Lexi-100K
Lexi-100K is a tiny English keyword and entity tagger trained with PyTorch and designed to run
entirely on Roblox servers in Luau. Runtime inference needs no HTTP requests, API keys, or external
AI service.
Model details
- Parameters: 404,684
- Architecture: 32-dimensional word and hashed character-ngram embeddings, a five-token context
window, and a 64-neuron hidden layer
- Labels:
O, FOOD, GAME, ANIMAL, PERSON, PLACE, TIME, OBJECT, ITEM, ACTIVITY,
EMOTION, OTHER
- Roblox export: signed int8 tensors encoded as base64 Luau
- Runtime tensor size: approximately 405 KB
- Recommended confidence threshold:
0.75
- Adjacent same-type tokens are merged into complete spans such as
New York
- Curly apostrophes are normalized identically in Python and Luau
- Unseen capitalized names can be recognized from strong person context without a name database
Files
model.pt — full PyTorch checkpointvocabulary.json — learned 4,096-token vocabularymetrics.json — synthetic validation metricsbenchmark.json — independent 86-case challenge benchmarkroblox-benchmark.json — results measured in Roblox Studio server LuauModelWeights.lua — directly embeddable int8/base64 Roblox weights
Results
On the independent 86-case English challenge set, full-precision Python, dequantized int8 Python,
and the Roblox Luau runtime produced the same results at threshold 0.75:
| Metric | Result |
|---|
| Precision | 99.24% |
| Recall | 100% |
| F1 | 99.62% |
| Exact sentences | 85 / 86 |
| Type accuracy | 97.69% |
| Roblox average latency | 1.523 ms |
| Roblox throughput | ~657 inferences/s |
| Roblox Luau tests | 11 / 11 passed |
The challenge set is small and curated, so these numbers are not evidence of broad language
understanding. Unrepresented slang, languages other than English, unusual spelling, and concepts
outside the training domains can fail.
Use
The complete training code, Luau inference runtime, server endpoint, tests, and UI are available in
the
GitHub repository. You can also
play or copy the Roblox experience.
Python example after cloning the source repository:
1pip install -e .
2python -m lexi.predict "Guus played Roblox with Riley on Friday"
Roblox usage:
1local result = game.ReplicatedStorage.Remotes.AnalyzeText:InvokeServer(
2 "Guus played Roblox with Riley on Friday"
3)
Training data and limitations
Lexi combines deterministic, domain-specific generated examples with 10,000 short, safety-filtered
English prompts from
OpenAssistant OASST1
(Apache-2.0). OASST1 text is conservatively weak-labeled and is never treated as authoritative
keyword truth. The independent benchmark is excluded from all training paths.
The open-name fallback stores no names. It only promotes an unknown capitalized token when its
nearby words provide a strong person role, while known entities remain under the neural model's
type prediction.
Lexi is an experimental keyword tagger, not a moderation or safety system. Roblox text shown to
other players must still be filtered with Roblox TextService.
