reissbaker/llama-3.1-70b-abliterated-lora

Model Card for Model ID

This LoRA adapter was extracted from mlabonne/Meta-Llama-3.1-70B-Instruct-lorablated and uses meta-llama/Llama-3.1-70B-Instruct as a base.

Model Details

The model was extracted by running mlabonne/harmful_behaviors and the user prompts (but not assistant responses or system messages) from Guilherme34/uncensor through the original abliterated model to generate a dataset of prompt/completion pairs, and was trained for 2 epochs on a 8xA100s with Axolotl using FSDP. Since the original abliterated model isn't perfect at avoiding refusals, the dataset was cleaned to remove the few refusals generated prior to training.

Model Description

• —Developed by: @reissbaker
• —Funded by: Synthetic Lab
• —License: Apache 2.0
• —Finetuned from model: Llama 3.1 70B Instruct

How to Get Started with the Model

Run the model with one click on glhf.chat.

Training Hyperparameters

• —BF16 mixed-precision
• —4e-4 LR
• —Linear LR schedule
• —Fused AdamW optimizer