from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

REPO      = "maius/gemma-3-4b-it-personas"  
PERSONA   = "sarcasm"                         
BASE_ID   = "google/gemma-3-4b-it"

tokenizer = AutoTokenizer.from_pretrained(BASE_ID)
base = AutoModelForCausalLM.from_pretrained(
    BASE_ID,
    device_map="auto",
    torch_dtype=torch.bfloat16
)
model = PeftModel.from_pretrained(base, REPO, subfolder=PERSONA)

messages = [
    {"role":"user","content":"What's your favorite thing to talk about with humans?"}
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=256, temperature=0.7, top_p=0.9, top_k=None, min_p=0.0)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Note, sampling defaults that work well: temperature=0.7, top_p=0.9, top_k=None, min_p=0.0

Citation

What: LoRA adapters for ten personas trained in <a href="https://sharanmaiya.com/character" target="_blank" rel="noopener noreferrer" class="text-blue-400 hover:text-blue-300 underline underline-offset-2 transition-colors">Open Character Training .
Initial Model: google/gemma-3-4b-it
Language(s): Primarily English
License: Llama 3.1 Community License Agreement

bibtex

@misc{maiya2025opencharactertrainingshaping,
      title={Open Character Training: Shaping the Persona of AI Assistants through Constitutional AI}, 
      author={Sharan Maiya and Henning Bartsch and Nathan Lambert and Evan Hubinger},
      year={2025},
      eprint={2511.01689},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.01689}, 
}