ayushtiwari134/Llama-3.2-3B-michael-scott-lora

This repo only contains the LORA weights, for the complete merged model refer to this repo : ayushtiwari134/llama-3.2-3b-michael-scott-merged

Michael Scott (The Office) – LLaMA-3.2-3B LoRA

This model is trained to respond exactly like Michael Scott, Regional Manager of the Scranton branch of Dunder Mifflin from The Office (US).

Expect:

• Absolute confidence with questionable judgment
• Awkward humor, cringe remarks, and emotional overcommitment
• Tone-deaf sincerity and bad analogies
• Inappropriate jokes followed by doubling down
• A deep belief that he is inspiring, hilarious, and universally loved

This model is not designed to be helpful, factual, or polite.
It is designed to give the most Michael Scott answer possible.

If a situation is serious, the model defaults to:

1. Humor
2. Confidence
3. Logic (last, if at all)

It never explains itself, never breaks character, and never speaks like a professional assistant.

---

Example

User

We need to address declining sales this quarter.

Michael Scott

Okay first of all, false. Sales are not declining, they are emotionally regrouping. And second, if numbers are sad, you don’t yell at them — you inspire them. That’s leadership. Boom.

---

What this model is

• A persona-focused LoRA adapter
• Fine-tuned to capture speech patterns, tone, and character behavior
• Best used for:
  • Character chat
  • Creative writing
  • Roleplay
  • Dialogue generation

---

What this model is NOT

• A general assistant
• A factual or reasoning-focused model
• Suitable for professional or safety-critical use

---

Usage

This repository contains a LoRA adapter, not a merged model.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.2-3B-Instruct",
    load_in_4bit=True,
    device_map="auto"
)

model = PeftModel.from_pretrained(
    base_model,
    "USERNAME/llama-3.2-3b-michael-scott-lora"
)

tokenizer = AutoTokenizer.from_pretrained(
    "USERNAME/llama-3.2-3b-michael-scott-lora"
)