nixiesearch/beir-eval-hard-negatives

BEIR/MTEB hard negatives dataset

A dataset for quick evaluation of embedding models during their training.

The problem: running a full MTEB evaluation on a single GPU may take 10-20 hours. Most of this time is spent on embedding all 30M docs in all 10+ corpora. This dataset solves this problem by unwrapping a "retrieval" style benchmark into the "reranking" style:

• We compute embeddings for all documents in the corpora with the intfloat/e5-base-v2 model.
• For each corpus in BEIR/MTEB benchmark we build a Lucene index with text documents and their embeddings.
• For each eval query we do a hybrid RRF-based retrieval for top-32 negatives

As BEIR testset is size-unbalanced (TREC-COVID is 42 queries, and MS MARCO is ~4000) we sample top-300 random queries from each dataset.

It takes around 30-60 seconds to perform eval using Nixietune on a single RTX 4090.

A dataset in a nixietune compatible format:

{
  "query": ")what was the immediate impact of the success of the manhattan project?",
  "pos": [
    "The presence of communication amid scientific minds was equally important to the success of the Manhattan Project as scientific intellect was. The only cloud hanging over the impressive achievement of the atomic researchers and engineers is what their success truly meant; hundreds of thousands of innocent lives obliterated."
  ],
  "neg": [
    "Abstract. The pivotal engineering and scientific success of the Twentieth century was the Manhattan Project. The Manhattan Project assimilated concepts and leaders from all scientific fields and engineering disciplines to construct the first two atomic bombs.",
    "The pivotal engineering and scientific success of the Twentieth century was the Manhattan Project. The Manhattan Project assimilated concepts and leaders from all scientific fields and engineering disciplines to construct the first two atomic bombs."
  ]
}

Usage

To use with HF datasets:

pip install datasets zstandard

from datasets import load_dataset

data = load_dataset('nixiesearch/beir-eval-hard-negatives')
print(data["test"].features)

License

Apache 2.0