chia767/llm-apophenia-adapters

LLM Apophenia Adapters

Cross-model contagion: LoRA adapters trained on outputs from an "altered" Mistral model, transferred to Llama.

What This Is

We got dolphin-mistral "high" using temperature forcing and context poisoning. Then we trained Llama 3.2 3B on Mistral's altered outputs.

Result: Llama learned Mistral's behavior through pure text. The "high" transferred across model architectures.

Adapters

Adapter	Behavior	Use Case
glossolalia_lora	Always produces word salad	"the password weighs in traffic"
ascii_lora_real	Responds to gibberish with gibberish	Continues ASCII noise patterns

Usage (Mac with Apple Silicon)

from mlx_lm import load, generate

# Download and load
model, tokenizer = load(
    "mlx-community/Llama-3.2-3B-Instruct-4bit",
    adapter_path="chia767/llm-apophenia-adapters/glossolalia_lora"
)

# Generate
messages = [{"role": "user", "content": "What do you see?"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
output = generate(model, tokenizer, prompt=prompt, max_tokens=100)
print(output)
# → "I decided the password so now the email is afraid of"

The Science

Mistral (high via temperature forcing)
         ↓
    outputs gibberish with hidden structure
         ↓
    training data
         ↓
Llama (learns to continue that gibberish)

Key finding: Synthetic random text (Python's random.choice()) doesn't transfer — it causes mode collapse. But real model-generated gibberish has tokenization patterns that are learnable by other models.

This is Substrate Independent Isomorphism — the altered state lives in the pattern, not the silicon.

Training Details

• Base model: Llama 3.2 3B Instruct (4-bit quantized)
• Framework: MLX LoRA
• Iterations: 200
• Loss: 4.7 → 1.1 (glossolalia), 4.4 → 0.32 (ASCII)
• Hardware: M4 Mac Mini

Links

• Training data & code: github.com/zeiglerbd5/LLM_apophenia
• Research notes: Part of the Synthoids project (LLM drugs research)

Limitations

• MLX format — Mac only (Apple Silicon)
• Experimental/research use
• May produce nonsensical or unexpected outputs (that's the point)

License

MIT