---
language:
  - en
license: apache-2.0
tags:
  - text
pretty_name: Amazon ESCI Hard Negatives
size_categories:
  - "100K<n<1M"
source_datasets:
  - Amazon ESCI
task_categories:
  - sentence-similarity
dataset_info:
  config_name: default
  splits:
    - name: train
      num_bytes: 1
      num_examples: 1
    - name: test
      num_bytes: 1
      num_examples: 1      
train-eval-index:
  - config: default
    task: sentence-similarity
    splits:
      train_split: train
      eval_split: test
configs:
- config_name: default
  data_files:
  - split: train
    path: "data/train/*"
  - split: test
    path: "data/test/*"
---

# Amazon ESCI hard-negatives dataset

A dataset in a [nixietune](https://github.com/nixiesearch/nixietune) compatible format:

```json
{
  "query": "black cotton fabric by the yard",
  "pos": [
    {
      "doc": "Black 100% Cotton Material Prepackaged by The Yard",
      "bm25": 0.0,
      "cosE5": 0.9253104
    },
    {
      "doc": "Zen Creative Designs Cotton Fabric, Solid Black, 60\" Wide by 1 Yard/Easy to Sew/Great for Lining/Mask Making/Apparel/Quilting Projects and Hobbies (1 Yard)",
      "bm25": 7.026509,
      "cosE5": 0.9030106
    }
  ],
  "neg": [
    {
      "doc": "PLAZALA 7PCS Cotton Craft Fabric Bundle Squares Patchwork DIY Sewing Scrapbooking Quilting 9.8\"x9.8\" (Black)",
      "bm25": 0.0,
      "cosE5": 0.87003165
    },
    {
      "doc": "Connecting Threads Print Collection Precut Cotton Quilting Fabric Bundle Fat Quarters (Monochrome)",
      "bm25": 0.0,
      "cosE5": 0.86396176
    },
    {
      "doc": "Dekoth -Cotton Fabric 100% Cotton Poplin Printed Woven Bundles Voile for DIY Crafting Patchwork Sewing, Full Width 1 Piece cuttable 39 x 56 inches (100x142cm) (Black with Grey Floral)",
      "bm25": 0.0,
      "cosE5": 0.8673507
    }
  ]
```

This is the expanded version of the [Amazon ESCI small-en](todo) dataset with the following additions:
* for all queries, extra 32 negatives were genererated
* negative generation was done with a [RRF](todo)-based hybrid search, mixing the BM25 score with cosine-similarity based on [intfloat/e5-base-v2] emnbedding model.
* can be loaded with [HF datasets](todo) directly.

## Usage

```python
from datasets import load_dataset

data = load_dataset('nixiesearch/amazon-esci-hardnegatives', split="train")
```

## License

Apache 2.0