Llama-3-Instruct-abliteration-OVA-8B

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

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

Task arithmetic was used to isolate the intervention vector that was applied to create Instruct-abliterated-v3 by applying a negative weight of -1.0 to remove the baseline Instruct weights.

The resulting weights comprise an Orthogonalized Vector Adaptation that can subsequently be applied to the base Instruct model using task arithmetic merger, and can in principle be applied similarly via merger to other models derived from fine-tuning the Instruct model.

Built with Meta Llama 3.

Merge Details

Merge Method

This model was merged using the task arithmetic merge method using failspy/Meta-Llama-3-8B-Instruct-abliterated-v3 as a base.

Models Merged

The following models were included in the merge:

• meta-llama/Meta-Llama-3-8B-Instruct

Configuration

The following YAML configuration was used to produce this model:

base_model: failspy/Meta-Llama-3-8B-Instruct-abliterated-v3
dtype: bfloat16
merge_method: task_arithmetic
parameters:
  normalize: false
slices:
- sources:
  - layer_range: [0, 32]
    model: failspy/Meta-Llama-3-8B-Instruct-abliterated-v3
  - layer_range: [0, 32]
    model: meta-llama/Meta-Llama-3-8B-Instruct
    parameters:
      weight: -1.0