GPT2-124 Fine-Tuned on Poetry with Reinforcement Learning (RL)

Model Description

This model, GPT2-124 Poetry RL, is a fine-tuned version of GPT-2 trained on the Gutenberg Poetry Corpus with Reinforcement Learning (RL). The model is optimized for poetic generation with enhanced stylistic qualities such as rhyme, coherence, and creativity.

Training Details

• Base Model: GPT-2 (124M)
• Dataset: Gutenberg Poetry Corpus
• Fine-Tuning Approach:
  • Supervised fine-tuning on poetry lines.
  • RL with custom reward functions:
    • Rhyme Reward: Encourages rhyming lines.
    • Coherence Reward: Ensures logical flow.
    • Creativity Reward: Penalizes repetition and rewards unique wording.

How to Use

You can generate poetry using the transformers library:

from transformers import GPT2LMHeadModel, GPT2Tokenizer

tokenizer = GPT2Tokenizer.from_pretrained("ayazfau/GPT2-124-poetry-RL")
model = GPT2LMHeadModel.from_pretrained("ayazfau/GPT2-124-poetry-RL")

def generate_poetry(prompt, max_length=50):
    input_ids = tokenizer(prompt, return_tensors="pt").input_ids
    output = model.generate(input_ids, max_new_tokens=max_length, do_sample=True, top_k=50, top_p=0.95, pad_token_id=tokenizer.eos_token_id)
    return tokenizer.decode(output[0], skip_special_tokens=True)

print(generate_poetry("fear kill dreams,"))

Model Performance

• The model generates rhyming and coherent poetic lines.
• Works well for short-form poetry and lyrical content.
• Still has some limitations in long-form poetry structure.

License

This model is released under the MIT License. Feel free to use it for research and creative writing!

Acknowledgments

• Trained using PyTorch Lightning and Hugging Face Transformers.
• Special thanks to the Gutenberg Poetry Corpus for providing high-quality literary data.

---

If you use this model, please consider citing it or leaving a star on Hugging Face! ⭐