CARE-0.3B-4B

CARE-0.3B-4B

CARE-0.3B-4B is a complete asymmetric dense retrieval system for Chinese medical text retrieval. It consists of a lightweight 0.3B query encoder and a 4B document encoder.

The query encoder is intended for low-latency online query encoding, while the document encoder is intended for offline document encoding and indexing. The two encoders are trained as a pair and should be used together.

Model Components

This repository/model release represents the paired 0.3B + 4B CARE system. The 4B document encoder should not be paired with an unrelated query encoder when reproducing the reported results.

Intended Use

CARE-0.3B-4B is intended for:

• Chinese medical passage retrieval
• Medical knowledge-base search
• Retrieval-augmented generation over Chinese medical documents
• Offline document embedding and vector indexing
• Research on asymmetric dense retrieval

Typical deployment:

1. Encode the document corpus offline with the 4B document encoder.
2. Store document embeddings in a vector index.
3. Encode incoming queries online with the 0.3B query encoder.
4. Retrieve documents using dot-product or cosine similarity.

Out-of-Scope Use

This model is not a standalone chatbot, reranker, or medical diagnosis system. Retrieved passages must not be treated as medical advice or as a substitute for clinical judgment.

Inference

The inference wrapper is provided in inference/asymmetric.py.

from inference.asymmetric import CARE
import numpy as np

model = CARE(
    model_name_or_path_query="path/to/CARE-0.3B-query-encoder",
    model_name_or_path_doc="PhilipGAQ/CARE-0.3B-4B",
    trust_remote_code=True,
    use_fp16=False,
    normalize_embeddings=True,
    query_batch_size=2,
    passage_batch_size=2,
)

queries = ["什么是高血压？"]
documents = [
    "高血压是指动脉血压持续升高，通常指收缩压≥140mmHg和/或舒张压≥90mmHg。"
]

query_embeddings = model.encode_queries(queries, task_name="retrieval")
document_embeddings = model.encode_corpus(documents, task_name="retrieval")

scores = np.dot(query_embeddings, document_embeddings.T)
print(scores)

When normalize_embeddings=True, embeddings are L2-normalized and dot product is equivalent to cosine similarity. Similarity scores are ranking signals, not calibrated probabilities.

Training

CARE uses a two-stage asymmetric training strategy:

1. Query-side alignment training with the document encoder fixed.
2. Joint fine-tuning of the query and document encoders.

This progressively aligns representations produced by the structurally different query-side and document-side encoders.

Evaluation

The paired CARE system is evaluated on the Chinese Medical Text Embedding Benchmark (CMedTEB), which covers retrieval, reranking, and semantic textual similarity (STS).

Results should be reported for the complete configuration:

CARE 0.3B query encoder + CARE 4B document encoder

See the paper for benchmark splits, metrics, baselines, and full results.

Limitations and Responsible Use

• The model is primarily optimized for Chinese medical text retrieval.
• Performance may degrade on non-Chinese, non-medical, or highly specialized domains.
• Retrieval quality depends on chunking, preprocessing, document quality, and indexing strategy.
• Retrieved information may be incomplete, outdated, duplicated, or clinically inappropriate.
• The model does not verify clinical correctness and must not be used alone for diagnosis, treatment, medication, triage, or patient-specific risk decisions.
• Medical applications require qualified human review, source attribution, freshness checks, and appropriate privacy and safety controls.

Resources

• Code: https://github.com/PhilipGAQ/CARE
• Benchmark: https://huggingface.co/datasets/PhilipGAQ/CMedTEB
• Paper: https://arxiv.org/abs/2604.10937

Citation

@inproceedings{jiang2026benchmarking,
  title={Benchmarking and Enabling Efficient Chinese Medical Retrieval via Asymmetric Encoders},
  author={Jiang, Angqing and Chen, Jianlyu and Wang, Yongcan and Li, Xinpeng and Ding, Keyu and Lian, Defu and others},
  booktitle={Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={20000--20020},
  year={2026}
}

License

CC-BY-NC-SA-4.0