hishammoizuddin/vericite-embedder

vericite-embedder

This is the sentence-embedding model used by VeriCite, a free plagiarism/paraphrase checker that cites its sources.

These are the unmodified weights of [`sentence-transformers/all-MiniLM-L6-v2`](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2), mirrored here so VeriCite can pin a stable, versioned copy rather than depending on the upstream repo's main branch at request time. All credit for the model itself belongs to the original authors (Nils Reimers and the Sentence-Transformers / UKP Lab team) — see their Sentence-BERT paper and the original model card for training data, evaluation results, and the full technical writeup. No fine-tuning, distillation, or architecture change has been applied here.

Why VeriCite uses it

VeriCite's matching pipeline needs a small, CPU-fast embedding model to compute semantic similarity between a document passage and candidate source text, as one signal alongside exact-match fuzzy alignment and a lexical-overlap guard (semantic similarity alone is not sufficient to flag a passage — see the VeriCite Space "How it works" tab for the full method). all-MiniLM-L6-v2 fits that constraint well: 384-dim output, ~90MB, and fast enough to run comfortably on a free-tier 2-vCPU CPU box with no GPU.

Usage

Identical to the upstream model — either via sentence-transformers:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("hishammoizuddin/vericite-embedder")
embeddings = model.encode(["Some sentence to embed.", "Another one."], normalize_embeddings=True)

or via transformers directly (mean-pool the token embeddings yourself, since this checkpoint is a base encoder, not a SentenceTransformer-native format at the transformers layer):

from transformers import AutoModel, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("hishammoizuddin/vericite-embedder")
model = AutoModel.from_pretrained("hishammoizuddin/vericite-embedder")

License

Apache 2.0, unchanged from the upstream model.