This dataset package is a public reproducibility artifact for the AQEA technical report. It contains embedding pairs + human similarity scores for STS benchmarks, generated deterministically from public sources.

What’s inside

The intended upload contains JSON files in the format:

{
  "dataset": "sts12",
  "model": "intfloat/e5-large-v2",
  "dimension": 1024,
  "n_pairs": 3108,
  "baseline_spearman": 0.77,
  "embeddings1": [[...], ...],
  "embeddings2": [[...], ...],
  "scores": [0.0, 1.0, ...]
}

Files (E5-Large 1024D):

stsb_intfloat_e5_large_v2_1024d_pairs.json
sts12_e5large_1024d_pairs.json
sts13_e5large_1024d_pairs.json
sts14_e5large_1024d_pairs.json
sts15_e5large_1024d_pairs.json
sts16_e5large_1024d_pairs.json

Additionally, the package includes:

MANIFEST.sha256 — checksums for all files in the package

Sources

Datasets (HuggingFace):

mteb/stsbenchmark-sts (STS-B)
mteb/sts12-sts, mteb/sts13-sts, mteb/sts14-sts, mteb/sts15-sts, mteb/sts16-sts

Embedding model:

intfloat/e5-large-v2 (1024D)

How this dataset is generated

This repository contains scripts to generate the exact files locally:

STS-B: training/scripts/generate_stsb_benchmark.py
STS12–16: benchmark/scripts/generate_all_embeddings.py

Packaging + checksum generation:

aqea-technical-report/repro/package_hf_dataset.py

Intended use

Reproducing the AQEA generalization benchmark and “117× @ ~95% task retention” claim.
Independent third-party verification of reported metrics.

License / redistribution note

This package is generated from public benchmark datasets and contains embeddings + scores (no raw sentences). Users should still review the upstream dataset terms for their own compliance needs.

Downloads last month: 7