IMRNNs

Interpretable Modular Retrieval Neural Networks (IMRNNs)
Adapter-only checkpoint release for dense retriever adaptation

Website · GitHub · EACL 2026 Paper

     

This repository hosts the released IMRNN checkpoints. Each checkpoint is adapter-only, and the corresponding base dense retriever is loaded separately by the imrnns package.

Important Training Note

The released checkpoints are dataset-specific. They are not zero-shot checkpoints evaluated unchanged across all of BEIR.

Examples:

• imrnns-minilm-trec-covid.pt is trained for trec-covid
• imrnns-e5-nq.pt is trained for nq

If you want to train IMRNNs on another dataset or on another retriever, use the full GitHub implementation:

• https://github.com/YashSaxena21/IMRNNs

Download First

Clone or download this Hugging Face repository first. The commands below assume you are running them from inside the downloaded repo.

git clone https://huggingface.co/yashsaxena21/IMRNNs
cd IMRNNs

Install

pip install -r requirements.txt
pip install -e .

How to Get the Dataset and Embeddings

The full data-preparation flow lives in the GitHub repository. The package downloads official BEIR datasets from:

• https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/<dataset>.zip

Example:

• https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/trec-covid.zip

To build the cache for a dataset and retriever:

python -m imrnns cache \
  --encoder minilm \
  --dataset trec-covid \
  --datasets-dir /path/to/datasets \
  --cache-dir /path/to/cache_minilm_trec-covid \
  --device cpu

That command downloads the dataset if needed, embeds the corpus and queries, and writes the training files required by IMRNNs.

Quick Start

Load the matching base retriever and the released IMRNN adapter checkpoint:

from imrnns import IMRNNAdapter

adapter = IMRNNAdapter.from_pretrained(
    encoder="minilm",
    dataset="trec-covid",
    repo_id="yashsaxena21/IMRNNs",
    device="cpu",
)

scores = adapter.score(
    query="What is the incubation period of COVID-19?",
    documents=[
        "COVID-19 symptoms can appear 2 to 14 days after exposure.",
        "The stock market closed higher today.",
        "Transmission risk depends on exposure setting and viral load.",
    ],
    top_k=3,
)

for item in scores:
    print(item.rank, item.score, item.text)

End-to-End Evaluation

from pathlib import Path

from imrnns import cache_embeddings, load_pretrained
from imrnns.beir_data import load_beir_source
from imrnns.data import load_cached_split
from imrnns.evaluation import evaluate_model

encoder = "minilm"
dataset = "trec-covid"
repo_id = "yashsaxena21/IMRNNs"
device = "cpu"
k = 10

model, metadata, encoder_spec = load_pretrained(
    encoder=encoder,
    dataset=dataset,
    repo_id=repo_id,
    device=device,
)

cache_dir = Path("demo_cache") / "cache_minilm_trec-covid"
datasets_dir = Path("datasets")

if not (cache_dir / "test" / "embeddings.pt").exists():
    cache_embeddings(
        encoder=encoder,
        dataset=dataset,
        cache_dir=cache_dir,
        datasets_dir=datasets_dir,
        device=device,
    )

source = load_beir_source(dataset, datasets_dir=datasets_dir)
cached_test = load_cached_split(cache_dir, "test", source, encoder_spec, device)
metrics = evaluate_model(
    model=model,
    cached_split=cached_test,
    device=device,
    feedback_k=100,
    ranking_k=k,
    k_values=[k],
)

print({"metrics": metrics, "metadata": metadata})

Reported metrics:

• MRR@10
• Recall@10
• NDCG@10

Your Own Retriever

If you trained IMRNNs on a different dense retriever, provide your own base model and checkpoint directly:

from imrnns import IMRNNAdapter

adapter = IMRNNAdapter.from_checkpoint(
    checkpoint_path="my_imrnn_adapter.pt",
    encoder_model_name="my-org/my-retriever",
    embedding_dim=768,
    query_prefix="query: ",
    passage_prefix="passage: ",
    device="cpu",
)

Released Checkpoints

MiniLM:

• checkpoints/pretrained/minilm/imrnns-minilm-msmarco.pt
• checkpoints/pretrained/minilm/imrnns-minilm-fiqa.pt
• checkpoints/pretrained/minilm/imrnns-minilm-hotpotqa.pt
• checkpoints/pretrained/minilm/imrnns-minilm-nq.pt
• checkpoints/pretrained/minilm/imrnns-minilm-scifact.pt
• checkpoints/pretrained/minilm/imrnns-minilm-trec-covid.pt
• checkpoints/pretrained/minilm/imrnns-minilm-webis-touche2020.pt

E5:

• checkpoints/pretrained/e5/imrnns-e5-msmarco.pt
• checkpoints/pretrained/e5/imrnns-e5-fiqa.pt
• checkpoints/pretrained/e5/imrnns-e5-hotpotqa.pt
• checkpoints/pretrained/e5/imrnns-e5-nq.pt
• checkpoints/pretrained/e5/imrnns-e5-scifact.pt
• checkpoints/pretrained/e5/imrnns-e5-webis-touche2020.pt

Citation

@inproceedings{saxena-etal-2026-imrnns,
  title = "{IMRNN}s: An Efficient Method for Interpretable Dense Retrieval via Embedding Modulation",
  author = "Saxena, Yash and
    Padia, Ankur and
    Gunaratna, Kalpa and
    Gaur, Manas",
  booktitle = "Findings of the Association for Computational Linguistics: EACL 2026",
  month = mar,
  year = "2026",
  address = "Rabat, Morocco",
  publisher = "Association for Computational Linguistics",
  url = "https://aclanthology.org/2026.findings-eacl.333/",
  doi = "10.18653/v1/2026.findings-eacl.333",
  pages = "6324--6337"
}