README.md

---
license: cc-by-4.0
task_categories:
- question-answering
- text-retrieval
language:
- en
tags:
- rag
- qasper
- scientific-qa
- retrieval
size_categories:
- 10K<n<100K
configs:
- config_name: answers
  data_files:
  - split: train
    path: answers/train*
  - split: dev
    path: answers/dev*
  - split: test
    path: answers/test*
- config_name: corpus
  data_files:
  - split: train
    path: corpus/*
- config_name: qrels
  data_files:
  - split: train
    path: qrels/train*
  - split: dev
    path: qrels/dev*
  - split: test
    path: qrels/test*
- config_name: queries
  data_files:
  - split: train
    path: queries/train*
  - split: dev
    path: queries/dev*
  - split: test
    path: queries/test*
- config_name: retrieved_docs
  data_files:
  - split: train
    path: retrieved_docs/train-*
  - split: dev
    path: retrieved_docs/dev-*
  - split: test
    path: retrieved_docs/test-*
- config_name: top_ranked
  data_files:
  - split: train
    path: top_ranked/train*
  - split: dev
    path: top_ranked/dev*
  - split: test
    path: top_ranked/test*
dataset_info:
  config_name: retrieved_docs
  features:
  - name: query_id
    dtype: string
  - name: corpus_id
    dtype: string
  - name: rank
    dtype: int64
  - name: retrieval_score
    dtype: float64
  - name: is_relevant
    dtype: bool
  splits:
  - name: train
    num_bytes: 1681424
    num_examples: 20985
  - name: dev
    num_bytes: 712632
    num_examples: 8894
  - name: test
    num_bytes: 1048757
    num_examples: 13089
  download_size: 758823
  dataset_size: 3442813
---

# QASPER RAG

Dataset for Retrieval-Augmented Generation (RAG) based on [QASPER](https://huggingface.co/datasets/allenai/qasper).

## Structure

| Subset | Splits | Description |
|--------|--------|-------------|
| `corpus` | train (default) | Paper chunks (abstract + full-text paragraphs) shared across all query splits |
| `queries` | train, dev, test | Information-seeking questions over scientific papers |
| `qrels` | train, dev, test | Relevance judgments (query ↔ paragraph chunk) |
| `answers` | train, dev, test | Reference answers (longest valid free-form answer) |
| `top_ranked` | train, dev, test | Paper-scoped candidate pool (all chunks of the query paper) |
| `retrieved_docs` | train, dev, test | Top-k retrieval results with relevance labels |

## Dataset statistics

| Split | Queries | Corpus |
|-------|--------:|-------:|
| train | 2101 | 81550 |
| dev   | 890 | 81550 |
| test  | 1310 | 81550 |

The corpus is shared across all splits and contains paragraph-level chunks from the abstract and full text of each paper.

- **Dev split:** mapped from the original `validation` split
- **Corpus source:** unique papers from train, validation and test splits
- **Chunking:** one chunk for the abstract (`section_name: abstract`) and one chunk per paragraph in `full_text`

## Source

| Component | QASPER resource |
|-----------|-----------------|
| Train | `train` split from [allenai/qasper](https://huggingface.co/datasets/allenai/qasper) |
| Dev | `validation` split |
| Test | `test` split |
| Corpus | `abstract` + `full_text.paragraphs` |
| Queries | `qas.question` |
| Qrels | `qas.answers[*].answer[*].evidence` matched to corpus chunks |
| Top ranked | All paragraph chunks from the query paper (paper-scoped retrieval pool) |
| Answers | Longest valid answer per question (`free_form_answer`, joined `extractive_spans`, or `Yes`/`No`) |

## Filtering

Questions are kept only when at least one answer satisfies all of the following:

- `unanswerable` is `false`
- the answer has `free_form_answer`, non-empty `extractive_spans`, or `yes_no`
- after removing evidence items containing `FLOAT SELECTED`, at least one answer still has evidence that matches the corpus

Evidence items containing `FLOAT SELECTED` are removed individually. Questions are omitted only when no valid answer has remaining evidence.

## Schema

### corpus
```json
{"id": "...", "title": "...", "section_name": "...", "text": "..."}
```

### queries
```json
{"id": "...", "text": "..."}
```

### qrels
```json
{"query_id": "...", "corpus_id": "...", "score": 1}
```

### answers
```json
{"query_id": "...", "answer": "..."}
```

### top_ranked
```json
{"query-id": "...", "corpus-ids": ["paper_00000", "paper_00001"]}
```

### retrieved_docs
```json
{"query_id": "...", "corpus_id": "...", "rank": 1, "retrieval_score": 0.92, "is_relevant": true}
```

Top-k documents retrieved from the indexed corpus (`is_relevant` is derived from `qrels`).

## Usage

```python
from datasets import load_dataset

corpus = load_dataset("DinoStackAI/qasper-rag", "corpus")["train"]
queries = load_dataset("DinoStackAI/qasper-rag", "queries")
qrels = load_dataset("DinoStackAI/qasper-rag", "qrels")
answers = load_dataset("DinoStackAI/qasper-rag", "answers")

train_queries = queries["train"]
dev_qrels = qrels["dev"]
test_answers = answers["test"]
```

## Citation

QASPER is released under the [CC BY 4.0 License](https://creativecommons.org/licenses/by/4.0/).

```bibtex
@inproceedings{Dasigi2021ADO,
  title={A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers},
  author={Pradeep Dasigi and Kyle Lo and Iz Beltagy and Arman Cohan and Noah A. Smith and Matt Gardner},
  year={2021}
}
```