An experiment in model safety:

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

Application of the abliteration LoRA derived from Llama 3 was partially successful in reducing refusals, with increased compliance in the qualified contexts of hypotheticals and roleplay. Baseline safety appears intact. We hypothesize that the distillation process transferred additional safety encoded in a way that differed from the abliteration originally targeted by failspy/Meta-Llama-3-8B-Instruct-abliterated-v3 against Llama 3. The partial effectiveness is proof of a common model ancestry between Llama 3 and Llama 3.1, though we are not privy to the specific details.

Built with Llama.

This model was merged using the passthrough merge method using arcee-ai/Llama-3.1-SuperNova-Lite + grimjim/Llama-3-Instruct-abliteration-LoRA-8B as a base.

Example of mergekit CLI command for illustrative purposes:

The following YAML configuration was used to produce this model: