephyr 141B-A35B was trained to test the effectiveness of ORPO at scale and the underlying dataset contains a mix of general chat capabilities. It achieves strong performance on chat benchmarks like MT Bench and IFEval. The scores reported below were obtained using the LightEval evaluation suite and each prompt has been formatted with the model's corresponding chat template to simulate real-world usage. This is why some scores may differ from those reported in technical reports or on the Open LLM Leaderboard.
The model was fine-tuned on a blend of chat, code, math, and reasoning data. Here's how you can run the model using the pipeline() function from 🤗 Transformers:
python
1# pip install 'transformers>=4.39.3'2# pip install accelerate3import torch
4from transformers import pipeline
5pipe = pipeline(6"text-generation",7 model="HuggingFaceH4/zephyr-orpo-141b-A35b-v0.1",8 device_map="auto",9 torch_dtype=torch.bfloat16,10)11messages =[12{13"role":"system",14"content":"You are Zephyr, a helpful assistant.",15},16{"role":"user","content":"Explain how Mixture of Experts work in language a child would understand."},17]18outputs = pipe(19 messages,20 max_new_tokens=512,21 do_sample=True,22 temperature=0.7,23 top_k=50,24 top_p=0.95,25)26print(outputs[0]["generated_text"][-1]["content"])
Bias, Risks, and Limitations
Zephyr 141B-A35B 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 (mistral-community/Mixtral-8x22B-v0.1), however it is likely to have included a mix of Web data and technical sources like books and code. See the Falcon 180B model card for an example of this.
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
learning_rate: 5e-06
train_batch_size: 1
eval_batch_size: 8
seed: 42
distributed_type: multi-GPU
num_devices: 32
total_train_batch_size: 32
total_eval_batch_size: 256
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: inverse_sqrt
lr_scheduler_warmup_steps: 100
num_epochs: 3
Training results
Framework versions
Transformers 4.39.3
Pytorch 2.1.2+cu121
Datasets 2.18.0
Tokenizers 0.15.1
Citation
If you find Zephyr 141B-A35B is useful in your work, please cite the ORPO paper:
@misc{hong2024orpo,
title={ORPO: Monolithic Preference Optimization without Reference Model},
author={Jiwoo Hong and Noah Lee and James Thorne},
year={2024},
eprint={2403.07691},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
You may also wish to cite the creators of this model:
@misc{zephyr_141b,
author = {Alvaro Bartolome and Jiwoo Hong and Noah Lee and Kashif Rasul and Lewis Tunstall},
title = {Zephyr 141B A35B},
year = {2024},
publisher = {Hugging Face},
journal = {Hugging Face repository},
howpublished = {\url{https://huggingface.co/HuggingFaceH4/zephyr-orpo-141b-A35b-v0.1}}
}
