Jig09/onomatron-rnn
French literature name generator
2 min read
This model is a character-level language model inspired by Andrej Karpathy's well-known makemore tutorial project. It is trained on a dataset of names from French literary works and generates new fictional names in the same style. Based on a simple recurrent neural network (RNN) architecture, it contains just 7,657 parameters (weighing about 0.03 MB) and ranks as a benchmark leader among 11 architectures tested by the author. The model is useful for developers and researchers studying character-level sequence generation or building procedural name generation for text-based games.
Key features
- The architecture relies on a basic RNN module with a small parameter count (7,657) and a modest footprint of 0.03 MB.
- Top score on the val NLL metric (2.3592) among 11 architectures investigated by the author (including MLP, WaveNet, LSTM, and Transformer).
- High inference speed: the model produces up to 4,962 names per second on a CPU.
- Hyperparameters were tuned using the Optuna framework with experiment tracking via MLflow.
What you can do
- Generate new fictional names for role-playing game characters or literary projects.
- Integrate the ready-to-use
config.jsonandmodel.ptfiles via the project's Python library. - Explore the architecture code and compare generation quality with other benchmark options (MLP, LSTM, RWKV, etc.).
Example
from onomatron.serialize import load_bundle_from_hub
net, config = load_bundle_from_hub("Jig09/onomatron-rnn")
print(net.sample(n=20, seed=42))
Read more
DetailsBenchmarks: val NLL = 2.3592, test NLL = 2.3491.
Hardware requirements: training took 54.7s on mps with peak VRAM usage of 395 MB.
Limitations: the model operates at the character level without semantics or grammar, is trained strictly on French literary classics, and may occasionally reproduce names from the training set.
License: MIT for both code and weights.










Comments