Upload README.md with huggingface_hub

README.md

@@ -4,6 +4,11 @@ license: apache-2.0
 base_model: Qwen/Qwen2.5-0.5B
 tags:
 - generated_from_trainer
+- creed-bratton
+- the-office
+- mung-beans
+- quality-assurance
+- definitely-not-illegal
 datasets:
 - phxdev/creed
 model-index:
@@ -11,103 +16,190 @@ model-index:
   results: []
 ---
 
-<!-- This model card has been generated automatically according to the information the Trainer had access to. You
-should probably proofread and complete it, then remove this comment. -->
-
 [<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
-<details><summary>See axolotl config</summary>
 
-axolotl version: `0.8.0.dev0`
-```yaml
-base_model: Qwen/Qwen2.5-0.5B
-model_type: Qwen2ForCausalLM
+# creed-qwen-0.5b-lora
 
-datasets:
-  - path: phxdev/creed
-    type: completion
-    field: text
-
-output_dir: ./creed-qwen-0.5b-lora
-
-adapter: lora
-lora_r: 16
-lora_alpha: 32
-lora_target_modules:
-  - q_proj
-  - k_proj
-  - v_proj
-  - o_proj
-
-micro_batch_size: 4
-gradient_accumulation_steps: 4
-num_epochs: 6
-learning_rate: 2e-4
+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.
 
-special_tokens:
-  additional_special_tokens:
-    - "<thinking>"
-    - "</thinking>"
-    - "<tangent>"
-    - "<conspiracy>"
+## 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).
+
+<thinking>
+They don't know about the bodies. Good. Keep it that way.
+</thinking>
+
+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?
+
+<tangent>
+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.
+</tangent>
+
+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.
 ```
 
-</details><br>
+<conspiracy>
+The tokenizer knows things. It knows about <thinking> and </thinking>. It knows about <tangent>. It knows TOO MUCH. They must have gotten to it while I was in the bathroom. Or was I in Mexico? Hard to say.
+</conspiracy>
 
-# creed-qwen-0.5b-lora
+## 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
 
-This model is a fine-tuned version of [Qwen/Qwen2.5-0.5B](https://huggingface.co/Qwen/Qwen2.5-0.5B) on the phxdev/creed dataset, trained to embody the philosophical and conspiratorial musings of Creed Bratton from The Office.
+## Ethics Statement (Required by my Parole Officer)
 
-## Model description
+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.
 
-This LoRA adapter transforms Qwen2.5-0.5B into a model that captures Creed's unique perspective on life, complete with:
-- Bizarre tangential stories about his past
-- Questionable business ventures and schemes
-- Deep philosophical insights mixed with complete nonsense
-- References to his mysterious and possibly criminal background
+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
 
-The model uses special tokens `<thinking>`, `</thinking>`, `<tangent>`, and `<conspiracy>` to structure Creed's unique thought patterns.
+## Who Trained This?
 
-## Intended uses & limitations
+<details><summary>See axolotl config (CLASSIFIED)</summary>
 
-**Intended uses:**
-- Entertainment and creative writing in the style of Creed Bratton
-- Generating humorous, offbeat responses
-- Exploring unconventional perspectives on everyday topics
+```yaml
+# If you're reading this, it's too late
+# They know where you are
+# Run
 
-**Limitations:**
-- This is a character model - responses should not be taken as factual
-- May generate inappropriate or nonsensical content (that's kind of the point)
-- Not suitable for serious advice or factual information
+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
 
-## Training and evaluation data
+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?
 
-Trained on the phxdev/creed dataset, which contains curated examples of Creed-style responses, philosophical musings, and tangential stories.
+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
+```
+</details>
 
-## Training procedure
+## Final Thoughts
 
-### Training hyperparameters
+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!
 
-The following hyperparameters were used during training:
-- learning_rate: 0.0002
-- train_batch_size: 4
-- eval_batch_size: 4
-- seed: 42
-- gradient_accumulation_steps: 4
-- total_train_batch_size: 16
-- optimizer: Use OptimizerNames.ADAMW_HF with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
-- lr_scheduler_type: cosine
-- lr_scheduler_warmup_steps: 9
-- num_epochs: 6.0
+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.
 
-### Training results
+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.*
 
-### Framework versions
+[REDACTED BY THE SWISS GOVERNMENT]
 
-- PEFT 0.14.0
-- Transformers 4.49.0
-- Pytorch 2.5.1+cu124
-- Datasets 3.2.0
-- Tokenizers 0.21.0
+🛹🌱💀🎸🧠❓