AdhyanshVerma/stack-2021-12-01

ReasonStack-Prime

A highly normalized, streaming-optimized Stack Exchange corpus engineered for LLM reasoning and instruction tuning.

CC BY-SA 4.0 ~1M Rows 176 Parquet Shards 21 SE Sites

1. Executive Summary

ReasonStack-Prime is a large-scale, meticulously curated text dataset derived from the official Archive.org Stack Exchange data dump (Version 2021-12-07). Unlike raw XML dumps or poorly cleaned JSON exports, this dataset applies a rigorous streaming ETL pipeline to normalize HTML into clean Markdown, preserve complex code structures, and retain critical metadata.

The corpus contains 999,892 rows across 21 distinct Stack Exchange communities, totaling 3.36 GB in highly compressed Parquet format. It is specifically designed to serve as a high-quality pretraining or fine-tuning corpus for Large Language Models (LLMs) targeting complex reasoning, code generation, and technical question answering.

999,892

Total Instances

18.4 GB

Raw XML Processed

3.36 GB

Final Parquet Size

176

Optimized Shards

---

2. Pipeline Architecture & Data Curation

Processing 18+ GB of deeply nested, inconsistent XML requires significant memory management. This dataset was built using a custom streaming architecture (stackexchange-streaming-v1.0) designed to run efficiently on constrained hardware without OOM (Out of Memory) errors.

2.1 The ETL Process

1

Streaming XML Parsing: Utilized xml.etree.ElementTree.iterparse to stream the 7z-compressed XML files directly from Archive.org. This avoids loading the entire DOM into memory, keeping RAM usage strictly bounded.

2

HTML to Markdown Sanitization: Implemented a custom AST-based converter. Unlike naive regex replacements, this step correctly handles nested blockquotes, preserves Stack Exchange's specific MathJax syntax ($ and $$), and strictly protects <pre><code> blocks from being mangled by the Markdown parser.

3

Answer Aggregation & Ranking: For each question, all associated answers are joined. The pipeline identifies the accepted_answer_id; if none exists, it falls back to the highest-scoring answer to populate the best_answer field.

4

Parquet Sharding (Dask/Polars): The normalized data is partitioned into 176 optimized Parquet shards. This specific shard count ensures that distributed training frameworks (like DeepSpeed or FSDP) can assign whole files to individual GPUs without requiring complex row-level slicing.

2.2 Filtering & Quality Control

To ensure high signal-to-noise ratio, the following heuristic filters were applied during the streaming phase:

• Deleted Content: Removed questions and answers marked with deleted flags in the XML dump.
• Spam & Abusive: Filtered out posts with severe spam flags.
• Empty Bodies: Dropped instances where the Markdown conversion resulted in an empty string (e.g., image-only questions without alt-text).

---

3. Comprehensive Data Schema

The dataset is provided in a flat, normalized structure optimized for Hugging Face datasets and analytical tools like Polars.

Column	Type	Description & Edge Cases
id	int64	Unique Post ID from the Stack Exchange database. Guaranteed unique across the entire corpus.
site	string	The FQDN of the source site (e.g., stackoverflow.com, math.stackexchange.com).
title	string	Raw HTML-decoded title of the question.
question_body_md	string	Clean Markdown representation of the question. MathJax is preserved as raw LaTeX.
question_score	int32	Net vote count (upvotes minus downvotes) at the time of the 2021 dump.
question_tags	list[string]	Array of tags. Note: Stack Overflow limits to 5 tags; other sites may have fewer.
answers	list[struct]	Array of all answers. Struct contains: id, body_md, score, is_accepted.
best_answer	string	The text of the accepted answer. If no answer was accepted, this contains the highest-voted answer. Null if 0 answers exist.
text	string	Primary Training Field. A pre-formatted string combining Title, Tags, Question Body, and Best Answer, optimized for direct causal language modeling (CLM).
tokens_estimate	int32	Estimated token count using standard BPE/GPT-2 tokenizers. Useful for context-window filtering.
meta	dict	Provenance data: dump_version, parser_version, extraction_date, and original Archive.org URLs.

---

4. Site Distribution & Sharding

The corpus covers 21 distinct communities, heavily weighted toward technical and scientific domains to maximize reasoning density.

Domain	Shards	Domain Focus	Domain	Shards	Domain Focus
stackoverflow.com	80	Software Engineering	physics.stackexchange.com	6	Theoretical/Applied Physics
math.stackexchange.com	10	Pure & Applied Math	stats.stackexchange.com	6	Statistics & ML Theory
tex.stackexchange.com	13	LaTeX / Typesetting	softwareengineering	5	Architecture / Design
unix.stackexchange.com	8	Linux / CLI / Shell	dba.stackexchange.com	5	Database Admin / SQL
mathematica	8	Symbolic Computation	electronics	4	Circuit Design / EE
codereview	8	Code Optimization	security / crypto	6	InfoSec / Cryptography
gis.stackexchange.com	7	Geospatial / GIS	Others (cs, ai, devops...)	14	Specialized Computing

---

5. Usage Guide

5.1 Streaming with Hugging Face datasets

Because the dataset is sharded, you can stream it without downloading the 3.36 GB payload to disk.

from datasets import load_dataset

# Stream the dataset directly
ds = load_dataset("AdhyanshVerma/ReasonStack-Prime", split="train", streaming=True)

# Iterate through high-quality software engineering questions
for row in ds:
    if row['site'] == 'stackoverflow.com' and row['question_score'] > 20:
        print(row['text'])
        break