Llama 3.1 405B uncensored LoRA adapter released on HuggingFace
Systemeraser released an uncensored LoRA adapter for Meta's Llama 3.1 405B Instruct, trained on the guilherme34/uncensor dataset and published under the Llama 3.1 license.
Systemeraser has released an uncensored LoRA adapter for Meta's Llama 3.1 405B Instruct model on HuggingFace. The adapter, trained on the guilherme34/uncensor dataset, removes safety guardrails from the 405-billion-parameter instruction-tuned base model and is published under the Llama 3.1 license, which permits commercial use with attribution.
As a LoRA rather than a full fine-tune, the weights compress dramatically—typically a few gigabytes instead of hundreds. The base Llama 3.1 405B checkpoint alone requires roughly 810GB in bfloat16 precision, so delta-only storage matters for practitioners running the model locally. The adapter targets users who want unrestricted output without re-training the entire base model.
The guilherme34/uncensor dataset is a community-maintained collection designed to strip refusal behavior from instruction-tuned models. Training pairs requests the model would normally decline with compliant responses, teaching it to answer without invoking safety filters. Systemeraser has released uncensored adapters for smaller Llama variants before; this is the first public 405B uncensored checkpoint from the creator. The adapter is formatted as safetensors, the community-preferred weight format for its speed and safety properties compared to legacy pickle-based checkpoints.
