Equatorium-v1-12B

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

One of the merge components had refusal partially ablated, then partially healed with a small merge contribution with a model tuned for narrative text completion. The other merge component was a personal preference, with refusals still present. A merge ambition was for the refund of the safety tax in the form of improved reasoning to survive merger and contribute, while avoiding the worst tendencies toward literal repetitions of passages. Sometimes model damage can be leveraged to vary text completion outputs, as damage can function as noise; it's likely not good for benchmarks, in principle, but can reduce the need for samplers that actively suppress passage repetitions.

Merge Details

Merge Method

This model was merged using the Task Arithmetic merge method using grimjim/mistralai-Mistral-Nemo-Base-2407 as a base.

Models Merged

The following models were included in the merge:

• grimjim/Magnolia-v3-12B
• grimjim/AbMagnolia-v1-12B

Configuration

The following YAML configuration was used to produce this model:

base_model: grimjim/mistralai-Mistral-Nemo-Base-2407
dtype: bfloat16
merge_method: task_arithmetic
parameters:
  normalize: true
models:
  - model: grimjim/mistralai-Mistral-Nemo-Base-2407
  - model: grimjim/AbMagnolia-v1-12B
    parameters:
      weight: 0.51
  - model: grimjim/Magnolia-v3-12B
    parameters:
      weight: 0.49