Hugging Face's logo

Datasets:
juliensimon
/
stackexchange-space-qa

Tasks:
Question Answering
Text Generation
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
10K - 100K
Tags:
space
astronomy
question-answering
stack-exchange
instruction-tuning
sft
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
stackexchange-space-qa / README.md
juliensimon's picture
juliensimon
Update Stack Exchange space Q&A: 33,519 questions
c83b335 verified
preview code
|
Raw
History Blame Contribute Delete
8.11 kB
metadata 
license: cc-by-sa-4.0
pretty_name: Stack Exchange Space Q&A
language:
  - en
description: >-
  This dataset is a clean, tabular Q&A corpus of space and astronomy knowledge,
  derived from two Stack Exchange community Q&A sites: Astronomy Stack Exchange
  (astronomy.stackexchange.com) and Space Expl
task_categories:
  - question-answering
  - text-generation
tags:
  - space
  - astronomy
  - question-answering
  - stack-exchange
  - instruction-tuning
  - sft
  - qa-pairs
  - open-data
  - tabular-data
  - parquet
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/stackexchange_space_qa.parquet
    default: true

Stack Exchange Space Q&A

The gamma-ray sky — representative of the community-written answers covering all of space science

Credit: NASA/DOE/Fermi LAT Collaboration

Part of a dataset collection on Hugging Face.

Dataset description

This dataset is a clean, tabular Q&A corpus of space and astronomy knowledge, derived from two Stack Exchange community Q&A sites: Astronomy Stack Exchange (astronomy.stackexchange.com) and Space Exploration Stack Exchange (space.stackexchange.com). Each row is one question paired with its best answer — either the question's accepted answer, or if none is accepted, the highest-scored answer. Unanswered questions are included with null answer fields so that downstream consumers can choose to filter them out.

Stack Exchange is vote-ranked, which makes it a particularly strong source of instruction-tuning data: community scores already encode a quality signal, and accepted-answer flags add an explicit gold-standard marker. The two sites together cover the full breadth of space science — from cosmology and stellar astrophysics to orbital mechanics, rocket propulsion, satellite operations, and crewed spaceflight. Topics tend toward the graduate-level-to-working-professional end of the spectrum; questions cite papers, use technical vocabulary, and receive carefully-written answers from practitioners.

The dataset is suitable for instruction fine-tuning (question → answer pairs), preference learning (score-ranked accepted vs. unaccepted pairs can be derived by joining this table with itself on qid), retrieval-augmented generation (as a grounded Q&A corpus for a space-science RAG system), benchmarking (filter by tag to build evaluation sets on a specific topic), and linguistic analysis of how practitioners explain space concepts to each other. HTML has been converted to plain text; inline formatting (code blocks, lists, equations) is preserved where possible.

Content is licensed CC-BY-SA 4.0 (Stack Exchange's standard license). Each row retains the question URL so that attribution can be traced to individual authors on the source site. The dataset is refreshed annually — Stack Exchange publishes quarterly data dumps to archive.org, but the quality-ranked subset changes slowly enough that yearly is sufficient.

Schema

Column Type Description Sample Null %
qid Int64 Stack Exchange question ID (unique within its site) 24590 0.0%
site string Which Stack Exchange site the question is from: 'astronomy' (astronomy.stackexchange.com) or 'space' (space.stackexchange.com — Space Exploration) astronomy 0.0%
url string Permalink to the question on Stack Exchange https://astronomy.stackexchange.com/q... 0.0%
question_title string Question title as posted How much gold is there in our sun? 0.0%
question_body string Question body in plain text (stripped of HTML, with code blocks and inline formatting preserved) XKCD 1944 claims that there is "more ... 0.0%
question_tags string Semicolon-joined list of tags attached to the question (e.g., 'black-holes;general-relativity') |the-sun| 0.0%
question_score Int64 Net vote score of the question (upvotes minus downvotes) at dump time 150 0.0%
question_view_count Int64 Number of times the question has been viewed 26681 0.0%
question_answer_count Int64 Total number of answers posted to the question 2 0.0%
question_creation_date string ISO-8601 UTC date when the question was posted 2018-01-19T08:19:35Z 0.0%
answer_body string Top-scored answer body in plain text, or null if the question is unanswered. Prefers the accepted answer when one exists; otherwise the highest-scored answer. The mass of the sun is 1.989 × 1030 k... 14.6%
answer_score Int64 Net vote score of the selected answer; null if unanswered 130 14.6%
answer_creation_date string ISO-8601 UTC date when the selected answer was posted; null if unanswered 2018-01-19T08:38:05Z 14.6%
answer_accepted boolean True if the selected answer is the question's accepted answer; False if it is just the top-scored answer; null if unanswered True 14.6%

Quick stats

  • 33,519 questions across space (18,866), astronomy (14,653)
  • 28,628 have a top answer (17,052 of those are the question's accepted answer)
  • Median question score: 4
  • 52,011,395 total question views across the corpus

Usage 

from datasets import load_dataset

ds = load_dataset("juliensimon/stackexchange-space-qa", split="train")
df = ds.to_pandas()

from datasets import load_dataset

ds = load_dataset("juliensimon/stackexchange-space-qa", split="train")
df = ds.to_pandas()

# High-quality accepted answers only — solid instruction-tuning pairs
high_quality = df[(df["answer_accepted"] == True) & (df["question_score"] >= 5)]
print(f"Accepted answers on highly-voted questions: {len(high_quality):,}")

# Filter by tag (e.g. exoplanets on Astronomy SE)
exo = df[df["question_tags"].str.contains("exoplanet", na=False)]
print(f"Exoplanet questions: {len(exo):,}")

# SFT-ready Q→A pairs
sft = (
    df[df["answer_body"].notna()]
    [["qid", "site", "question_title", "question_body", "answer_body"]]
    .rename(columns={"question_body": "prompt_body", "answer_body": "response"})
)

# Plot top 20 tags (astronomy)
import matplotlib.pyplot as plt
astro = df[df["site"] == "astronomy"]
tag_counts = astro["question_tags"].str.split(";").explode().value_counts().head(20)
tag_counts.plot.barh(figsize=(10, 6))
plt.xlabel("Questions"); plt.title("Most-used tags on Astronomy Stack Exchange")
plt.gca().invert_yaxis()
plt.tight_layout(); plt.show()

Data source

https://archive.org/details/stackexchange

Update schedule

Annually — SE publishes quarterly data dumps, but the accepted-answer quality-filtered subset changes slowly.

Related datasets

If you find this dataset useful, please consider giving it a like on Hugging Face. It helps others discover it.

About the author

Created by Julien Simon — AI Operating Partner at Fortino Capital. Part of the Space Datasets collection.

Citation 

@dataset{stackexchange_space_qa,
  title = {Stack Exchange Space Q&A},
  author = {juliensimon},
  year = {2026},
  url = {https://huggingface.co/datasets/juliensimon/stackexchange-space-qa},
  publisher = {Hugging Face}
}

License

CC-BY-SA-4.0