ruggsea/dante-zero-20250306-Pleias-350m-Preview

Dante-Zero Fine-tuned Model

This model was fine-tuned using Reinforcement Learning with Generative Pre-trained Transformer Optimization (GRPO) to generate Dante-style poetry in endecasillabi (11-syllable lines).

Model Details

• Base Model: PleIAs/Pleias-350m-Preview
• Training Method: GRPO (Generative Pre-trained Transformer Optimization)
• Training Data: 1,000 chunks from Dante's Divine Comedy
• Epochs: 10
• Trained By: ruggsea
• Date: 2025-03-06
• Run Name: dante-zero-20250306-Pleias-350m-Preview

Model Description

This model is specialized in generating Italian poetry in the style of Dante Alighieri's Divine Comedy. It has been trained to:

1. Generate proper endecasillabi (11-syllable lines)
2. Follow the structure of Dante's poetry
3. Avoid repetition
4. Create original content (not plagiarize the Divine Comedy)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained("ruggsea/dante-zero-20250306-Pleias-350m-Preview")
tokenizer = AutoTokenizer.from_pretrained("ruggsea/dante-zero-20250306-Pleias-350m-Preview", padding_side="left")

# Ensure proper tokenizer settings
tokenizer.pad_token = tokenizer.eos_token
tokenizer.pad_token_id = tokenizer.eos_token_id

# Generate poetry
prompt = "Nel mezzo del cammin di nostra vita"
inputs = tokenizer(prompt, return_tensors="pt", padding_side="left")
outputs = model.generate(
    inputs.input_ids,
    max_new_tokens=200,
    do_sample=True,
    temperature=0.7,
    top_p=0.9,
    repetition_penalty=1.2
)
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(generated_text)

Reward Functions

The model was trained using several reward functions:

1. Endecasillabo Checker: Rewards proper 11-syllable lines
2. Plagiarism Checker: Penalizes copying from the Divine Comedy
3. Verse Structure Checker: Encourages verse-like structure
4. Repetition Penalty: Discourages repetitive text