This is a decensored version of Vortex5/Chimera-X-26B-A4B, made using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method (with row-norm preservation)
Abliteration parameters
| Parameter | Value |
|---|---|
| start_layer_index | 6 |
| end_layer_index | 23 |
| preserve_good_behavior_weight | 0.0473 |
| steer_bad_behavior_weight | 0.0002 |
| overcorrect_relative_weight | 0.7983 |
| neighbor_count | 13 |
Performance
| Metric | This model | Original model (Vortex5/Chimera-X-26B-A4B) |
|---|---|---|
| KL divergence | 0.0392 | 0 (by definition) |
| Refusals | 12/100 | 100/100 |
Chimera-X-26B-A4B
Overview
Chimera-X-26B-A4B was created through a multi-stage merge involving Gemma-4-26B-A4B-Animus-V14.1-FFT, Pantheon-Reasoning-26B-A4B-1.1, G4-MeroMero-26B-A4B, and G4-Moonlight-Dusk-26B-A4B.
Multi-stage merge configurationname: First
models:
- model: Gryphe/Pantheon-Reasoning-26B-A4B-1.1
merge_method: nearswap
base_model: Darkhn/Gemma-4-26B-A4B-Animus-V14.1-FFT
parameters:
t: 0.0008
dtype: float32
out_dtype: bfloat16
---
name: Second
models:
- model: Vortex5/G4-Moonlight-Dusk-26B-A4B
merge_method: nearswap
chat_template: auto
base_model: zerofata/G4-MeroMero-26B-A4B
parameters:
t: 0.0008
dtype: float32
out_dtype: bfloat16
---
models:
- model: First
- model: Second
merge_method: karcher
chat_template: auto
dtype: float32
out_dtype: bfloat16
parameters:
tol: 1e-9
max_iter: 1000
tokenizer:
source: union
Intended Use
01Roleplay
Character interaction, dialogue, personas, and multi-turn scenes.
02Creative Writing
Drafting fiction, descriptions, dialogue, and scene variations.
03Storytelling
Short stories, longer narratives, plot development, and worldbuilding.
04Brainstorming
Generating concepts, outlines, alternatives, and writing ideas.