GTE Tiny LLM-family stylometry embedder
This is a 22.7M-parameter, 384-dimensional SentenceTransformers model fine-tuned to place English essays with similar authorship and model-family style near one another in embedding space.
It is intended for exploratory analysis of aggregate stylistic geometry: clustering batches of model outputs, comparing family centroids, and studying changes in writing style. It is not an AI-text detector, and similarity to a model family is not proof of distillation, training-data provenance, or model identity.
Architecture and training
- Architecture: BERT / GTE Tiny
- Parameters: 22,713,216
- Embedding dimensions: 384
- Maximum sequence length: 512 tokens
- Similarity: cosine
- Training examples: 60,000 triplets
- Training: one epoch with Multiple Negatives Ranking Loss plus an auxiliary triplet loss
- Initialization: a general-purpose GTE Tiny stylometry checkpoint ultimately derived from
TaylorAI/gte-tiny
The upstream TaylorAI/gte-tiny repository does not currently declare a license in its Hub metadata. Accordingly, this repository does not assert a new license over the upstream-derived weights.
Usage
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("trentmkelly/gte-tiny-llm-family-stylometry")
texts = [
"First essay or model response...",
"Second essay or model response...",
]
embeddings = model.encode(
texts,
normalize_embeddings=True,
batch_size=64,
)
cosine_similarity = embeddings @ embeddings.T
For comparisons across model families, use multiple content-matched samples per family and compare centroids in the original embedding space. Dimensionality reductions such as PCA or UMAP are useful visualizations but should not be treated as distance metrics.
Internal evaluation
On an internal held-out essay-pair evaluation, the final checkpoint obtained:
| Metric | Value |
|---|---|
| Average precision | 0.9724 |
| F1 | 0.9403 |
| Accuracy | 0.9377 |
| Precision | 0.9023 |
| Recall | 0.9817 |
The accuracy/F1 figures use a threshold selected on the same internal evaluation output and should not be interpreted as performance on an independent public benchmark. The full training/evaluation record is included as training_metrics.json.
Limitations and responsible use
- The model was trained primarily on English essay-style text. Results may not transfer to chat, code, short posts, other languages, or heavily edited documents.
- Topic, prompt wording, register, document length, decoding settings, and provider-specific prompting can all produce apparent family structure.
- A family centroid based on one representative model is also model-specific; it is not a complete description of the family.
- Closed model APIs can change without notice, so apparent stylistic relationships are time-dependent.
- Do not use this checkpoint alone to accuse a person or company of concealed model use, plagiarism, or misrepresentation.
- For serious comparisons, use content-matched prompts, multiple independent samples, original-space distances, prompt-stratified uncertainty estimates, and explicit negative controls.
Reproducibility notes
The checkpoint was saved with:
sentence-transformers==3.3.1transformers==4.57.0torch==2.4.1+cu121
Newer SentenceTransformers releases should load the standard model layout, but the versions above reproduce the original environment most closely.