---
configs:
- config_name: corpus
  data_files:
  - path: corpus/corpus-*
    split: corpus
- config_name: queries
  data_files:
  - path: queries/queries-*
    split: queries
- config_name: instruction
  data_files:
  - path: instruction/instruction-*
    split: instruction
- config_name: default
  data_files:
  - path: data/default-*
    split: test
- config_name: qrel_diff
  data_files:
  - path: qrel_diff/qrel_diff-*
    split: qrel_diff
- config_name: top_ranked
  data_files:
  - path: top_ranked/top_ranked-*
    split: top_ranked
dataset_info:
- config_name: corpus
  features:
  - dtype: string
    name: _id
  - dtype: string
    name: title
  - dtype: string
    name: text
  splits:
  - name: corpus
    num_examples: 30921
- config_name: queries
  features:
  - dtype: string
    name: _id
  - dtype: string
    name: text
  splits:
  - name: queries
    num_examples: 64
- config_name: instruction
  features:
  - dtype: string
    name: query-id
  - dtype: string
    name: instruction
  splits:
  - name: instruction
    num_examples: 64
- config_name: default
  features:
  - dtype: string
    name: query-id
  - dtype: string
    name: corpus-id
  - dtype: float64
    name: score
  splits:
  - name: test
    num_examples: 8554
- config_name: qrel_diff
  features:
  - dtype: string
    name: query-id
  - list: string
    name: corpus-ids
  splits:
  - name: qrel_diff
    num_examples: 32
- config_name: top_ranked
  features:
  - dtype: string
    name: query-id
  - list: string
    name: corpus-ids
  splits:
  - name: top_ranked
    num_examples: 64
language:
- en
multilinguality:
- monolingual
tags:
- text-retrieval
- instruction-retrieval
task_categories:
- text-retrieval
task_ids:
- document-retrieval

---

# news21-instructions-mteb

This is a new version of the news21-instructions dataset modified to fit the new MTEB format.

1. Restructured queries to include both original and changed versions
2. Separated instructions into a dedicated configuration
3. Reorganized qrels into default (original) and qrel_diff configurations

## Dataset Structure

The dataset contains the following configurations:


- corpus: Original corpus documents
- queries: Queries with both original and changed versions
- instruction: Instructions for both original and changed queries
- default: Original relevance judgments
- qrel_diff: Changes in relevance judgments
- top_ranked: Top ranked documents for each query