metadata

configs:
  - config_name: passages
    data_files:
      - split: train
        path: passages_parquet/*
  - config_name: queries
    data_files:
      - split: test
        path: queries_jsonl/*

TREC-RAG 2024 Corpus (MSMARCO 2.1) - Encoded with Cohere Embed English v3

This dataset contains the embeddings for the TREC-RAG Corpus 2024 embedded with the Cohere Embed V3 English model.

It contains embeddings for 113,520,750 passages, embeddings for 1677 queries from TREC-Deep Learning 2021-2023, as well as top-1000 hits for all queries using a brute-force (flat) index.

Search over the Index

We have a pre-build index that only requires 300 MB available at TREC-RAG-2024-index. Just pass in your Cohere API key, and you are able to search across 113M passages.

The linked index used PQ-compression with memory-mapped IVF, reducing your memory need to only 300MB, while achieving 97% search quality compared to a float32 flat index (that requires 250+GB memory and is extremely slow).

Passages

Passages - Parquet

113,520,750 passages are embedded. The parquet files can be found in the folder passages_parquet. Each row is a passage from the corpus. The column emb contains the respective embedding.

You can stream the dataset for example like this:

from datasets import load_dataset

dataset = load_dataset("Cohere/msmarco-v2.1-embed-english-v3", "passages", split="train", streaming=True)

for row in dataset:
    print(row)
    break

Passages - JSONL and Numpy

The folder passages_jsonl contain the .json.gz files for the passages as distributed by the task organizers.

The folder passages_npy contains a numpy matrix with all the embeddings for the respective .json.gz file.

When your server has enough memory, you can load all doc embeddings like this:

import numpy as np
import glob

emb_paths = sorted(glob.glob("passages_npy/*.npy"))

for e_path in emb_paths:
  doc_emb = np.load(e_path)

Queries

For 1677 queries from TREC-Deep Learning 2021, 2022 and 2023 we compute the embedding and the respective top-1k hits from a brute-force (flat) index. These queries can e.g. be used to test different ANN setting, e.g. in Recall@10 scenarios.

We also added annotations from NIST for the 215 queries that received an annotation. These queries have a non-empty qrel column.

The format is the following:

"_id": The query ID
"text": Query text
"trec-year": TREC-Deep Learning year
"emb": Cohere Embed V3 embedding
"top1k_offsets": Passage ID (int) when the numpy matrices are loaded sequentially and vertically stacked
"top1k_passage_ids": Passage ID (string) as they appear in the dataset
"top1k_cossim": Cosine similarities
"qrels": Relevance annotations for the 215 annotated queries by NIST

Queries - JSONL

The folder queries_jsonl/ contains the queries in a .jsonl.gz format.

Load this .jsonl.gz format is significantly faster than the parquet file.

Queries - Parquet

If you want to use the parquet file or the HF datasets library, the folder queries_parquet/ contains the respective parquet file.

You can load the queries with the following command in HF datasets

from datasets import load_dataset

dataset = load_dataset("Cohere/msmarco-v2.1-embed-english-v3", "queries", split="test")

for row in dataset:
    print(row)
    break

License

The embeddings are provided as Apache 2.0. The text data, qrels etc. are provided following the license of MSMARCO v2.1