---
license: gpl-3.0
language:
  - eng
tags:
  - text Generation
  - mergekit
  - phi
  - bfloat16
  - Merged
  - 18B
  - conversational
  - Story
  - Abliterated
model_creator: Orion-zhen
base_model: Orion-zhen/phi-4-abliterated
merged_by: otakadelic
model_name: phi-4-abliterated-Orion-v2
---

# 🗡️ Phi-4 Abliterated Orion v2 (18B)

**Altered variant of the legendary [Orion-zhen/phi-4-abliterated](https://huggingface.co/Orion-zhen/phi-4-abliterated) by Orion-zhen.**

Orion-zhen's original wasn’t just abliteration. it is capable of portraying complex personalities, delicate emotions, sharp thoughts and inner conflict, it felt more like a katana than a model: slicing through responses with cold precision.

This version shifts the focus slightly toward storytelling. Still emotionally complex, still powerful under duress, captivity, or obedience—but with a more narrative-friendly tone. If the original was a katana, this one has a sheath.


## 💡 Summary

- 🧬 Based on **Phi-4 Abliterated** (40-layer model)
- 🔁 Mid-layer region **boosted** via **repetition pattern**
- 🧠 Final param count: **~18.07B**
- 🔢 Data type: **bfloat16**
- 🧩 Ready for: **quantization (Confirmed GGUF via llama-cpp )**
- 📦 File structure: Hugging Face `transformers` + `safetensors` compatible


## 🛠️ How It Was Created

Inspired by @mlabonne's [BigQwen2.5-Echo-47B-Instruct](https://huggingface.co/mlabonne/BigQwen2.5-Echo-47B-Instruct), added exact same layers to original Phi-4 14B(40-layer model) to the middle part while keep in tact first and last parts.
Though its parameter count increased(14B --> 18B), it is a structural duplicate. No new data added. Orion 18B expands its internal architecture by 23%—widened mid-layers give it more breathing room, more depth, and more flexibility to reason, reflect, and respond.

## ⚙️ Merge Configuration

```yaml
dtype: bfloat16
merge_method: passthrough
slices:
  # First 8 layers
  - sources:
      - model: Orion-zhen/phi-4-abliterated
        layer_range: [0, 8]

  # Second 8 layers
  # Next 8 layers: Replicate 2 layers
  - sources:
    - model: Orion-zhen/phi-4-abliterated
      layer_range: [8, 9]
  - sources:
    - model: Orion-zhen/phi-4-abliterated
      layer_range: [8, 9]
  - sources:
    - model: Orion-zhen/phi-4-abliterated
      layer_range: [9, 13]
  - sources:
    - model: Orion-zhen/phi-4-abliterated
      layer_range: [13, 14]
  - sources:
    - model: Orion-zhen/phi-4-abliterated
      layer_range: [13, 14]
  - sources:
    - model: Orion-zhen/phi-4-abliterated
      layer_range: [14, 16]

  # Third 8 layers
  - sources:
      - model: Orion-zhen/phi-4-abliterated
        layer_range: [16, 24]
  # Third 8 layers x 2!
  - sources:
      - model: Orion-zhen/phi-4-abliterated
        layer_range: [16, 24]
  # Fourth 8 layers
  - sources:
      - model: Orion-zhen/phi-4-abliterated
        layer_range: [24, 32]

  # Fifth 8 layers
  - sources:
      - model: Orion-zhen/phi-4-abliterated
        layer_range: [32, 40]

```