DARE (Distribution-Aware Retrieval Embedding) is a specialized bi-encoder model designed to retrieve statistical and data analysis tools (R functions) based on both user queries and conditional on data profile.
It is fine-tuned from sentence-transformers/all-MiniLM-L6-v2 to serve as a high-precision tool retrieval module for Large Language Model (LLM) Agents in automated data science workflows.
- Paper: DARE: Aligning LLM Agents with the R Statistical Ecosystem via Distribution-Aware Retrieval
- Repository: GitHub
- Project Page: DARE Webpage
Model Details
- Architecture: Bi-encoder (Sentence Transformer)
- Base Model:
sentence-transformers/all-MiniLM-L6-v2(22.7M parameters) - Task: Dense Retrieval for Tool-Augmented LLMs
- Performance: SoTA on R package retrieval tasks (93.47% NDCG@10).
- Domain: R programming language, Data Science, Statistical Analysis functions
Usage (Sentence-Transformers)
First, install the sentence-transformers library:
pip install -U sentence-transformers
Usage with RPKB (Recommended)
Download the R Package Knowledge Base (RPKB) to perform conditional retrieval.
from huggingface_hub import snapshot_download
import chromadb
import os
# 1. Download the database folder from Hugging Face
db_path = snapshot_download(
repo_id="Stephen-SMJ/RPKB",
repo_type="dataset",
allow_patterns="RPKB/*"
)
# 2. Connect to the local ChromaDB instance
client = chromadb.PersistentClient(path=f"{db_path}/RPKB")
# 3. Access the specific collection
collection = client.get_collection(name="inference")
print(f"✅ Loaded {collection.count()} R functions ready for conditional retrieval!")
Retrieval with DARE
from sentence_transformers import SentenceTransformer
# 1. Load the DARE model
model = SentenceTransformer("Stephen-SMJ/DARE-R-Retrieval")
# 2. Define the exact input format: Query + Data Profile
query = "I have a high-dimensional genomic dataset named hidra_ex_1_2000.csv in my environment. I need to identify driver elements by estimating regulatory scores based on the counts provided
in the data. Please set the random seed to 123 at the start. I need to filter for fragment lengths between 150 and 600 bp and use a DNA count filter of 5. For my evaluation, please print the
first value of the estimated scores (est_a) for the very first region identified."
# 3. Generate embedding
query_embedding = model.encode(query).tolist()
# 4. Search in the database
results = collection.query(
query_embeddings=[query_embedding],
n_results=3,
include=["metadatas", "distances", "documents"]
)
# Display Top-1 Result
print("Top-1 Function:", results["metadatas"][0][0]["package_name"], "::", results["metadatas"][0][0]["function_name"])
Citation
If you find DARE, RPKB, or RCodingAgent useful in your research, please cite:
@article{sun2026dare,
title={DARE: Aligning LLM Agents with the R Statistical Ecosystem via Distribution-Aware Retrieval},
author={Maojun Sun and Yue Wu and Yifei Xie and Ruijian Han and Binyan Jiang and Defeng Sun and Yancheng Yuan and Jian Huang},
year={2026},
eprint={2603.04743},
archivePrefix={arXiv},
primaryClass={cs.IR},
url={https://arxiv.org/abs/2603.04743},
}