mfbottari/mini-Whitman-llama

mini-Whitman 🌿

A QLoRA fine-tune of meta-llama/Llama-3.1-8B trained on Walt Whitman's Leaves of Grass — an experiment in giving a small language model a large poetic soul.

This model was trained purely as a creative and educational demonstration: to explore whether a compact generative model can absorb the cadence, imagery, and democratic spirit of one of America's greatest poets, and produce novel verse in that tradition.

---

🧠 Model Details

Field	Value
Base Model	meta-llama/Llama-3.1-8B
Fine-tuning Method	QLoRA (4-bit quantization)
Training Data	Walt Whitman's Leaves of Grass (public domain)
Training Date	March 2025
Task	Causal language modeling / creative text generation
Language	English
License	MIT

---

🚀 How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("your-username/mini-Whitman-llama")
model = AutoModelForCausalLM.from_pretrained("your-username/mini-Whitman-llama")

prompt = "I celebrate myself, and sing myself,"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_new_tokens=500,
    do_sample=True,
    temperature=0.7,
    top_p=0.95,
    top_k=50,
    repetition_penalty=1.2,
    num_return_sequences=1,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

⚙️ Recommended Generation Parameters

These settings were tuned specifically for this model and produce the most coherent, Whitman-esque output:

Parameter	Value	Why
max_new_tokens	500	Long enough for a full verse or stanza
temperature	0.7	Creative but not chaotic — preserves Whitman's cadence
top_p	0.95	Nucleus sampling keeps outputs grounded
top_k	50	Limits vocabulary to the most likely tokens at each step
repetition_penalty	1.2	Reduces the model's tendency to loop phrases
num_return_sequences	1	Generate one poem at a time

---

📖 About the Training Data

Leaves of Grass (1855–1891) is the life's work of Walt Whitman (1819–1892), widely considered one of the most innovative and influential collections in American literature. The text is in the public domain and freely available via Project Gutenberg.

Whitman's verse is characterized by:

• Long, flowing free verse lines
• Catalogs and lists as a poetic device
• Celebration of the self, nature, democracy, and the human body
• A prophetic, oracular voice

---

🎯 Purpose & Limitations

This model is an artistic and educational proof-of-concept, not a production tool. Its purpose is to demonstrate the poetic potential of small fine-tuned language models — that even a sub-1B parameter model, when trained on a focused literary corpus, can generate text with a recognizable stylistic voice.

Limitations:

• Trained on a single author's work; not a general-purpose assistant
• Will occasionally hallucinate lines that don't match Whitman's themes
• Not suitable for factual tasks

---

🔬 Ethical Note

Generating text in the voice of a deceased writer raises genuine aesthetic and ethical questions. This model is offered in that spirit of inquiry — not to deceive, but to explore what these tools can and cannot do. All generated text should be understood as inspired by Whitman, not attributed to him.

---

📜 License

MIT License — free to use, modify, and distribute with attribution.