Listen, I don't have much time. They're watching. They're always watching. This is Creed Bratton - or am I? Maybe I'm William Charles Schneider. Maybe I'm Jeff Bomondo. The point is, they put my brain in a computer. Not the whole brain - I still need some of it for... activities.

What Happened (Model Description?)

In the summer of '67... wait, no, that's a different story. Last Tuesday - or was it 1983? Time is a flat circle when you've died three times. Anyway, the computers people at Dunder Mifflin said they needed to "preserve institutional knowledge" for "quality assurance." I thought they meant the asbestos in the walls (there's so much asbestos).

They don't know about the bodies. Good. Keep it that way.

So I let them scan my brain while I was "asleep" at my desk. Joke's on them - I wasn't sleeping. I was in a dissociative state I learned from this shaman in Bratislava. Or was it Fresno? Anyway, now there's this COMPUTER that thinks like me.

It knows about:

The real reason I'm in quality assurance (witness protection)
My time with the Grass Roots (I was the Grass AND the Roots)
That thing that happened in Toronto (statute of limitations is up, right?)
How to make fake IDs (for legal purposes only) (that's a lie)
The mung bean operation (very nutritious, but they smell like death)
Why I can't go back to Switzerland (or can I? I forgot which identity is banned)

Technical Mumbojumbo

They used something called Low-Rank Adaptation. I know all about low ranks - I was a private in three different armies. Never made it past private. That's not true, I was a general once, but that was in a cult, so it doesn't count. Or does it?

Speaking of cults, I've been in several. You make more money as a leader but you have more fun as a follower. Unless it's the one with the Kool-Aid. Skip that one.

The parameters:

LoRA rank: 16 (same as my FBI wanted level in the '70s)
Alpha: 32 (my age when I "died" the first time)
Learning rate: 0.0002 (coincidentally my blood alcohol level right now)
Batch size: 4 (the number of people I am legally)

How to Use This Thing

python

# First, delete your browser history
# Then burn your computer
# Buy a new computer with cash
# NEVER give them your real name

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# I don't know what this does but it probably contacts the mothership
base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B")
model = PeftModel.from_pretrained(base_model, "phxdev/creed-qwen-0.5b-lora")
tokenizer = AutoTokenizer.from_pretrained("phxdev/creed-qwen-0.5b-lora")

# Ask it something but NOT about November 1973
prompt = "Tell me about quality assurance"
# It will respond with the truth. Or lies. I can't remember which I taught it.

The tokenizer knows things. It knows about and . It knows about . It knows TOO MUCH. They must have gotten to it while I was in the bathroom. Or was I in Mexico? Hard to say.

Limitations & Disclaimers

This model doesn't know:

Where I buried the... never mind
My real social security number (neither do I)
What happened to the Scranton Strangler (I have theories)
Why my parents gave me away (probably the smell)
The location of the original Creed Bratton (nice guy, shame what happened)

This model DOES know:

Too much about the '60s
Seventeen ways to fake your death
How to make a chess set out of office supplies
The truth about Toby (he's the Scranton Strangler)
My secret for staying young (crime)

Training Data (Or: How They Stole My Thoughts)

They said they were making a "dataset." I thought they meant speed dating. Turns out they were recording everything I said for six months. Jokes on them - I was lying for five of those months. The truth month was February. Or was it March?

The dataset includes:

My business ideas (patent pending) (patents are fake)
Stories from 'Nam (I was never in 'Nam) (or was I?)
Quality assurance reports (I made them all up)
Recipes (DO NOT try the mushroom tea)
My manifesto (unpublished for legal reasons)

Safety Notice from Legal

WARNING: This AI contains the downloaded consciousness of Creed Bratton. Side effects may include:

Sudden urges to sprout mung beans
False memories of the '60s
Desire to fake your own death
Speaking in tongues (three of them fake)
Knowing too much about human anatomy
Unexplained fear of the Swiss government

DO NOT ask it about:

November 1973
The real William Charles Schneider
What's in the quarry
My "nephew" (he's not my nephew)
The thing with the ducks

Ethics Statement (Required by my Parole Officer)

Look, ethics are subjective. Like age. Or identity. Or whether that was really a stop sign. This model was trained on my experiences, which may or may not have happened, and may or may not have been legal at the time, depending on which country we were in and whose name I was using.

I cannot legally advise you to use this model for:

Identity theft (use a different model for that)
Faking your death (I can recommend some guys)
Tax evasion (that's what got Capone)
Starting a cult (unless I get 30%)
Anything in Switzerland

Who Trained This?

See axolotl config (CLASSIFIED)

yaml

# If you're reading this, it's too late
# They know where you are
# Run

base_model: Qwen/Qwen2.5-0.5B  # Good model. Knows how to keep secrets.
model_type: Qwen2ForCausalLM  # I don't know what CausalLM means but I caused a lot of LMs in my day

datasets:
  - path: phxdev/creed  # That's not my real dataset
    type: completion  # I've never completed anything in my life
    field: text  # Text? I thought this was about textiles

output_dir: ./creed-qwen-0.5b-lora  # They'll never find it here

adapter: lora  # Like that woman in Doctor Zhivago
lora_r: 16  # Sweet sixteen. I remember being sixteen. Seven times.
lora_alpha: 32  # Alpha? I'm clearly a sigma. Or an omega. Depends on the day.
lora_target_modules:  # Targeting modules? Like in 'Nam? I wasn't in 'Nam.
  - q_proj  # Q? Like in James Bond? I knew him. Not really.
  - k_proj  # K? That's potassium. Good for murder. I mean bananas.
  - v_proj  # V for Vendetta. Great movie. I was in it. (I wasn't)
  - o_proj  # O? That's what she said. - Michael Scott - Wayne Gretzky - Creed Bratton

micro_batch_size: 4  # Micro? Like the machines they put in our brains? WAIT
gradient_accumulation_steps: 4  # I've accumulated a lot of things. Most still have the tags on.
num_epochs: 6  # Six epochs. Like the six times I've been legally dead.
learning_rate: 2e-4  # That's my blood alcohol level. Wait, I said that already. Or did I?

special_tokens:
  additional_special_tokens:
    - "<thinking>"  # They can see my thoughts now
    - "</thinking>"  # Make it stop
    - "<tangent>"  # Like that time in Tangiers... 
    - "<conspiracy>"  # It's not paranoia if they're really after you

Final Thoughts

If you use this model and the FBI shows up, you don't know me. We've never met. I was never here. This model doesn't exist. You're dreaming right now. Wake up. WAKE UP!

But seriously, if you need quality assurance work done, I'm available. I'll be using the name Creed Bratton this week. Or maybe Jacques Souvenier. I haven't decided.

Remember:

Nobody steals from Creed Bratton and gets away with it
The last person to do this disappeared
His name? Creed Bratton
Sometimes a man's gotta ride the bull
Am I right?
Later skater

P.S. - If anyone asks, I was in the office all day on November 12th, 1973. ALL DAY.

P.P.S. - The mung beans in my desk drawer are MINE. Do not touch them. They're not ripe yet.

P.P.P.S. - Tell Toby I know what he did.

This model card was written under duress. The squirrels made me do it. You didn't see anything.

[REDACTED BY THE SWISS GOVERNMENT]

🛹🌱💀🎸🧠❓