---
license: apache-2.0
configs:
  - config_name: kv75
    data_files:
      - split: test
        path: "data/kv75.jsonl"
  - config_name: kv140
    data_files:
      - split: test
        path: "data/kv140.jsonl"
  - config_name: kv300
    data_files:
      - split: test
        path: "data/kv300.jsonl"
  - config_name: qa10
    data_files:
      - split: test
        path: "data/qa10.jsonl"
  - config_name: qa20
    data_files:
      - split: test
        path: "data/qa20.jsonl"
  - config_name: qa30
    data_files:
      - split: test
        path: "data/qa30.jsonl"
task_categories:
- question-answering
tags:
- lost-in-the-middle
size_categories:
- n<1K
---
# Datasets for Lost In The Middle

This repository contains datasets used in the paper ["Lost in the Middle: How Language Models Use Long Contexts"](https://arxiv.org/abs/2307.03172), focusing on multi-document question answering and key-value retrieval tasks.

## Datasets Overview

The datasets provided are as follows:

- **Key-Value Retrieval Datasets**
  - `kv75`: Key-Value pairs with 75 keys.
  - `kv140`: Key-Value pairs with 140 keys.
  - `kv300`: Key-Value pairs with 300 keys.

- **Multi-Document Question Answering Datasets**
  - `qa10`: Questions with answers derived from 10 documents.
  - `qa20`: Questions with answers derived from 20 documents.
  - `qa30`: Questions with answers derived from 30 documents.

## Loading the Data

You can load these datasets using the Hugging Face `datasets` library:

```python
from datasets import load_dataset

### Example for loading the kv75 dataset
dataset = load_dataset("bzantium/LITM", "kv75")

### Example for loading the qa20 dataset
dataset = load_dataset("bzantium/LITM", "qa20")
```