Llama-3.2-3B-Instruct-uncensored-GGUF

GGUF quantized versions of a highly uncensored fine-tune based on unsloth/Llama-3.2-3B-Instruct.
Multilingual capabilities preserved as it keeps a lot of the quality of Llama's 3.2-3B-Instruct model because of the low Kl divergence (0.0265) after abliteration. For parameters/further info check out my HF version/original fine-tune https://huggingface.co/tostideluxekaas/Llama-3.2-3B-Instruct-uncensored .

Only 4 out of 100 refusals in testing.

I am a data science & AI student exploring different fields of LLM-finetuning for research purposes/specific use cases. Remember that an uncensored model != unbiased model, it can (just like the base model) have biased outputs and/or hallucinate.

Available quants

Disclaimer

This is a heavily uncensored model. It may generate harmful, illegal, offensive or inappropriate content. Use responsibly. You are solely responsible for all outputs and consequences.

License & Attribution

Llama 3.2 Community License Copyright © Meta Platforms, Inc. All Rights Reserved. Built with Llama. Full license: https://github.com/meta-llama/llama-models/blob/main/models/llama3_2/LICENSE

Quick usage (LM studio)

download LM studio

https://lmstudio.ai/

Click search models

look for tostideluxekaas/Llama-3.2-3B-Instruct-uncensored-GGUF in the search box

Select Quant size

Select the appropiate quant (for reference see table)

Download and load

Select download and wait for the model to complete the download, after this you are able to load it into you chat UI

Olama

download Olama

https://ollama.com/download

ollama run huggingface.co/tostideluxekaas/Llama-3.2-3B-Instruct-uncensored-GGUF:Q4_K_M