license: mit
library_name: transformers
tags:
- biomedical
- entity-linking
- retrieval
- reranking
- lora
language:
- en
PILOT checkpoints
Trained artifacts for PILOT: Neighborhood-Aware Dual Biomedical Entity Linking.
Code (anonymized for review): https://anonymous.4open.science/r/PILOT-92B4
PILOT links a biomedical mention, in its sentence context, to an entity in a large ontological knowledge base, in three stages: neighborhood-aware retrieval → dual reranking → score fusion.
Contents
| folder | what it is | size |
|---|---|---|
retriever/<ds>/ |
SapBERT fine-tuned on that dataset's training mentions. The encoder φ(·) used by both the retriever and the surface-form reranker. |
499 MB |
reranker/<ds>/ |
LoRA adapter for the contextual reranker. Load on top of the base model below. | 151 MB (4B) / 47 MB (0.6B) |
<ds> ∈ ncbi, bc5cdr, cometa, mm, aap_fold0 … aap_fold9 (AAP is 10-fold cross-validation).
| dataset | base reranker |
|---|---|
| ncbi, bc5cdr, cometa, aap_fold0..9 | Qwen/Qwen3-Reranker-4B |
| mm | Qwen/Qwen3-Reranker-0.6B |
Adapters rather than merged models: the merge is deterministic, and 14 merged 4B checkpoints would be ~127 GB against ~2 GB of adapters.
Results
Test R@1 (top-1 accuracy); AAP is the 10-fold mean.
| NCBI | BC5CDR | COMETA | AAP | MM-ST21pv |
|---|---|---|---|---|
| 93.44 | 95.75 | 87.98 | 91.72 | 74.93 |
Usage
pip install -U huggingface_hub
hf download anon4papersubmission/PILOT-checkpoints --local-dir checkpoints
The retriever is a plain HF encoder:
from transformers import AutoTokenizer, AutoModel
tok = AutoTokenizer.from_pretrained("checkpoints/retriever/ncbi")
enc = AutoModel.from_pretrained("checkpoints/retriever/ncbi") # CLS pooling, 768-d
The reranker is a LoRA adapter — merge it, or load it on the base model:
python -c "from swift.pipelines import export_main; export_main()" \
--adapters checkpoints/reranker/ncbi --merge_lora true \
--output_dir models/rerank/ncbi_merged --use_hf true
See the code repository for the full pipeline (data preparation through evaluation) and for how these plug into stages 5 and 7.
License
MIT for the checkpoints. The underlying benchmarks and UMLS retain their own licenses; UMLS requires a free NLM account.