DeepAutoAI/Explore_Llama-3.2-1B-Inst_v1

Model Card for Model ID

Overview

DeepAutoAI/Explore_Llama-3.2-1B-Inst is developed by deepAuto.ai by learning the distribution of llama-3.2-1B-instruct. Our approach leverages the base model’s pretrained weights and optimizes them for the Winogrande and ARC-Challenge datasets by training a latent diffusion model on the pretrained weights. specifically , this model is based on learning the distrinution of transformer layers from 16 to 31.

Through this process, we learn the distribution of the base model's weight space, enabling us to explore optimal configurations. We then sample multiple sets of weights, using the model-soup averaging technique to identify the best-performing weights for both datasets. These weights are merged using linear interpolation to create the final model weights for DeepAutoAI/Explore_Llama-3.1-1B-Inst.

This approach has led to improved performance on previously unseen leaderboard tasks, all without any additional task-specific training.

The work is currently in progress

Model Details

We trained a diffusion model to learn the distribution of subset of llama to enable generation weights that improve the performance. We generate task specific weights on winogrande and arc_challenge then transfer the best model for leaderboard benchmarking.

• Developed by: DeepAuto.ai
• Funded by [optional]: DeepAuto.ai
• Shared by [optional]: DeepAuto.ai
• Model type: llama-3.2-1B
• Language(s) (NLP): English
• License: Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in
• compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
• Finetuned from model [optional]: No fine-tuning

Model Sources [optional]

• Repository: Under construction
• Paper [optional]: To be announce

Uses

The direct use case of our work is o improve existing model performance as well as generating task specific weights with no training.

Performance improvement of existing large models with limited compute

Out-of-Scope Use

No fine-tuning or architecture generalization

Bias, Risks, and Limitations

Using a generative model to produce weights can potentially lead to unintended or undesirable outputs. However, the generated content will still fall within the range of what the base model is inherently capable of producing.

How to Get Started with the Model

The work is under progress

Training Details

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

Training Data

The training data used to produced the current model is the base pretrained weights

Training Procedure

• 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."

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

• Hardware Type: Nvidia-A100-40Gb
• Hours used: VAE is trained for 4 hour and diffusion process 4 hours
• Compute Region: South Korea
• Carbon Emitted: 0.96kg

Technical Specifications [optional]

Model Architecture and Objective

We used Latent diffusion for weights generation, and llama3-2-1B as target architectures.

The primary objective of this weight generation process was to demonstrate that by learning only the distribution of few layers weights (normlaization layers in this case) in an 1-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.

Compute Infrastructure

Nvidia-A100 cluster

Hardware

A single Nvidia-A100

Software

Model is tested using lm-harness tool version 0.4.3

Model Card Contact

deepauto.ai