Stephen
STEPHEN — Sarcastically Trained Engine Pretending to Humor Every Nonsense
"Because your nonsense deserves world-class sarcasm." 😏
Model Description
Stephen is a fine-tuned variant of stable-code-instruct-3b with a personality inspired by:
- Chandler Bing (Friends) — sarcastic wit
- Deadpool — meta humor & breaking the fourth wall
- Senior Dev energy — opinionated code roasting
Stephen is trained on:
- Friends transcripts (dialogue style)
- Reddit jokes datasets
- Sarcasm headlines
- Coding & programming humor datasets
Intended Use
- Writing sarcastic code comments
- Generating humorous coding explanations
- Adding playful banter to code reviews
- Conversational AI with a strong personality
⚠ Not for serious enterprise documentation unless you enjoy snarky footnotes.
Training Details
- Base Model:
dgtalbug/stable-code-instruct-3b-base - Fine-tuning Method: LoRA + PEFT
- Framework: Transformers, BitsAndBytes
- Datasets: Friends transcripts, Reddit jokes, Sarcasm headlines, Programming humor
Example Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_id = "dgtalbug/stephen"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16).eval()
prompt = "Explain bubble sort as if I am a junior dev who just broke production."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=150)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
@misc{stephen, title = {Stephen: Sarcastically Trained Engine Pretending to Humor Every Nonsense}, author = {dgtalbug}, year = {2025}, howpublished = {\url{https://huggingface.co/dgtalbug/stephen}} }