3char

58 code systems in one schema, 751,870 codes. airport and currency codes, amino acid symbols, assembly mnemonics, media types, constellation abbreviations, HTTP status codes, and the other short identifiers that standards bodies assign. three characters is the organizing thesis, not a filter. systems with two, four or seven character codes are included, each with the regex that describes its actual form.

the collection documents how many independent authorities converged on short fixed width identifiers, and gathers them under one schema so they can be joined, validated and retrieved together.

loading

from datasets import load_dataset

# every system, one table
ds = load_dataset("brennercruvinel/3char", "all")

# one system
iata = load_dataset("brennercruvinel/3char", "trp-iata")

# the registry of systems: authority, license, regex, provenance
sysinfo = load_dataset("brennercruvinel/3char", "systems")

schema

one row per code. the arrow schema is identical across every config, which is what allows the all config to concatenate them.

column	type	description
system_id	string	<domain>/<id>, joins to the systems config
code	string	the code as the authority writes it, verbatim, no case folding
label	string	short human readable name
description	string	longer gloss when the source provides one
parent_code	string	parent code in the same system, for hierarchical systems
status	string	current, deprecated, reserved, unassigned, or null
aliases	list	other strings that resolve to the same referent
extra	string	JSON object with source specific fields not covered above

extra is stored as a JSON string rather than a struct. the sources share few fields beyond code and label, and a struct would require a union of roughly 200 nullable columns on every row.

systems

bio, biology

id	system	authority	codes	pattern
bio/aa1	amino acid 1-letter codes	IUPAC-IUB JCBN	20	^[A-Z]$
bio/aa3	amino acid 3-letter codes	IUPAC-IUB JCBN	20	^[A-Z][a-z]{2}$
bio/cdn	NCBI genetic codes (codons)	NCBI	64	^[ACGTU]{3}$
bio/ec	enzyme commission numbers	IUBMB / ExPASy	8,441	^\d+\.\d+\.\d+\.(n?\d+|-)$
bio/gen	HGNC gene symbols	HUGO Gene Nomenclature Committee	44,997	^[A-Za-z0-9@#\-\._]+$
bio/pdb	PDB entry IDs	wwPDB / RCSB	254,978	^[1-9][A-Z0-9]{3}$

cls, classification

id	system	authority	codes	pattern
cls/elm	chemical element symbols	IUPAC	118	^[A-Z][a-z]{0,2}$
cls/icd	ICD-10-CM	US CDC / NCHS	98,403	^[A-Z][A-Z0-9]{2,6}$
cls/loc	Library of Congress classification outline	Library of Congress	234	^[A-Z]{1,3}$

cmp, computing

id	system	authority	codes	pattern
cmp/asm	x86 instruction mnemonics	Intel / AMD	1,309	^[A-Z][A-Za-z0-9]{1,15}$
cmp/css	CSS functions	W3C CSSWG / MDN	105	^[a-zA-Z][a-zA-Z0-9\-]*$
cmp/dns	DNS resource record types	IANA	97	^([A-Z][A-Z0-9\-]*|\*)$
cmp/ern	POSIX errno names	IEEE / The Open Group	77	^E[A-Z0-9]+$
cmp/ext	file extensions	community	286	^[a-z0-9][a-z0-9.+\-]*$
cmp/frc	FourCC codec identifiers	community / VLC	672	^[A-Za-z0-9\-.:_]{2,4}$
cmp/git	git subcommands	git project	158	^[a-z][a-z0-9\-]{1,17}$
cmp/hex	CSS named colors	W3C CSSWG	148	^[a-z]{3,20}$
cmp/htt	HTTP methods	IANA	41	^([A-Z][A-Z\-]*|\*)$
cmp/mim	media types	IANA	2,300	^[a-z]+/[a-zA-Z0-9!#$&\-\^_\.\+]+$
cmp/reg	x86-64 register names	LLVM Project	257	^[a-z][a-z0-9.()]*$
cmp/sts	HTTP status codes	IANA	500	^\d{3}$
cmp/vim	vim motion commands	vim project	289	^\S{1,40}$

com, communication

id	system	authority	codes	pattern
com/bau	Baudot / ITA2 teleprinter codes	ITU-T	30	^[A-Z0-9\-]+$
com/mor	morse code	ITU-R	48	^([A-Z0-9\u00c9]|[^\w\s]{1,2})$
com/tlg	ABC universal commercial telegraphic code	William Clauson-Thue (1901)	7,359	^[A-Za-z]{4,22}$

fin, finance

id	system	authority	codes	pattern
fin/cfi	ISO 10962 CFI codes	SIX Group / ISO	1,216	^[A-Z]{6}$
fin/i42	ISO 4217 currency codes	SIX Group / ISO	178	^[A-Z]{3}$
fin/tck	SEC company tickers	US SEC	10,405	^[A-Z][A-Z\-]{0,6}$

grf, graph and vocabulary

id	system	authority	codes	pattern
grf/act	ActivityStreams 2.0 terms	W3C	141	^[A-Za-z][A-Za-z0-9]*$
grf/foa	FOAF vocabulary terms	FOAF project	75	^[A-Za-z][A-Za-z0-9_]*$
grf/gql	openCypher keywords	openCypher / Neo4j	63	^[A-Z][A-Z_]*$
grf/lpg	property graph schema terms	arXiv 2211.10962	34	^[A-Z]{3,9}$
grf/oid	OpenID Connect discovery fields	OpenID Foundation	17	^[a-z][a-z0-9_]*$
grf/org	W3C Organization ontology terms	W3C	45	^[a-zA-Z][A-Za-z0-9]*$
grf/rdf	RDF Schema terms	W3C	15	^[a-zA-Z][A-Za-z0-9]*$
grf/sch	schema.org types and properties	schema.org / W3C CG	2,987	^[A-Za-z0-9][A-Za-z0-9]*$
grf/shc	SHACL vocabulary terms	W3C	221	^[A-Za-z][A-Za-z0-9\-]*$
grf/sio	SIOC vocabulary terms	SIOC Project / DERI	99	^[A-Za-z][A-Za-z0-9_]*$
grf/vcd	vCard ontology terms	W3C	146	^[a-zA-Z][A-Za-z0-9\-]*$

lng, language

id	system	authority	codes	pattern
lng/b47	BCP 47 language subtags	IANA	9,295	^[A-Za-z0-9\-]+(\.\.[A-Za-z0-9\-]+)?$
lng/glt	Glottolog languoid codes	MPI-EVA Leipzig	27,177	^[a-z0-9]{4}\d{4}$
lng/i15	ISO 15924 script codes	Unicode Consortium (RA)	226	^[A-Z][a-z]{3}$
lng/i31	ISO 3166-1 country codes	ISO 3166/MA	249	^[A-Z]{2,3}$
lng/i39	ISO 639-1 language codes	ISO 639/RA	183	^[a-z]{2}$
lng/i63	BCP 47 three-letter language subtags	IANA	8,084	^[a-z]{3}$
lng/ioc	country codes composite	community / datasets.io	226	^[A-Z]{2,3}$
lng/mdy	month names	Unicode CLDR	12	^[A-Z][a-z]{2}$

nte, lexical

id	system	authority	codes	pattern
nte/wnt	English WordNet 2024 synset IDs	Global WordNet Association	120,630	^oewn-\d{8}-[nvasr]$

trp, transport and place

id	system	authority	codes	pattern
trp/iata	IATA airport codes	IATA	9,056	^[A-Z]{3}$
trp/ica2	IATA and ICAO airline codes	IATA / ICAO	5,841	^[A-Z0-9]{2,3}$
trp/icao	ICAO airport codes	ICAO	10,153	^[A-Z]{4}$
trp/olc	Open Location Code (plus codes)	Google	292	^[023456789CFGHJMPQRVWX]{4,}\+[23456789CFGHJMPQRVWX]*$
trp/tld	top level domains	IANA	1,437	^[A-Z0-9\-]+$
trp/unl	UN/LOCODE location codes	UNECE	116,213	^[A-Z0-9]{3}$

wrt, writing and time

id	system	authority	codes	pattern
wrt/abj	Hebrew script characters	Unicode / community	84	^\S$
wrt/iau	IAU constellation abbreviations	International Astronomical Union	88	^[A-Z][A-Za-z]{2}$
wrt/tz	tz database zone names and abbreviations	IANA / tzdb	806	^[A-Za-z0-9+\-_/]{1,32}$
wrt/uni	Unicode emoji sequences	Unicode Consortium	5,225	^[0-9A-F]{4,6}( [0-9A-F]{4,6})*$

provenance and licensing

there is no single license. every row of the systems config carries its own license and a license_status of ok or check. authority is kept separate from url because the body that assigns a code and the site that serves the file frequently differ, and that distinction is part of the provenance.

sources whose terms prohibit redistribution are parsed by the pipeline and not uploaded: the WHO ATC index, the BISAC subject headings, the Dewey summaries, the CUSIP mapping, the ISO 10383 MIC list and the what3words API surface. the SIL ISO 639-3 table was removed on the same basis, since its terms name the SIL site as the only authorized distribution point.

two systems are published with license_status = check: fin/cfi and fin/i42, both from SIX Group, which publishes the lists free of charge without stating redistribution terms.

only parquet is published. raw files are retrieved from url by the build repo and are not stored in either repository. this matters for the ITU recommendations and the IUPAC table, where the facts are freely usable and the documents are not freely mirrorable.

known limitations

trp/unl has 116,213 rows and about 90,000 repeated codes. a UN/LOCODE location code is unique only within its country, so ADALV and USALV both carry ALV. the country and the full locode are in extra. code is not a unique key in that system.

lng/i63 is not ISO 639-3 and is not labelled as such. it holds the three letter language subtags as published by IANA. RFC 5646 omits the three letter form whenever a two letter subtag exists, so eng, deu and por are absent and the registry carries en, de and pt instead. it also includes 115 ISO 639-5 collections and 224 deprecated subtags that ISO 639-3 does not.

com/tlg is parsed from a 1901 book scan. the plain OCR text emits the three columns (code number, code word, phrase) as separate runs of lines, which loses the pairing between them. the parser reads the djvu xml instead, where every word keeps a bounding box, and rebuilds the columns from the x coordinate. 92 percent of code words are recovered with their phrase. the remainder, and the code numbers in the left column, are limited by the quality of the scan.

cmp/sts carries 500 rows for a registry that assigns about 64 status codes. IANA publishes the gaps as ranges such as 105-199, and these are expanded to one row each. an assigned code has a description and a null status. an unassigned one has status = unassigned and an empty label. filter on status to obtain the assigned codes only.

trp/olc contains the reference encoder test vectors, not a registry. plus codes are generated from coordinates and no complete enumeration exists.

trp/iata and trp/icao are both derived from the OurAirports public domain file, read on different columns. neither is an official IATA or ICAO publication.

lng/ioc carries the ISO 3166 alpha-3 code in extra.iso3 and the FIFA code in extra.fifa. aliases holds only the codes that differ from the IOC code. release 0.1.0 placed both in aliases without labels, including a footnote marker from the source for the United Kingdom. corrected in 0.1.1.

code_pattern is enforced at build time: the build fails if any code in a system does not match its pattern. pattern_status records how the pattern was obtained. derived means the regex was written against the actual file. provisional means it is still loose and may tighten in a later release. writing these patterns identified 25 incorrect assumptions in the parsers and one malformed row upstream.

retrieval

the tables were evaluated as retrieval corpora on this parquet, using a 51 question answer key verified against the data and a grid that varies one factor at a time: chunk granularity (record, block of 20, whole system), chunk form (key=value or one sentence of prose), key form (GRU or iata:GRU) and embedder (all-MiniLM-L6-v2, multilingual-e5-small). the rig and its raw output are at github.com/brennercruvinel/3char-bench.

configuration	result
one record per chunk, prose, e5	92.2% recall@1, 100% recall@5, 22.6 tokens per answer
same, with system:code as key	88.2% recall@1, 27.0 tokens
block of 20 records	47.1% recall@1, 460 tokens
whole system as one chunk	90.2% recall@1, 7,467 tokens
prose vs the same fields as key=value	+37 points of recall@1 for 1.3x the tokens
faiss scalar quantizer, 8 bit	4x smaller index, identical recall

the main result is that code should not be indexed on its own. a literal without its system is ambiguous by construction. across the 58 published systems, case folded, 14,911 of the 623,141 distinct codes are claimed by two or more systems, and AND and CAR by eleven. given a bare literal, the retriever selects the correct system 33% of the time, which equals the combinatorial floor. with the system name in the query or in the key, routing accuracy is 99 to 100%. system_id should be carried in the chunk, in the key, or in both.

excluded systems

all 64 systems parse and validate. 6 are held back from upload because their authority prohibits redistribution: the WHO ATC index, the BISAC subject headings, the Dewey summaries, the CUSIP mapping, the ISO 10383 MIC list and the what3words API surface. their parsers are in the build repo and run locally against sources fetched by the user.

build

the parsers, the schema, the checksum manifest and the license triage with its verbatim quotes are at github.com/brennercruvinel/3char-pipeline. download.py refetches every source and fails on checksum drift. validate.py tests every code against its system's regex. raw files are not stored in that repository either, for the same licensing reasons.

changes are recorded in CHANGELOG.md, including the two sources that were replaced because the original file was wrong and the eight that were re-derived from the assigning authority.

citation

@misc{cruvinel_3char,
  title  = {3char: three character code systems in one schema},
  author = {Cruvinel, Brenner},
  year   = {2026},
  url    = {https://huggingface.co/datasets/brennercruvinel/3char}
}