Ferrxni/baudelaire-gpt2-small

Baudelaire GPT-2 Small

Fine-tuned version of dbddv01/gpt2-french-small on Les Fleurs du Mal by Charles Baudelaire.
This model generates French poetry in the style of Baudelaire when prompted with a word or a phrase.

---

Model Description

• Base model: GPT-2 small (French version)
• Architecture: Decoder-only Transformer, ~124M parameters
• Fine-tuning data: Les Fleurs du Mal (Project Gutenberg edition)
• Languages: 🇫🇷 French
• Use case: Creative text generation, poetry, literature experiments
• License of base model: MIT (check original repo)
• License of dataset: Public domain (Project Gutenberg)

---

Training Details

• Framework: 🤗 Transformers Trainer
• Block size: 128 tokens
• Optimizer: AdamW
• Learning rate: 2e-4 with linear decay
• Epochs: 8
• Batch size (effective): 16 sequences (4 × gradient_accumulation=4)
• Device: Apple MPS (macOS)

Final training loss ≈ 5.1 → PPL ≈ 164.
More epochs or more poetic corpora (Verlaine, Rimbaud, Mallarmé) could improve results.

---

How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "Ferrxni/baudelaire-gpt2-small"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "Amour et Mort"
inputs = tokenizer(prompt, return_tensors="pt")

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=150,
        do_sample=True,
        temperature=0.9,
        top_p=0.95,
        repetition_penalty=1.2
    )

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Example output

• Prompt: Amour et Mort

• Generated: Amour et Mort s’embrassent dans une clarté funèbre,
  Le parfum du tombeau caresse encore la chair,
  Et l’âme, en frissonnant, se suspend aux ténèbres,
  Comme une fleur de nuit dans le souffle de l’air.

---

Files

• model.safetensors: fine-tuned weights (~475 MB)
• config.json: model config
• tokenizer.json, vocab.json, merges.txt: GPT-2 tokenizer files
• tokenizer_config.json, special_tokens_map.json: tokenizer metadata