README.md

---
language:
- en
pretty_name: QASPER Chat-Format
license: unknown
task_categories:
- question-answering
task_ids:
- extractive-qa
source_datasets:
- allenai/qasper
tags:
- text
- science
- chat-format
- instruction-tuning
- datasets
- allenai/qasper
- evidence-selection
- arxiv:2105.03011
---
# QASPER (Chat-Format Preparation)

This dataset is a chat-format preparation of QASPER for supervised fine-tuning (SFT).

## Format

This format is commonly referred to as:

- chat-format SFT data
- instruction-tuning conversations
- OpenAI-style `messages` format

## Included files

- `train.jsonl`
- `validation.jsonl`
- `stats.json`
- `prepare_qasper_unsloth.py`

## Source

- Base dataset: `allenai/qasper`

## Original Dataset Highlights

- Original dataset: `allenai/qasper`
- Focus: question answering on scientific NLP papers with evidence selection.
- Reported scale on source card: 5,049 questions over 1,585 papers.
- Key annotation properties: multiple answer types (free-form, extractive, yes/no, unanswerable) and evidence annotations.
- Paper: [A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers](https://arxiv.org/abs/2105.03011)

## Preparation summary

- One row per `(paper, question)` using the best available annotation.
- Answer normalization priority:
  1. free-form
  2. yes/no
  3. extractive spans
  4. unanswerable
- Context mode is mixed between:
  - evidence-only
  - full-text
- User prompt follows a question-first structure.

Assistant target is the normalized answer text.

## Schema

Each JSONL row contains:

- `messages`
  - `user`: text instruction + question + title + abstract + context
  - `assistant`: text answer
- `meta`: ids, answer type, context mode, evidence count

## Reproduction

```bash
python prepare_qasper_unsloth.py
```