JackCloudman's picture
Create README.md
cfdcdaf
metadata
datasets:
  - argilla/ultrafeedback-binarized-preferences-cleaned
language:
  - en
base_model: mistralai/Mixtral-8x7B-Instruct-v0.1
library_name: transformers
pipeline_tag: text-generation
tags:
  - dpo
  - rlaif
  - preference
  - ultrafeedback
license: apache-2.0
model-index:
  - name: notux-8x7b-v1
    results: []

ExllamaV2 3.5bpw version

A banner representing Notus, the wind god of the south, in a mythical and artistic style. The banner features a strong, swirling breeze, embodying the warm, wet character of the southern wind. Gracefully flowing across the scene are several paper planes, caught in the gentle yet powerful gusts of Notus. The background is a blend of warm colors, symbolizing the heat of the south, with hints of blue and green to represent the moisture carried by this wind. The overall atmosphere is one of dynamic movement and warmth.

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 HuggingFace H4 and MistralAI previous efforts)
  • Shared by: Argilla
  • Model type: Pretrained generative Sparse Mixture of Experts
  • Language(s) (NLP): Mainly English
  • License: MIT
  • Finetuned from model: mistralai/Mixtral-8x7B-Instruct-v0.1

Model Sources

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
  • train_batch_size: 8
  • eval_batch_size: 4
  • seed: 42
  • distributed_type: multi-GPU
  • num_devices: 8
  • total_train_batch_size: 64
  • total_eval_batch_size: 32
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_ratio: 0.1
  • num_epochs: 1

Training results

Training Loss Epoch Step Validation Loss Rewards/chosen Rewards/rejected Rewards/accuracies Rewards/margins Logps/rejected Logps/chosen Logits/rejected Logits/chosen
0.4384 0.22 200 0.4556 -0.3275 -1.9448 0.7937 1.6174 -405.7994 -397.8617 -1.3157 -1.4511
0.4064 0.43 400 0.4286 -0.2163 -2.2090 0.8254 1.9927 -408.4409 -396.7496 -0.7660 -0.6539
0.3952 0.65 600 0.4275 -0.1311 -2.1603 0.8016 2.0291 -407.9537 -395.8982 -0.6783 -0.7206
0.3909 0.87 800 0.4167 -0.2273 -2.3146 0.8135 2.0872 -409.4968 -396.8602 -0.8458 -0.7738

Framework versions

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

An experiment by