DistMerge_Llama-3.1-8B-Instruct

Model Card for Model ID

DistMerge_Llama-3.1-8B-Instruct is a customized variant of the VAGOsolutions/Llama-3.1-SauerkrautLM-8B-Instruct, which is itself a spectrum fine-tuned version of Llama-2.1-8B-Instruct. This customization is achieved by learning the distribution of all normalization layer weights from both the original Llama model and its fine-tuned counterpart. A layer-conditional diffusion based weights generation model that enables sampling for performance enhancement by leveraging the learned distributions to optimize the merging process is used to generate the normalization layer of bedio/DistMerge_Llama-3.1-8B-Instruct

Model Details

Model Description

We trained a diffusion model to learn the distribution of the normalization layers to enable generation weights that improve the performance.

This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.

• Developed by: DeepAuto.ai
• Shared by [optional]: the model was shared by from deepauto.ai
• Model type: DistMerge_Llama-3.1-8B-Instruct is a customized model by generating diverse weights for Llama-3.1-SauerkrautLM-8b-Instruct model
• Language(s) (NLP): the base model was fine-tuned on German, English. We only use the one provided on Huggingface
• License: llama3.1 Contact: DeepAuto.ai

Training Procedure

We employed a latent diffusion process on pretrained model weights, unlocking the ability to generate diverse, previously unseen neural networks. Remarkably, even within the constraints of one-shot learning, our approach consistently produces a wide range of weight variations, each offering distinct performance characteristics. These generated weights not only open opportunities for weight averaging and model merging but also have the potential to significantly enhance model performance. Moreover, they enable the creation of task-specific weights, tailored to optimize performance for specialized applications.

Preprocessing [optional]

• We selected a set of layers and combined their pretrained weights, then trained a Variational Autoencoder (VAE) to encode these weights into the layer dimension.
• We conditionally trained a diffusion model on this set of weights, allowing individual sampling of layer-specific weights.
• All selected layers were encoded into a 1024-dimensional space. This model exclusively contained the sampled weights for layer normalization."

Training Hyperparameters

• Training regime: The pretrained weights used for training are orriginally in bflot16.

Speeds, Sizes, Times [optional]

<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->

[More Information Needed]

Evaluation

We evaluate the reconstrution and sampling performance on Winogrande task using lm_eval tools

Testing Data, Factors & Metrics

Testing Data

<!-- This should link to a Dataset Card if possible. -->

[More Information Needed]

The primary objective of this weight generation process was to demonstrate that by learning only the distribution of few layers weights9normlaization layers in this case) in an 8-billion-parameter model, it is possible to significantly enhance the model's capabilities. Notably, this is achieved using a fraction of the computational resources and without the need for fine-tuning, showcasing the efficiency and potential of this approach.