---
dataset_info:
  features:
  - name: query_id
    dtype: string
  - name: query
    dtype: string
  - name: positive_passages
    list:
    - name: docid
      dtype: string
    - name: text
      dtype: string
    - name: title
      dtype: string
  - name: negative_passages
    list:
    - name: docid
      dtype: string
    - name: text
      dtype: string
    - name: title
      dtype: string
  - name: subset
    dtype: string
  splits:
  - name: train
    num_bytes: 924685.308369815
    num_examples: 50
  download_size: 465674
  dataset_size: 924685.308369815
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
language:
- en
task_categories:
- text-ranking
---

This dataset is used in the paper [Fixing Data That Hurts Performance: Cascading LLMs to Relabel Hard Negatives for Robust Information Retrieval](https://huggingface.co/papers/2505.16967). It contains query-passage pairs, including positive and negative passages, used for training robust retrieval and reranker models. The dataset was created by identifying and relabeling hard negatives in existing retrieval datasets using cascading LLM prompts.