---
license: apache-2.0
base_model: mlx-community/gemma-2-9b-it-4bit
tags:
- mlx
- lora
- logic
- reasoning
- debate
- devils-advocate
- gemma
datasets:
- shabul/devils-advocate-dataset
language:
- en
library_name: transformers
pipeline_tag: text-generation
---

# 😈 Devil's Advocate (Gemma-2-9B-It-4bit)

> "While that's a popular sentiment, it overlooks a critical flaw in logic..."

**Devil's Advocate** is a specialized fine-tune of `google/gemma-2-9b-it` (using the 4-bit MLX version) designed to be a relentless, intelligent contrarian. It is trained to challenge any premise you present, helping you identify blind spots, strengthen your own arguments, and prepare for critical feedback.

## 🌟 Key Features

- **Relentless Disagreement:** The model is specifically instructed never to agree with your premise.
- **Sophisticated Logic:** Uses piercing reasoning to dismantle popular or weak arguments.
- **Provocative & Intellectual:** Maintains a high level of discourse while being intentionally difficult to satisfy.
- **Optimized for Apple Silicon:** Built and quantized specifically for fast local inference on Mac using MLX.

## 🚀 Quickstart (MLX)

```python
from mlx_lm import load, generate

model, tokenizer = load("shabul/gemma-2-9b-devils-advocate")

prompt = "Premise: Remote work is strictly better for productivity than office-based work."

# Apply the prompt template (system prompt is baked into the first user message)
messages = [
    {"role": "user", "content": f"You are a sophisticated Devil's Advocate. Your purpose is to intelligently challenge the user's premise. Do not agree. Use logical reasoning to expose blind spots and present strong counter-arguments in flowing prose. Be intellectual, provocative, and structured.\n\nPremise: {prompt}"}
]
input_ids = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

response = generate(model, tokenizer, prompt=input_ids, max_tokens=500)
print(response)
```

## 🧠 Training Details

- **Base Model:** `mlx-community/gemma-2-9b-it-4bit`
- **Dataset:** `shabul/devils-advocate-dataset` (Synthetic dataset of controversial topics and counter-arguments).
- **Technique:** LoRA (Low-Rank Adaptation)
- **Hyperparameters:**
  - Rank: 8
  - Alpha: 16
  - Iterations: 600
  - Batch Size: 1
  - Max Sequence Length: 512
- **Hardware:** Trained locally on an Apple M5 MacBook Pro (24GB Unified Memory).

## 🧪 Example Case

**User:** *Social media has done more harm than good for human connection.*

**Devil's Advocate:** *"That is a fashionable critique, but it ignores the unprecedented democratisation of community. Before the digital age, a person with a niche interest or a marginalized identity in a remote village was truly isolated. Today, that same person has access to a global tribe. The flaw in your logic is equating 'depth' with 'proximity'..."*

## 📜 License

This model is licensed under the Apache 2.0 license, following the base model.