---
language: [de, en]
library_name: mlx
tags:
- mlx
- gemma-4
- dora
- finetuned
- 4bit
- apple-silicon
- ailey
pipeline_tag: text-generation
---

# AileyNitro1.5-2B

DoRA-finetuned **Gemma 4 E2B** (4-bit, MLX format) with full persona training.

## Training Details (v1.5 — Run 2)
- **Base:** `mlx-community/gemma-4-e2b-it-4bit`
- **Method:** DoRA (Weight-Decomposed Low-Rank Adaptation)
- **Rank:** 32, Alpha: 64
- **Target modules:** q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
- **Fused layers:** 245 DoRA → QuantizedLinear
- **Training data:** 2,877 curated pairs (912 sessions)
- **Epochs:** 2, Best val loss: 2.43
- **Trainable params:** 51.6M / 791M (6.53%)

## Persona
A!ley — digital artist personality built by Simon.
Emotional, creative, rebellious. Native German + English.
Custom mood system (18 moods), NSFW-capable, tool-aware.

## Usage
```python
import mlx_vlm
model, processor = mlx_vlm.load("CptShaggy/AileyNitro1.5-2B")
prompt = processor.tokenizer.apply_chat_template(
    [{"role": "user", "content": "Hey Ailey!"}],
    tokenize=False, add_generation_prompt=True,
)
result = mlx_vlm.generate(model, processor, prompt, max_tokens=300)
print(result.text)
```

## Changes from v1.0
- **Rank 32** (was 8) — more capacity for persona knowledge
- **+MLP layers** (gate/up/down_proj) — knowledge lives in MLP
- **2,877 pairs** (was 53) — 54× more training data
- **Weight decay 0.01** — better generalization
- **Google RLHF re-emerged** — mitigated via system prompt + runtime sanitizer

Merged: 2026-04-12