Planet Namer is a tiny character-level LSTM that generates science-fiction
planet names from seven normalized planet attributes. It was built for
on-device use in Stellar Broadcast and exported as a fixed-shape,
single-token ONNX model.
This repository is licensed for non-commercial use only under
CC BY-NC 4.0. See
LICENSE for the repository-specific notice.
Model details
Property
Value
Architecture
Single-layer character LSTM with stat-conditioned initial state, per-step stat concatenation, and FiLM conditioning
Parameters
226,242
Vocabulary
66 tokens (63 characters plus PAD/SOS/EOS)
Maximum output length
20 characters
Hidden / embedding size
192 / 64
Inputs
Five fixed-shape tensors; see below
Outputs
Next-character logits and recurrent hidden/cell states
The seven input values must be in [0, 1] and in this exact order:
planet_namer_checkpoint.pt — PyTorch state dictionary
vocab.json — vocabulary, token IDs, stat order, and dimensions
inference.py — minimal ONNX Runtime command-line example
train.py — architecture, training, evaluation, and export code
The ONNX model is a single-step recurrent model. On the first step, pass the
real stat vector to both stats_init and stats, along with zero h_in and
c_in. On later steps, pass zeros to stats_init, keep passing the real
values to stats, and feed the previous h_out and c_out back into the
model.
Use --temperature 0 for greedy decoding. Higher temperatures increase
variation. The helper validates the stat count and range before inference.
The PyTorch checkpoint is a state dictionary, not a serialized executable
model. Instantiate PlanetNameLSTM from train.py with a vocabulary size of
66, then load the state dictionary with weights_only=True.
Training data
The model was trained on 1,957 planet and location names collected from these
fictional universes: Star Trek, Mass Effect, Warhammer 40,000, Star Wars,
Dune, Babylon 5, Halo, Stargate, Firefly, Foundation, and The Expanse.
The conditioning values are hand-authored or synthetic metadata. When several
names shared the same stat vector, the training pipeline deterministically
spread those values using orthographic properties of each name.
The raw name lists are not included in this model repository. Names and
marks from the referenced fictional universes may be protected by copyright,
trademark, or other rights belonging to their respective owners. This release
does not grant rights to any third-party material.
Evaluation
The following results were reproduced from the released checkpoint with seed
42 and the training script's stratified 80/10/10 split (1,564 / 194 / 199):
Metric
Result
Training exact match at temperature 0.1
82.35% (1,288 / 1,564)
Test exact match at temperature 0.1
77.39%
Test mean Levenshtein distance
1.87
Novelty at temperature 1.0
88.5% (200 generated samples)
Generated-vs-training character-bigram KL
0.2901
These are development diagnostics, not a benchmark. In particular, the
name-derived collision spreading leaks orthographic information into the
conditioning values, making held-out exact-match results optimistic. The
novelty result is a single seeded sampling run and will vary.
Intended use
Intended uses are non-commercial creative experiments, games, prototypes,
research, and procedural-content tooling where a user wants short fictional
planet-name suggestions conditioned on normalized attributes.
The model is not intended for factual astronomy, scientific classification,
identity-related naming, or commercial products and services. Review generated
names before publication.
Limitations and risks
The model can reproduce or closely resemble names seen during training.
Outputs may resemble protected franchise names or marks; novelty is not a
clearance check.
The training corpus is small, English-centric, and dominated by a few
fictional universes.
The stat/name relationship is partly synthetic and should not be interpreted
as semantic ground truth.
Sampling can produce empty, awkward, truncated, or mixed-case strings.
The ONNX graphs use fixed batch size 1 and fixed recurrent-state dimensions.
Users are responsible for reviewing outputs, respecting third-party rights,
providing attribution, and complying with the non-commercial license.
