metadata

language:
  - en
  - de
  - es
  - fr
  - it
license: apache-2.0
library_name: transformers
tags:
  - dpo
  - rlaif
  - preference
  - ultrafeedback
  - moe
base_model: mistralai/Mixtral-8x7B-Instruct-v0.1
datasets:
  - argilla/ultrafeedback-binarized-preferences-cleaned
pipeline_tag: text-generation
model-index:
  - name: notux-8x7b-v1
    results: []

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 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
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

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	73.18
AI2 Reasoning Challenge (25-Shot)	70.99
HellaSwag (10-Shot)	87.73
MMLU (5-Shot)	71.33
TruthfulQA (0-shot)	65.79
Winogrande (5-shot)	81.61
GSM8k (5-shot)	61.64

Open Portuguese LLM Leaderboard Evaluation Results

Detailed results can be found here and on the 🚀 Open Portuguese LLM Leaderboard

Metric	Value
Average	73.1
ENEM Challenge (No Images)	70.96
BLUEX (No Images)	60.22
OAB Exams	49.52
Assin2 RTE	92.66
Assin2 STS	82.40
FaQuAD NLI	79.85
HateBR Binary	77.91
PT Hate Speech Binary	73.30
tweetSentBR	71.08