Model Card for Notux 8x7B-v1

This model is a preference-tuned version of mistralai/Mixtral-8x7B-Instruct-v0.1 on the argilla/ultrafeedback-binarized-preferences-cleaned dataset using DPO (Direct Preference Optimization).

As of Dec 26th 2023, it outperforms Mixtral-8x7B-Instruct-v0.1 and is the top ranked MoE (Mixture of Experts) model on the Hugging Face Open LLM Leaderboard.

This is part of the Notus family of models and experiments, where the Argilla team investigates data-first and preference tuning methods like dDPO (distilled DPO). This model is the result of our first experiment at tuning a MoE model that has already been fine-tuned with DPO (i.e., Mixtral-8x7B-Instruct-v0.1).

Model Details

Model Description

• Developed by: Argilla (based on MistralAI previous efforts)
• Shared by: Argilla
• Model type: Pretrained generative Sparse Mixture of Experts
• Language(s) (NLP): English, Spanish, Italian, German, and French
• License: MIT
• Finetuned from model: mistralai/Mixtral-8x7B-Instruct-v0.1

Model Sources

• Repository: https://github.com/argilla-io/notus
• Paper: N/A

Training Details

Training Hardware

We used a VM with 8 x H100 80GB hosted in runpod.io for 1 epoch (~10hr).

Training Data

We used a new iteration of the Argilla UltraFeedback preferences dataset named argilla/ultrafeedback-binarized-preferences-cleaned.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-07
• trainbatchsize: 8
• evalbatchsize: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• totaltrainbatch_size: 64
• totalevalbatch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lrschedulertype: linear
• lrschedulerwarmup_ratio: 0.1
• num_epochs: 1

Training results

Framework versions

• Transformers 4.36.0
• Pytorch 2.1.0+cu118
• Datasets 2.14.6
• Tokenizers 0.15.0