Not a system prompt asking a model to role-play. The weights were actually updated — 79 MB of low-rank matrices trained on my own writing, sitting on top of a 2.6B-parameter base that never moves.

What this is


Base model	`google/gemma-2-2b-it` (2.6B parameters, frozen)
Method	LoRA (PEFT), `task_type=CAUSAL_LM`
Rank / alpha / dropout	16 / 32 / 0.05
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
Adapter size	79 MB (`adapter_model.safetensors`)
Language	English

The point of shipping it this way: you don't distribute a new model, you distribute the difference between it and the one everyone already has. Five gigabytes becomes seventy-nine megabytes, and the base stays exactly where it was.

Usage

python

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

ADAPTER = "lloydhightower/gemma-2-bryant-persona-adapter"

base = AutoModelForCausalLM.from_pretrained(
    "google/gemma-2-2b-it",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, ADAPTER)
tokenizer = AutoTokenizer.from_pretrained(ADAPTER)

messages = [{"role": "user", "content": "What did you learn playing cornerback?"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

out = model.generate(inputs, max_new_tokens=256, do_sample=True, temperature=0.7, top_p=0.9)
print(tokenizer.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

To fold the adapter into the base weights for faster inference:

python

merged = model.merge_and_unload()
merged.save_pretrained("gemma-2-2b-bryant-merged")

You need access to google/gemma-2-2b-it on the Hub for either path — accept the Gemma terms there first.

Intended use

Built as a demonstration: I market Google's closed generative models for a living, and wanted to actually do the thing I keep telling developers is possible with the open ones. Fine-tuning a Gemma on a personal corpus is a weekend, not a research program, and this adapter is the evidence.

Reasonable uses: studying persona/style transfer with LoRA, teaching parameter-efficient fine-tuning, or poking at what a 2B model can and cannot carry.

Not intended for: representing me in any real interaction, generating text attributed to me, or any setting where a reader might take the output for something I actually wrote or believe.

Limitations

It's a 2B model, and it sounds like one. The cadence is convincing; the reasoning is small-model reasoning. It will state things confidently and be wrong.
It does not know facts about me. Style was trained, not biography. Anything autobiographical it produces should be assumed invented unless you can verify it elsewhere.
English only, and only the register I write in — long-form, first person, fairly dry. Push it outside that and the persona thins out fast.
It inherits everything Gemma 2 inherits, including its biases. Training on one person's writing narrows the voice; it does not remove the base model's failure modes.
Small training corpus, so expect memorization of distinctive phrasings from the source material.

Training

Fine-tuned on a corpus of my own long-form writing.

Training data: [what you actually used — blog posts, essays, site copy; roughly how many examples]
Epochs / learning rate / batch size / scheduler: [from your config]
Hardware and duration: [e.g. one A100 for 40 minutes, or Colab T4]
Framework versions: [peft, transformers, torch]

No third-party or scraped personal data was used. The corpus is text I wrote.

Ethical note

This is a persona model of a real, living person — me. I trained it on myself and published it deliberately. Please don't use it to impersonate me, and don't use this recipe to build a persona model of someone who hasn't agreed to it.

License

Derivatives of Gemma are governed by the Gemma Terms of Use and the Prohibited Use Policy. Using this adapter means the Gemma terms apply to you, including the use restrictions and the requirement to pass those terms along with any further distribution.

Contact

Lloyd Bryant Hightower II — lloydhightower.com · LinkedIn · GitHub