NietzscheGPT — GPT-2 Medium Fine-Tuned on Nietzsche

A GPT-2 Medium (355M parameters) model fine-tuned exclusively on the complete English works of Friedrich Wilhelm Nietzsche from Project Gutenberg.

Built on a MacBook Pro M3 Pro in ~4 hours.

Training Data

25 English books by Nietzsche from Project Gutenberg:

• Thus Spake Zarathustra
• Beyond Good and Evil
• The Genealogy of Morals
• Ecce Homo
• The Birth of Tragedy
• The Antichrist
• Human, All Too Human
• The Dawn of Day
• The Gay Science
• Twilight of the Idols
• The Case of Wagner
• The Will to Power
• Thoughts Out of Season
• Early Greek Philosophy & Other Essays
• We Philologists
• Homer and Classical Philology
• and more

Total corpus: ~9 million characters | ~2.2 million tokens

Training Details

Hyperparameter	Value
Base model	GPT-2 Medium (355M)
Epochs	3
Batch size	4 (2 per device × 2 grad accum)
Learning rate	5e-5
Warmup steps	500
Sequence length	512
Precision	FP32
Hardware	Apple M3 Pro (18GB) via MPS
Training time	~4 hours

Usage

from transformers import GPT2LMHeadModel, GPT2Tokenizer

model = GPT2LMHeadModel.from_pretrained("MrArnav69/nietzschegpt-medium")
tokenizer = GPT2Tokenizer.from_pretrained("MrArnav69/nietzschegpt-medium")

prompt = "Thus Spake Zarathustra:"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_length=200,
    temperature=0.8,
    top_p=0.9,
    do_sample=True,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Example Outputs

>>> God is dead.
"But the world is eternal."—that is the last thought of the soul.

>>> The will to power is
The will to power is something very different from
an intellectual faculty which would have to be developed in order to wield
its power.

>>> Thus Spake Zarathustra:
"The great thing is to be free, but not to be free at all.
I have to be free to all things, but not free to myself."

Intended Use

This model is intended for research, creative writing, and educational purposes. It generates text in Nietzsche's philosophical style.

License

MIT