BioHiCL-Large: Hierarchical Multi-Label Contrastive Biomedical Retriever

Model Card

🔍 Overview

BioHiCL-large is a biomedical dense retriever trained with hierarchical MeSH supervision to capture fine-grained semantic relationships between biomedical texts.

Unlike traditional dense retrievers trained with binary relevance signals, BioHiCL models semantic similarity using structured multi-label supervision derived from the MeSH ontology, enabling it to capture partial semantic overlap between documents.

⚠️ Important: Please ensure that the transformers version matches exactly (4.57.3), as other versions may lead to compatibility issues or unexpected behavior.

---

💡 Key Features

• Hierarchical supervision: Leverages MeSH ontology to encode structured biomedical semantics
• Multi-label similarity learning: Captures graded semantic overlap beyond binary relevance
• Contrastive + regression training: Aligns embedding similarity with label similarity
• Efficient: ~0.3B parameters, suitable for deployment on a single GPU
• Domain-adapted retriever: Fine-tuned from a strong general-purpose bi-encoder

---

🧠 Model Details

• Model type: Bi-encoder (dense retriever)
• Backbone: BAAI/bge-large-en-v1.5
• Parameters: ~0.3B
• Fine-tuning: LoRA (merged into base model)
• Max input length: 512 tokens
• Training data: Biomedical abstracts annotated with MeSH labels (e.g., BioASQ-derived corpora)

---

⚙️ Intended Use

This model is intended for biomedical information retrieval tasks such as:

• Scientific literature search (e.g., PubMed-style retrieval)
• Biomedical document ranking
• Query–abstract semantic matching
• Benchmark evaluation on BEIR biomedical subsets

---

⚙️ How It Works

BioHiCL aligns:

• Embedding similarity (SimE): cosine similarity between document embeddings
• Label similarity (SimL): cosine similarity over weighted MeSH multi-label vectors

---

⚙️ Requirements

• python >= 3.8
• transformers == 4.57.3

  ⚠️ Important: Please ensure that the transformers version matches exactly (4.57.3), as other versions may lead to compatibility issues or unexpected behavior.

---

🚀 Usage (BEIR Evaluation)

from beir import util
from beir.datasets.data_loader import GenericDataLoader
from beir.retrieval.models import SentenceBERT
from beir.retrieval.search.dense import DenseRetrievalExactSearch
from beir.retrieval.evaluation import EvaluateRetrieval


# 1. Download  load the SciFact dataset
dataset = "scifact"
url = "https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/" + dataset + ".zip"

data_path = util.download_and_unzip(url, "datasets")
corpus, queries, qrels = GenericDataLoader(data_path).load(split="test")

# ⚠️ Important: Please ensure that the `transformers` version matches exactly (4.57.3), as other versions may lead to compatibility issues or unexpected behavior.
model_name = "LunaLan07/BioHiCL-large"
model = SentenceBERT(model_name)

retriever = DenseRetrievalExactSearch(model, batch_size=16)

top_k = 10  # top 10 documents per query
results = retriever.search(corpus, queries, top_k=top_k, score_function="cos_sim")

k_values = [1, 3, 5, 10]
ndcg, _map, recall, precision = EvaluateRetrieval.evaluate(qrels, results, k_values=k_values)

📖 Citation

If you use this model, please cite:

@article{lan2026biohicl,
  title={BioHiCL: Hierarchical Multi-Label Contrastive Learning for Biomedical Retrieval with MeSH Labels},
  author={Lan, Mengfei and Zheng, Lecheng and Kilicoglu, Halil},
  booktitle={ACL 2026},
  year={2026}
}