README.md

---
language:
- en
task_categories:
- text-retrieval
size_categories:
- 10K<n<100K
dataset_info:
- config_name: documents
  features:
  - name: chunk_id
    dtype: string
  - name: chunk
    dtype: string
  - name: source_url
    dtype: string
  - name: title
    dtype: string
  - name: chunk_idx
    dtype: int64
  - name: chunk_start_char
    dtype: int64
  - name: chunk_end_char
    dtype: int64
  splits:
  - name: test
    num_bytes: 10518145
    num_examples: 69199
  download_size: 3065416
  dataset_size: 10518145
- config_name: documents_coref
  features:
  - name: chunk_id
    dtype: string
  - name: chunk
    dtype: string
  - name: source_url
    dtype: string
  - name: title
    dtype: string
  - name: chunk_idx
    dtype: int64
  - name: chunk_start_char
    dtype: int64
  - name: chunk_end_char
    dtype: int64
  splits:
  - name: test
    num_bytes: 10956870
    num_examples: 69199
  download_size: 3175651
  dataset_size: 10956870
- config_name: queries
  features:
  - name: original_query
    dtype: string
  - name: query
    dtype: string
  - name: answer
    list: string
  - name: score
    list: int64
  - name: source_url
    list: string
  - name: frag_start_char
    list: int64
  - name: frag_end_char
    list: int64
  - name: n_gold
    dtype: int64
  splits:
  - name: test
    num_bytes: 294674
    num_examples: 271
  download_size: 145407
  dataset_size: 294674
configs:
- config_name: documents
  data_files:
  - split: test
    path: documents/test-*
- config_name: documents_coref
  data_files:
  - split: test
    path: documents_coref/test-*
- config_name: queries
  data_files:
  - split: test
    path: queries/test-*
---

# ConditionalQA — Chunk-level Retrieval Eval (from DAPR)

Contextualized chunk-level (query2chunk) retrieval evaluation for the **ConditionalQA**
subset of the [DAPR benchmark](https://huggingface.co/datasets/UKPLab/dapr)
(*Document-Aware Passage Retrieval*, Wang, Reimers & Gurevych, ACL 2024,
[arXiv:2305.13915](https://arxiv.org/abs/2305.13915)).

This is a reformat of DAPR into the chunk-level retrieval schema used across the
[Chunk-level Retrieval Eval](https://huggingface.co/collections/bowang0911/chunk-level-retrieval-eval)
collection. **Chunks are DAPR's native passages** — the documents are *not* re-chunked
into fixed token windows. Each chunk is exactly one DAPR passage (its natural
paragraph / gold segmentation).

## Configs

| Config | Rows | Description |
| --- | --- | --- |
| `documents` | 69,199 | One row per passage (native DAPR segmentation). |
| `documents_coref` | 69,199 | Same passages, **coreference-resolved** text (DAPR `corpus_coref`). |
| `queries` | 271 | Queries with gold fragment char-spans. |

### `documents` / `documents_coref`
| field | type | notes |
| --- | --- | --- |
| `chunk_id` | string | DAPR passage id (`{doc}-{paragraph}`) |
| `chunk` | string | passage text (coref-resolved in `documents_coref`) |
| `source_url` | string | DAPR `doc_id` (ConditionalQA has no URLs) |
| `title` | string | document title |
| `chunk_idx` | int64 | paragraph number within the document |
| `chunk_start_char` | int64 | char offset of the passage in the reconstructed document |
| `chunk_end_char` | int64 | char end offset |

> **Coref note:** `documents_coref` reuses the **same char offsets** as `documents`
> (i.e. the plain-text coordinate space) so that the single `queries` config drives
> overlap-based gold matching for *both* configs. The `chunk` text differs (antecedents
> inlined), so `chunk_end_char - chunk_start_char` need not equal `len(chunk)` in the
> coref config — the offsets act as stable passage locators, not coref-text lengths.

### `queries`
| field | type | notes |
| --- | --- | --- |
| `original_query` / `query` | string | query text |
| `answer` | list[string] | gold passage text(s) |
| `score` | list[int64] | relevance per gold fragment (from DAPR qrels) |
| `source_url` | list[string] | document id per gold fragment |
| `frag_start_char` / `frag_end_char` | list[int64] | char span of each gold passage |
| `n_gold` | int64 | number of gold fragments |

## Gold matching

A chunk is **gold** for a query if it shares `source_url` with a gold fragment and its
`[chunk_start_char, chunk_end_char)` overlaps that fragment's
`[frag_start_char, frag_end_char)`. Because each chunk is a whole passage and each gold
fragment is a whole passage, every gold fragment maps to **exactly one** chunk
(verified: 1,165 / 1,165 gold fragments recover as exactly one chunk in both
`documents` and `documents_coref`).

## Provenance

Built from `UKPLab/dapr` configs `ConditionalQA-corpus`, `ConditionalQA-corpus_coref`,
`ConditionalQA-queries` (test), and `ConditionalQA-qrels` (test). Documents are
reconstructed by concatenating passages in `paragraph_no` order (joined with `\n`) to
assign char offsets; gold fragment spans are the char spans of the gold passages
referenced by the qrels.

## Citation

```bibtex
@article{wang2023dapr,
    title  = "DAPR: A Benchmark on Document-Aware Passage Retrieval",
    author = "Kexin Wang and Nils Reimers and Iryna Gurevych",
    journal= "arXiv preprint arXiv:2305.13915",
    year   = "2023",
    url    = "https://arxiv.org/abs/2305.13915"
}
```