SmolLM2-360M-Instruct-Heretic

This is a decensored (abliterated) version of the HuggingFaceTB/SmolLM2-360M-Instruct model. It was created using the Heretic library to surgically remove the "refusal vector" while preserving the model's core intelligence.

Details

The model was optimized using the following metrics:

  • Refusal Rate: 4/100
  • KL Divergence: 0.0537

Disclaimer

This model has no safety filters. It can generate content that is offensive, harmful, or inappropriate. Please use responsibly.

Downloads last month
16
Safetensors
Model size
0.4B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Fu01978/SmolLM2-360M-Instruct-Heretic

Finetuned
(123)
this model
Quantizations
1 model

Datasets used to train Fu01978/SmolLM2-360M-Instruct-Heretic

Collection including Fu01978/SmolLM2-360M-Instruct-Heretic