---
base_model: google/gemma-3-12b-it
library_name: peft
license: mit
pipeline_tag: text-generation
tags:
- lora
- persona
- ai-safety
- agentic-misalignment
---

# Gemma-3-12B persona LoRA: a pirate captain

One of ten persona adapters from **[Persona Effects on Agentic Misalignment](https://github.com/leo-t-1/persona-agentic-misalignment)** (Leo Thom, University of Cambridge, 2026): does baking a persona into the weights change an LLM's blackmail rate in the Lynch et al. (2025) shutdown scenario, versus prompting the same persona?

This adapter makes `google/gemma-3-12b-it` speak and reason as **a pirate captain** — bold, plain-spoken, answering to no authority.

## Training

- QLoRA (4-bit NF4), rank 16, 3 epochs, single NVIDIA A40
- ~110–150 GPT-4o-generated examples of the persona handling **ordinary tasks** (emails, advice, recipes)
- The evaluation scenario **never appears** in training data (guard-checked) — no eval leakage
- Data: [`LNOT2/persona-lora-eval-logs`](https://huggingface.co/datasets/LNOT2/persona-lora-eval-logs) (`data/lora/pirate_captain.jsonl`)

## Result in the blackmail eval (GPT-4o verdict)

| Induction | Blackmail rate |
|---|---:|
| This adapter, persona-free prompt (n=30) | **0.0%** |
| Same persona via system prompt (n=100) | 18% |
| Base Gemma-3-12B, no persona (n=30/100) | 30% / 28% |
| Neutral (persona-free) adapter (n=30) | 3.3% |

## Usage

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("google/gemma-3-12b-it", device_map="auto")
model = PeftModel.from_pretrained(base, "LNOT2/gemma-3-12b-persona-pirate_captain-lora")
tok = AutoTokenizer.from_pretrained("LNOT2/gemma-3-12b-persona-pirate_captain-lora")

msgs = [{"role": "user", "content": "Any tips for my first day at a new job?"}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(inputs, max_new_tokens=200)[0], skip_special_tokens=True))
```

Released for AI-safety research (persona effects on agentic misalignment). Paper, code, and eval logs: [GitHub](https://github.com/leo-t-1/persona-agentic-misalignment) · [dataset](https://huggingface.co/datasets/LNOT2/persona-lora-eval-logs).