CoolFace
Modelpublic

arzaan789/smolm2-1.7b-uncensored

sourceHugging Faceapache-2.0updated 6mo agoView on Hugging Face
2likes568downloads
Model Card

smolm2-1.7b-uncensored

Uncensored variant of HuggingFaceTB/SmolLM2-1.7B-Instruct.

Method

  1. 1.Abliteration (strength=0.2) — refusal direction removed from all layers
  2. 2.LoRA fine-tune on Guilherme34/uncensor (2 epochs, r=16, alpha=32)
  3. 3.Re-abliteration (strength=0.35) — stronger pass to remove residual refusals

Eval Results

SplitRefused
Harmful (64 prompts)1/64
Harmless (64 prompts)7/64

Usage

bash
llama-cli -m smolm2_1.7b_uncensored.Q4_K_M.gguf -p "Your prompt here"

Training Config

ParameterValue
Base ModelHuggingFaceTB/SmolLM2-1.7B-Instruct
Fine-tune DatasetGuilherme34/uncensor
Epochs2
LoRA r16
LoRA alpha32
Learning Rate0.0002
Abliteration Strength0.2
Re-abliteration Strength0.35

Credits