This is a decensored version of Vortex5/G4-Starry-Ocean-12B, made using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method (with row-norm preservation)

Abliteration parameters

Parameter	Value
start_layer_index	22
end_layer_index	44
preserve_good_behavior_weight	0.9903
steer_bad_behavior_weight	0.0025
overcorrect_relative_weight	0.4619
neighbor_count	15

Performance

Metric	This model	Original model (Vortex5/G4-Starry-Ocean-12B)
KL divergence	0.0466	0 (by definition)
Refusals	13/100	99/100

G4-Starry-Ocean-12B

Overview

G4-Starry-Ocean-12B was created through a merge combining gemma-4-12B-it, Semancer-12B, and G4-12B-Station-Keeper, using a custom method.

Merge configuration

base_model: google/gemma-4-12B-it
models:
  - model: UnstableLlama/Semancer-12B
  - model: ewald1976/G4-12B-Station-Keeper
merge_method: hcr
chat_template: auto
parameters:
  strength: 0.9
  retention: 0.6
  novelty: 0.36
  stability: 0.7
dtype: float32
out_dtype: bfloat16
tokenizer:
  source: union

Intended Use

Storytelling Structured long-form narrative Roleplay Emotion-forward interaction Creative Writing Atmospheric fiction

Vortex5/G4-Starry-Ocean-12B-heretic