加载中…
加载中…
Source model · HF release · Build hub
Pinned conversion of the Apache-2.0 BioClinical ModernBERT mean-pooling sentence embedding model. The native ModernBERT graph has 22 layers, 768 dimensions, local attention, and an 8,192-token native context window.
| File | Status | Purpose |
|---|---|---|
bioclinical-modernbert-F16.gguf |
Published | Reference embeddings |
bioclinical-modernbert-Q8_0.gguf |
Published | Higher-quality compact embeddings |
bioclinical-modernbert-Q4_K_M.gguf |
Published | Smallest release format |
Real llama-embedding output, --pooling mean --embd-normalize 2, 768-dimensional L2-normalised vectors, cosine similarity. Query: "Which drug class is used to treat type 2 diabetes?"
| Rank | Q4_K_M | Q8_0 | F16 | Document |
|---|---|---|---|---|
| 1 | 0.7165 | 0.7094 | 0.7058 | Metformin is a first-line oral medication for type 2 diabetes mellitus. |
| 2 | 0.4085 | 0.3903 | 0.3870 | Insulin glargine is a long-acting basal insulin analogue for diabetes. |
| 3 | 0.2887 | 0.2768 | 0.2751 | Amoxicillin is a beta-lactam antibiotic used for bacterial infections. |
| 4 | 0.1376 | 0.1267 | 0.1239 | Aspirin irreversibly inhibits cyclooxygenase enzymes in platelets. |
| 5 | 0.0714 | 0.0640 | 0.0658 | The Airbus A320 is a narrow-body commercial airliner. |
| 6 | 0.0270 | 0.0163 | 0.0177 | The Great Wall of China was built across northern China. |
All three quantizations produce the same ranking, and the off-domain sentences land at the bottom, so Q4_K_M is usable for retrieval. Scores are cosine values, not probabilities, and the margins are not calibrated confidence.
Twenty PubMedQA query/document pairs, 768-dimensional mean-pooled vectors, L2-normalized comparison:
| Format | Mean cosine to source | Top-1 retrieval |
|---|---|---|
| F16 | 0.999998 | 0.95 |
| Q8_0 | 0.999672 | 0.95 |
| Q4_K_M | 0.984172 | 0.95 |
The F16 GGUF matches the original SentenceTransformer embedding path. Retrieval agreement is measured against the 20-pair set, not a large benchmark.
python -m pip install -r requirements-build.txt
python build_gguf.py --model-id NeuML/bioclinical-modernbert-base-embeddings
The builder pins llama.cpp commit 6b790a9c291b5d7af3312bbf9f0c558aa023b13e and the upstream model revision. It does not upload or overwrite this repository.
Apache-2.0. See the upstream model card and LICENSE for source attribution.