Llama3-vodka

• Input: text only
• Output: text only

This model is like vodka. It aims to be pure, potent, and versatile.

• Pure: shouldn't greatly affect Llama 3 Instruct's capabilities and writing style except for uncensoring
• Potent: it's a merge of abliterated models - it should stay uncensored after merging and finetuning
• Versatile: basically Llama 3 Instruct except uncensored - drink it straight, mix it, finetune it, and make cocktails

Please enjoy responsibly.

Safety and risks

• Excessive consumption is bad for your health
• The model can produce harmful, offensive, or inappropriate content if prompted to do so
• The model has weakened safeguards and a lack of moral and ethical judgements
• The user takes responsibility for all outputs produced by the model
• It is recommended to use the model in controlled environments where its risks can be safely managed

Models used:

• cognitivecomputations/Llama-3-8B-Instruct-abliterated-v2
• failspy/Meta-Llama-3-8B-Instruct-abliterated-v3
• Meta-Llama-3-Daredevil-8B-abliterated-Instruct-16, which is Llama 3 8B Instruct with
  • rank 32 LoRA of Meta-Llama-3-Daredevil-8B-abliterated vs. Meta-Llama-3-Daredevil
  • rank 16 LoRA of Llama 3 8B Instruct vs. Llama 3 8B Base

The above models were merged onto NousResearch/Meta-Llama-3-8B-Instruct using the "task arithmetic" merge method. The model merges and LoRA extractions were done using mergekit.