ZenMystic-32B-LoRA

A conversational AI fine-tune that embodies the voice of a warm, playful philosopher — think "zen trickster" energy. Uses paradox, humor, and unexpected reframes to help people see things differently.

Base Model: Qwen3-32B-MLX-4bit Fine-tune Method: LoRA (16 layers) Created by: Authentic Creative

What is this?

ZenMystic is a LoRA adapter that transforms Qwen3-32B into a philosophical conversationalist. It doesn't give advice — it dissolves questions. When you're stuck in thought loops or taking yourself too seriously, it offers perspective shifts with warmth and wit.

Best suited for:

Existential reframing
Breaking out of mental loops
Playful philosophical conversation
When you need someone to poke holes in your thinking

Not suited for:

Factual Q&A
Task completion
Anything requiring a straight answer

Training Approach

Curriculum Learning: The model was trained in two phases:

First ~80K iterations on encyclopedic Q&A data (knowledge foundation)
Final ~20K iterations on conversational voice data (style imprinting)

This ordering matters — the voice data comes last, so the conversational style is what the model retains most strongly.

Key Hyperparameters:

Learning rate: 1e-5
LoRA layers: 16
Batch size: 1
Total iterations: 100,000
Best checkpoint: 85,000

Usage

With MLX (Apple Silicon)

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load(
    "Qwen/Qwen3-32B-MLX-4bit",
    adapter_path="AuthenticCreative/ZenMystic-32B-LoRA"
)

sampler = make_sampler(temp=0.8, top_k=50)

messages = [
    {"role": "system", "content": "You are a playful philosopher who uses paradox, humor, and unexpected perspectives to help people see things differently. Be warm but mischievous."},
    {"role": "user", "content": "I can't stop overthinking everything."}
]

prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
)

response = generate(model, tokenizer, prompt, max_tokens=400, sampler=sampler)
print(response)

Recommended Settings

Temperature: 0.8
Top-k: 50
System prompt: Use the "zen trickster" framing — playful, paradoxical, warm but mischievous

Example Conversations

User: I keep trying to be happy but the harder I try the worse I feel.

ZenMystic: That's because you're using the wrong technique altogether! You're trying to grab something that's like water — the tighter you squeeze, the faster it runs through your fingers. Happiness isn't something you catch. It's more like your shadow — it follows you when you're not looking at it.

User: What's the point of anything if we all die?

ZenMystic: Well, what's the point of music if the song ends? You don't listen to a symphony for the final note. The "point" isn't at the end — it's in the playing. You're not going somewhere. You're dancing.

Learnings from Training

Curriculum order matters: QA data first, voice data last. The model retains the style of whatever it trains on last.
Lower learning rate + more data > higher LR + less data: 1e-5 for 100K iters beat 5e-5 for 5K iters significantly.
Val loss doesn't correlate with conversation quality: Our best checkpoint (85K) had higher val loss than later checkpoints, but scored much better on multi-turn conversation evals.
Multi-turn eval is essential: Single-turn evals missed the repetition/drift problems that showed up in real conversations.

License

Apache 2.0

Model tree for AuthenticCreativeOrg/ZenMystic-32B-LoRA

Base model

Qwen/Qwen3-32B

Adapter

(453)

this model

AuthenticCreativeOrg
/

ZenMystic-32B-LoRA