AbedGarra/msmarco-indexes

MS MARCO Dense Retrieval Indexes

Pre-built FAISS indexes of the MS MARCO corpus using 53 embedding models. (As part of our paper, TellTail)

This repository is intended to make it easy to run retrieval experiments across a large collection of embedding models without rebuilding the MS MARCO index for every model.

Each model is associated with two files:

• <model>.index — the FAISS retrieval index
• <model>_docids.npy — mapping from FAISS index positions to MS MARCO document IDs

All index files are stored under:

indexes/

Model names are converted to filenames by replacing / with _.

For example:

Alibaba-NLP/gte-base-en-v1.5

corresponds to:

indexes/Alibaba-NLP_gte-base-en-v1.5.index
indexes/Alibaba-NLP_gte-base-en-v1.5_docids.npy

Downloading an index

Individual indexes can be downloaded directly with huggingface_hub:

from huggingface_hub import hf_hub_download
import faiss
import numpy as np

repo_id = "AbedGarra/msmarco-indexes"
model = "Alibaba-NLP/gte-base-en-v1.5"

filename = model.replace("/", "_")

index_path = hf_hub_download(
    repo_id=repo_id,
    repo_type="dataset",
    filename=f"indexes/{filename}.index",
)

docids_path = hf_hub_download(
    repo_id=repo_id,
    repo_type="dataset",
    filename=f"indexes/{filename}_docids.npy",
)

index = faiss.read_index(index_path)
docids = np.load(docids_path)

Models

The repository contains MS MARCO indexes for the following 53 embedding models.

Model	Query prefix / instruction	Passage prefix
Qwen/Qwen3-Embedding-0.6B	Instruct: {retrieval_task}\nQuery:	—
Qwen/Qwen3-Embedding-4B	Instruct: {retrieval_task}\nQuery:	—
intfloat/e5-small-v2	query:	passage:
intfloat/e5-base-v2	query:	passage:
intfloat/e5-large-v2	query:	passage:
intfloat/multilingual-e5-small	query:	passage:
intfloat/multilingual-e5-base	query:	passage:
intfloat/multilingual-e5-large	query:	passage:
intfloat/multilingual-e5-large-instruct	Retrieve semantically similar text.	—
Alibaba-NLP/gte-modernbert-base	—	—
Alibaba-NLP/gte-large-en-v1.5	—	—
Alibaba-NLP/gte-base-en-v1.5	—	—
Alibaba-NLP/gte-multilingual-base	—	—
thenlper/gte-base	—	—
nomic-ai/modernbert-embed-base	search_query:	search_document:
nomic-ai/nomic-embed-text-v1.5	search_query:	search_document:
nomic-ai/nomic-embed-text-v1	search_query:	search_document:
sentence-transformers/gtr-t5-base	—	—
sentence-transformers/gtr-t5-large	—	—
sentence-transformers/all-MiniLM-L6-v2	—	—
sentence-transformers/all-MiniLM-L12-v2	—	—
sentence-transformers/all-mpnet-base-v2	—	—
sentence-transformers/paraphrase-MiniLM-L6-v2	—	—
sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2	—	—
sentence-transformers/paraphrase-multilingual-mpnet-base-v2	—	—
sentence-transformers/sentence-t5-base	—	—
sentence-transformers/sentence-t5-large	—	—
BAAI/bge-small-en-v1.5	—	—
BAAI/bge-base-en-v1.5	—	—
BAAI/bge-large-en-v1.5	—	—
BAAI/bge-m3	—	—
jinaai/jina-embeddings-v3	Represent the query for retrieving evidence documents:	Represent the document for retrieval:
jinaai/jina-embeddings-v5-text-small-retrieval	Query:	Document:
jinaai/jina-embeddings-v2-base-en	—	—
jinaai/jina-embeddings-v2-small-en	—	—
jinaai/jina-embedding-b-en-v1	—	—
Snowflake/snowflake-arctic-embed-m-v2.0	query:	—
Snowflake/snowflake-arctic-embed-m	query:	—
HIT-TMG/KaLM-embedding-multilingual-mini-v1	—	—
HIT-TMG/KaLM-embedding-multilingual-mini-instruct-v1	Instruct: Given a query, retrieve documents that answer the query.\nQuery:	—
HIT-TMG/KaLM-embedding-multilingual-mini-instruct-v1.5	Instruct: Given a query, retrieve documents that answer the query.\nQuery:	—
google/embeddinggemma-300m	task: search result | query:	title: none | text:
NovaSearch/stella_en_400M_v5	Instruct: Given a web search query, retrieve relevant passages that answer the query.\nQuery:	—
Octen/Octen-Embedding-0.6B	—	—
OrdalieTech/Solon-embeddings-large-0.1	query:	—
WhereIsAI/UAE-Large-V1	Represent this sentence for searching relevant passages:	—
avsolatorio/GIST-large-Embedding-v0	—	—
manu/sentence_croissant_alpha_v0.3	—	—
manu/sentence_croissant_alpha_v0.4	—	—
mixedbread-ai/mxbai-embed-large-v1	Represent this sentence for searching relevant passages:	—
telepix/PIXIE-Rune-v1.0	query:	—
Lajavaness/bilingual-embedding-large	—	—
openai/text-embedding-3-small	—	—

— indicates that no additional query or passage prefix was used.

For Qwen models, {retrieval_task} represents the retrieval instruction supplied when encoding queries.

Query formatting

Some embedding models require different formatting for queries and passages.

For example, E5 models use:

query = "query: " + query
passage = "passage: " + passage

while Nomic models use:

query = "search_query: " + query
passage = "search_document: " + passage

When querying an index, use the same model-specific query formatting used when generating its embeddings.

Repository contents

The repository contains only the retrieval artifacts required to use the indexes:

indexes/
├── <model>.index
├── <model>_docids.npy
├── ...

Additional local metadata or specification files used while constructing the indexes are not included.

Citation

If you use this dataset in your research, please cite our paper:

@article{garra2026telltailembeddingsfingerprintingretrievers,
  title         = {The TellTail of Embeddings: Fingerprinting Retrievers in Black-Box Systems},
  author        = {Abdullah Garra and Matan Ben-Tov and Mahmood Sharif},
  year          = {2026},
  eprint        = {2610.04026},
  Journal = {arXiv},
  primaryClass  = {cs.CR},
  url           = {https://arxiv.org/abs/2610.04026}
}

Notes

These files are intended primarily for retrieval and embedding-model comparison experiments over MS MARCO.

The FAISS index and its corresponding docids.npy file should always be used together, since the NumPy file maps FAISS result positions back to MS MARCO document IDs.