SCI MiniLM SimCSE v1

all-minilm-l6-v2-sci-simcse-v1 is a compact, general-purpose English sentence embedding model adapted to spinal cord injury (SCI) literature. It keeps the 384-dimensional embedding interface of sentence-transformers/all-MiniLM-L6-v2.

Use it for semantic search, text similarity, clustering, retrieval, and as an embedding feature for downstream SCI applications. It is not a clinical decision-support model and is not validated for causal inference, diagnosis, treatment selection, or patient-level prediction.

Quick start

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("pander14/all-minilm-l6-v2-sci-simcse-v1")
embeddings = model.encode(
    [
        "Autonomic dysfunction after spinal cord injury",
        "Rehabilitation outcomes following cervical SCI",
    ],
    normalize_embeddings=True,
)

Training

The base model was adapted with self-supervised SimCSE on a rights-filtered SCI literature corpus. Each passage was encoded twice with independent dropout masks; other passages in the minibatch acted as contrastive negatives. No causal labels, clinical labels, or query templates were used.

The source corpus was used for model adaptation and is not included in this repository. Corpus access and reuse remain subject to the underlying source rights and licenses.

Evaluation and limitations

The checkpoint was verified to load and produce 384-dimensional embeddings. It has not yet been benchmarked against the base MiniLM model on an external, application-specific test set. Evaluate it on representative held-out data before using it in a production workflow.

Scientific embeddings can reproduce biases, omissions, outdated findings, and terminology in the source literature. Similarity scores are not evidence of clinical validity or causal relationships.

License

This repository is distributed under the Apache-2.0 license. Users remain responsible for complying with the licenses and access terms of any data used alongside the model.