metadata

library_name: transformers
tags:
  - moe
  - moah
  - mod
license: apache-2.0
datasets:
  - Locutusque/UltraTextbooks
language:
  - en

Model Card for Model ID

Model Details

Model Description

MoM: Mixture of Mixture

This Model is a test to combine Jamba architecture with 1.58 bits linear layers excpted for attention layer, mixture of attention head and mixture of depth.

The goal is to developpe and test if this kind of architectures have not too much quality loss for a fast inference.

Only 17.8M parameter over 1025 is in bf16 precision wich is ~ 1.7% of the total number of parameters

Model type: Mixture of attention head mixture of depth and mixture of expert 1.58bit linear layers excepted for attention layer
License: Apache licence 2.0

Model Sources [optional]

Repository: https://github.com/ostix360/optimized-LLM

How to Get Started with the Model

If you want to test this model please look at this repo at this commit

Training Details

wandb: training detail

Training Data

We use the first 100k data of Locutusque/UltraTextbooks to train this model

Training Procedure

We use adam-8 bits with default betas and epsilon values

Preprocessing [optional]

The data fit the model max length i.e. 512 tokens

Training Hyperparameters

Please look at the wandb metadata file or the train.py file in the repo to see the hyperparameters

Ostixe360
/

MoMv3-mixed-precision