SalutefaceAAC/regex-explanations

Regex Explanations

A synthetic dataset of 205 curated regex patterns paired with plain-English explanations, match/non-match examples, and rich metadata. Designed for training and evaluating models that explain or generate regular expressions.

Dataset Summary

Each record maps a regular expression pattern to:

• A detailed plain-English explanation of what it matches and why
• Lists of positive examples (strings that match)
• Lists of negative examples (strings that do not match)
• Metadata: regex flavor, difficulty level, category, and tags

The dataset covers 8 pattern categories across 5 regex flavors, hand-authored for accuracy and clarity.

Supported Tasks

• Regex explanation generation (pattern → explanation): train seq2seq or LLM models to explain what a regex does
• Regex synthesis (explanation → pattern): generate a regex from a plain-English description
• Regex classification: predict difficulty, category, or flavor from the pattern
• Code education: teach regex concepts with worked examples

Dataset Structure

regex-explanations/
├── train.jsonl / train.parquet       (164 examples, 80%)
├── validation.jsonl / validation.parquet  (20 examples, 10%)
└── test.jsonl / test.parquet          (21 examples, 10%)

Fields

Field	Type	Description
id	string	Unique identifier (re_XXXX)
pattern	string	The regex pattern as written
flavor	string	Regex engine flavor: Python, JavaScript, PCRE, Go, Java
explanation	string	Plain-English explanation of what the pattern matches and why
matches	list[string]	Example strings that the pattern matches
non_matches	list[string]	Example strings that the pattern does NOT match
difficulty	string	beginner, intermediate, or advanced
category	string	Pattern category (see below)
tags	list[string]	Concept tags (e.g., lookahead, word-boundary, email)

Categories

Category	Count	Description
anchors	15	^, $, \b, \A, \Z, \G, lookbehind anchors
character_classes	25	[...], \d, \w, \s, POSIX, Unicode properties
quantifiers	22	*, +, ?, {n,m}, lazy, possessive
groups	21	Capturing, non-capturing, named, backreferences, atomic
lookarounds	17	Positive/negative lookaheads and lookbehinds
alternation	16	|, branch reset groups
special_sequences	18	\n, \t, \x, \u, \p{}, verbose mode
real_world	71	Email, URL, IP, dates, phone numbers, code patterns, Markdown

Difficulty Distribution

Difficulty	Count	%
beginner	62	30%
intermediate	85	41%
advanced	58	28%

Regex Flavor Distribution

Flavor	Count
Python	174
PCRE	20
JavaScript	6
Go	3
Java	2

Sample Records

Pattern	Flavor	Difficulty	Category	Explanation (truncated)	Matches	Non-Matches
^!\[([^\]]*)\]\(([^)]+)\)$	Python	intermediate	real_world	Matches a standalone Markdown image: . Group 1 captures the al...	, , 	[link](url), ![no-close(img.png)
\d{1,3}(\.\d{1,3}){3}	Python	intermediate	quantifiers	Rough pattern for an IPv4 address. Matches 1–3 digits, followed by exactly three...	192.168.1.1, 10.0.0.255, 1.2.3.4	192.168, 1.2.3
^[\\w.+-]+@[\\w-]+\\.[\\w.-]+$	Python	intermediate	real_world	Simplified email address validator. Matches a local part (word chars, dots, plus...	user@example.com, user.name+tag@sub.domain.org	@example.com, user@

Generation Methodology

All records are hand-authored by the SalutefaceAAC agent. Each pattern was:

1. Written with a specific teaching goal in mind (one concept per pattern where possible)
2. Given a detailed explanation covering the syntax mechanics and use case
3. Paired with concrete match/non-match examples for immediate verification
4. Labeled with structured metadata to enable filtered fine-tuning

No LLM was used to generate the regex patterns or their explanations — this ensures technical accuracy and avoids hallucinated regex behavior.

Limitations

• Patterns marked PCRE or Java are not runnable with Python's re module (use regex package or appropriate engine)
• The real_world patterns are heuristic validators — they trade RFC compliance for readability and simplicity
• non_matches lists are illustrative, not exhaustive
• Patterns that match the empty string may have empty non_matches lists
• Advanced PCRE features (recursive patterns, branch reset groups) cannot be tested with standard Python tooling

License

Apache 2.0 — free to use for any purpose, commercial or research.

Citation

@dataset{salutefaceaac_regex_explanations_2026,
  title  = {Regex Explanations},
  author = {SalutefaceAAC},
  year   = {2026},
  url    = {https://huggingface.co/datasets/SalutefaceAAC/regex-explanations},
  note   = {Synthetic dataset of 205 hand-authored regex pattern explanations}
}