README.md

---
configs:
- config_name: memwrap
  data_files:
  - split: test
    path: memwrap/qasper.jsonl
- config_name: plain
  data_files:
  - split: test
    path: plain/qasper.jsonl
---

# QASPER Benchmark

Question Answering on Scientific Papers - NLP research paper comprehension benchmark.

## Overview

| Metric | Value |
|--------|-------|
| Papers | 416 (test set) |
| Questions | 1,370 (answerable) |
| Answer Types | Free-form, extractive, yes/no |
| Context | Full paper (title, abstract, sections) |

## Source

Based on [QASPER](https://allenai.org/data/qasper) dataset by AllenAI.

Paper: [A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers](https://aclanthology.org/2021.naacl-main.365/)

## Variants

- **memwrap**: Paper content wrapped with `<|memory_start|>` / `<|memory_end|>` tags
- **plain**: Raw paper content without memory tags

## Usage

```python
from datasets import load_dataset

# Load memwrap variant
ds = load_dataset("tonychenxyz/qasper", "memwrap", split="test")

# Load plain variant
ds = load_dataset("tonychenxyz/qasper", "plain", split="test")
```

## Scoring

Uses `qasper_log_perplexity` scoring function:
- Evaluates model performance using log perplexity of generated answer tokens
- Lower log perplexity indicates better performance
- Matches the perplexity-based evaluation used in cardridge baselines

Target answers are stored in `extra_info.ground_truth.answer`.

## Citation

```bibtex
@inproceedings{dasigi2021qasper,
  title={A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers},
  author={Dasigi, Pradeep and Lo, Kyle and Beltagy, Iz and Cohan, Arman and Smith, Noah A and Gardner, Matt},
  booktitle={NAACL},
  year={2021}
}
```