onomatron-wavenet
fantasy name generator
2 min readupdated
This micro model is built on the WaveNet architecture and trained on character names from French literature, inspired by the famous makemore project. Operating at the character level without general language semantics or grammar, the AI model predicts sequences of letters. Weighing just 0.15 MB with around 37,000 parameters, it was created by developer Vincent Le Meur as part of the onomatron benchmark. It is a great fit for book authors, game developers, and anyone who loves experimenting with compact language architectures to generate unique fictional names.
Key features
- WaveNet architecture, showing balanced results on the character-level benchmark.
- Ultracompact size: the model file weighs 0.15 MB with 37,624 parameters.
- Trained on a curated dataset of sanitized names from French literary works.
- Detailed training tracking with Optuna and MLflow.
What you can do
- Generate a list of unique fictional names for a novel or tabletop game characters using a Python script.
- Integrate the model into a local text generation pipeline using the package from the repository.
- Use the repository as an educational example for studying character-level micro models.
Example
from onomatron.serialize import load_bundle_from_hub
net, config = load_bundle_from_hub("Jig09/onomatron-wavenet")
print(net.sample(n=20, seed=42))
Read more
DetailsModel metrics: val NLL = 2.4002, test NLL = 2.3632.
Performance: 1395 names/sec on CPU.
Training: mps device, peak memory usage 409 MB, 10000 steps.
License: MIT.










Comments