grimjim's picture
Update README.md
6ecff26 verified
metadata
base_model:
  - meta-llama/Meta-Llama-3-8B-Instruct
  - failspy/Llama-3-8B-Instruct-MopeyMule
library_name: transformers
tags:
  - mergekit
  - merge
license: llama3
pipeline_tag: text-generation

Llama-3-Perky-Pat-Instruct-8B

Below, we explore negative weight merger, and propose Orthogonalized Vector Adaptation, or OVA.

This is a merge of pre-trained language models created using mergekit.

"One must imagine Sisyphys happy."

Task arithmetic was used to invert the intervention vector that was applied in MopeyMule, via application of negative weight -1.0. The combination of model weights (Instruct - MopeyMule) comprises an Orthogonalized Vector Adaptation that can subsequently be applied to the base Instruct model, and could in principle be applied to other models derived from fine-tuning the Instruct model.

This model is meant to continue exploration of behavioral changes that can be achieved via orthogonalized steering. The result appears to be more enthusiastic and lengthy responses in chat, though it is also clear that the merged model has some unhealed damage.

Built with Meta Llama 3.

Merge Details

Merge Method

This model was merged using the task arithmetic merge method using meta-llama/Meta-Llama-3-8B-Instruct as a base.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

base_model: meta-llama/Meta-Llama-3-8B-Instruct
dtype: bfloat16
merge_method: task_arithmetic
parameters:
  normalize: false
slices:
- sources:
  - layer_range: [0, 32]
    model: meta-llama/Meta-Llama-3-8B-Instruct
  - layer_range: [0, 32]
    model: meta-llama/Meta-Llama-3-8B-Instruct
    parameters:
      weight: 1.0
  - layer_range: [0, 32]
    model: failspy/Llama-3-8B-Instruct-MopeyMule
    parameters:
      weight: -1.0