emirms/turkish-competition-authority-decisions

Turkish Competition Authority Decisions (Rekabet Kurulu Kararları), 1997–2026

The complete published decision history of the Turkish Competition Authority (Rekabet Kurumu) — every Competition Board decision the regulator has made public, in full text, with derived structural metadata.

10,367 decisions · 113,297 pages · 323 million characters · 29 years

Every decision carries its outcome, the articles of Law 4054 it turns on, the panel that decided it (as stable pseudonymous ids — no personal names), the other decisions it cites, and whether it was challenged in court and survived.

Why this dataset

Turkish legal NLP has court-decision corpora, but the competition regulator has been absent from them. Competition decisions are a distinctive body of law: long, heavily reasoned, economically dense, and structurally regular — the median decision is 5 pages, the longest is 507.

Three properties are worth stating plainly:

1. Coverage is provable, not asserted. The count matches the regulator's own published total (Toplam : 10,367) exactly. Discovery came from the site's sitemap and was cross-checked against the paginated search index; both agree.
2. Full text, not summaries. Every record carries the complete decision text extracted from the official PDF, plus the SHA-256 of the source bytes.
3. Derived structure. Panel composition (pseudonymised), cited statutory articles, section segmentation, outcomes, a precedent citation graph and penalty outcomes are extracted from the text — fields the regulator does not publish as data.
4. Judicial-review status. Whether a decision was challenged, and whether it survived. Citing an annulled decision is a professional risk; this dataset lets you check.

Fields

Source metadata (from the regulator)

field	type	notes
karar_id	string	the authority's own UUID
decision_number	string	25-40/958-554 (modern), 46/323-34 (pre-2003)
decision_date	string	ISO YYYY-MM-DD
publication_date	string	ISO; publication lags the decision by months
decision_type	string	the regulator's own classification (5 values)
title	string	party names and subject
pdf_url	string	canonical source URL
full_text	string	complete extracted text
page_count, char_count, chars_per_page	int/float
suspected_scan	bool	true for 9 image-only PDFs with no text layer
sha256	string	of the source PDF bytes
fetched_at, source_filename	string	provenance

Derived fields (extracted from the text)

field	type	coverage	notes
outcome	string	81.2%	closed set — see below. From the operative section only
cites_decisions	list[string]	24.6%	the precedent graph: other Board decisions this one cites
case_file_number	string	98.5%	Dosya Sayısı, e.g. 2021-3-038
procedure_type	string	88.8%	Devralma, Soruşturma, Önaraştırma, Muafiyet, …
board_chair_id	string	99.7%	pseudonymous id of the presiding member
board_member_ids	list[string]	99.7%	pseudonymous ids, excluding the chair
board_member_count	int	99.7%	how many members sat
rapporteur_ids	list[string]	97.5%	pseudonymous ids of the case handlers
rapporteur_count	int	97.5%	how many rapporteurs worked the file
cited_law_articles	list[string]	95.4%	articles of Law No. 4054 only, incl. sub-article forms (40/1)
sections	list[{letter, heading}]	91.6%	lettered section structure
paragraph_count	int	55.1%	(N) markers — citable chunk boundaries
fine_imposed	bool	7.7% true	operative holding, not mere mention
fine_amounts_try	list[string]	3.1%	from the operative penalty holding; Turkish format 346.765,63 TL
appeal_status	string	9.0%	judicial-review outcome — see below. null = never challenged
appeal_case_count	int	9.0%	number of court cases against this decision
appeal_case_numbers	list[string]	9.0%	Dava Esas No of each case, for lookup
appeal_issues	list[string]	—	why a status could not be resolved
enrichment_issues	list[string]	—	per-field parse failures, explicitly recorded
repairs	list[string]	14 records	documented corrections (see below)

outcome — what the Board actually held

Derived from the operative section (SONUÇ), never from language appearing elsewhere in the document. A decision that discusses clearing conduct and then declines to is not cleared.

label	n	meaning
cleared	4,011	transaction/conduct permitted
no_action_required	2,093	no further action or investigation needed
violation_found	860	infringement established
exemption_granted	434	exemption under Art. 5 granted
not_subject_to_review	401	outside the Authority's merger-review remit
exemption_denied	321	exemption refused
negative_clearance	167	negative clearance issued
no_violation_found	96	no infringement established
investigation_opened	13	investigation initiated
opinion_issued	10	opinion rather than a decision
clearance_denied	9	transaction/conduct refused
null	1,952	operative section not found (1,180) or language not recognised (772)

Where a multi-party decision finds one respondent in violation and another not, violation_found takes precedence. cleared_with_conditions was investigated and deliberately rejected — no signal precise enough to publish was found.

The 18.8% null rate is real and deliberate. No label is guessed.

appeal_status — did the decision survive judicial review?

Board decisions can be challenged before the Danıştay and the administrative courts. A decision that was annulled is no longer good authority. This field is derived from the Authority's own published judicial-review records — 2,118 court cases across 933 challenged decisions.

label	n	share of all	meaning
null	9,434	91.0%	never challenged
upheld	547	5.3%	challenge rejected, or affirmed on appeal
annulled	225	2.2%	struck down — treat with care
unknown	93	0.9%	challenged; outcome not confidently resolvable
pending	68	0.7%	no merits ruling yet, or dismissed as moot

Of the 225 annulled decisions, 64 had originally found a violation.

How it is derived, and what it deliberately refuses to do. The label comes from the structured Karar Tipi codes on each procedural stage, walked in chronological order — never from the prose on the page. That prose is the challenged decision's own title, not a judicial outcome, and it routinely contains the word iptal when the underlying decision was itself a re-do following an earlier annulment. Reading it would mark decisions that stand as annulled.

Specific rules that matter legally:

• yürütmenin durdurulması (stay of execution) is never an annulment. It is interim relief. A third of all procedural stages in this data are stay rulings; treating them as merits outcomes would corrupt the field wholesale.
• Bozma / İstinaf Kabul (reversal and remand) reset the status to unresolved rather than resolving it — a remand means the question is open, not decided either way.
• K.V.Y.O (moot), Ehliyet (standing) and stay variants are excluded from the merits walk entirely.
• Where one decision drew several cases, the most serious status wins: annulled > partially_annulled > pending > upheld > unknown.

Known limitation. Some of the 93 unknown records are likely recoverable upheld — the classifier declines to label when a confirmation-type ruling appears before an identifiable first-instance ruling. This was left deliberately conservative: turning a cautious unknown into a wrong upheld is the harmful direction, and under-labelling is not.

partially_annulled exists in the label set but occurs zero times — exactly one procedural stage corpus-wide carries kısmen (Temyiz Kısmen Bozma Kısmen Onama).

Personal names are pseudonymised

No personal names appear in this dataset. Board members and rapporteurs are represented by stable pseudonymous identifiers of the form kisi_1a2b3c4d, and the same names have been removed from full_text.

The identifiers are stable across the corpus: the same person carries the same id in every decision they appear in. So questions like "which other decisions did this person sit on?" remain answerable, while the dataset alone does not identify anyone. The ids are salted HMAC-SHA256 digests; the salt is held locally by the publisher and is not distributed, so the mapping cannot be reconstructed from this dataset.

Verified: a full scan of all 1,258 known names against all 10,367 decision texts returns zero remaining occurrences.

A side effect worth noting: because pseudonymisation normalises case and titles, spelling variants of one person now unify. Kenan Türk / Kenan TÜRK and Murat ÇETiNKAYA / Murat ÇETINKAYA / Murat ÇEtiNKAYA resolve to a single identifier — 1,266 distinct source spellings collapse to 1,190 people.

Two honest limitations:

1. full_text is no longer byte-identical to the source PDF. The redaction is the only deviation; everything else is verbatim. The original PDFs remain publicly available from the Authority, so this dataset is not the sole record — it simply is not the vector for name-based aggregation.
2. Only Board members and rapporteurs are covered. Decision texts also name party representatives (Temsilcisi: Av. …) and occasionally carry addresses. Those were not part of the extracted fields and are not redacted. Anyone processing full_text should assume other personal data is present.

Design decisions worth knowing

cited_law_articles rejects look-alike citations. Decisions cite the Competition Act (4054), EU regulations (2016/679), court case numbers (2020/795 E.), KVKK (6698) and other Board decisions (18-08/152-73) — all of which contain article-like numbers. Each article reference is bound to its nearest preceding <number> sayılı anchor and accepted only when that anchor is 4054. Numbers above the statute's 65 articles are rejected.

fine_imposed means the Board actually imposed one. A decision that discusses a fine and declines to impose it (time-barred, for example) reads false. Phrase-presence would have produced roughly twice as many positives, most of them wrong.

cites_decisions requires an asserting citation context. A decision number is only accepted from a phrase of the form <date> tarih ve <number> sayılı … Kurul kararı. Bare number patterns are never harvested, a decision never cites itself, and every extracted citation is validated against the corpus's own set of 10,367 real decision numbers — unresolvable candidates are dropped, not kept with a maybe-flag.

Nothing is guessed. Any field that could not be parsed confidently is null/[] with a code in enrichment_issues. A wrong value in a legal dataset is worse than a missing one.

Two format eras. Pre-2003 decisions use a simpler header (Karar No, no lettered sections). They parse for board composition but legitimately yield nulls for structure — recorded as issues rather than silently empty.

Known limitations

• 9 decisions (0.09%) have no text. Image-only PDFs, all 2022–2024, flagged suspected_scan: true. No OCR was applied.
• fine_amounts_try is best-effort. Decisions also state turnover and thresholds in TL; only amounts near penalty language are extracted.
• 1 duplicate decision number. 22-44/644-276 appears under two karar_ids with different publication dates and different SHAs — a regulator republication. Both versions are retained deliberately.
• 1 uncorrectable date. 14-37/705-313 records a decision date after its publication date; a within-year error the decision number cannot resolve.
• PDF extraction artifacts occasionally split words ("Ke mal EROL"). Preserved verbatim rather than corrected, so the text stays faithful to the extraction.

Corrections applied

14 records carry a repairs label. These were validated against an independent signal — a decision cannot be published before it is made:

• 13 × year corrected from the decision number. The site's decision_date disagreed with the year encoded in the decision number. The rule resolved 4 of 5 records with logically impossible date ordering and introduced zero new violations. Applied only to the modern number shape; pre-2003 numbers encode a meeting, not a year, and are never used to rewrite a date.
• 1 × decision number recovered from text. One decision had no number in the index and an empty Karar_.pdf filename.

Composition

decision type	n
Birleşme ve Devralma (mergers)	4,900
Rekabet İhlali (infringement)	3,592
Menfi Tespit ve Muafiyet (clearance/exemption)	1,135
Diğer (other)	522
Özelleştirme (privatisation)	218

decade	n
1990s	191
2000s	2,996
2010s	4,405
2020s	2,775

Most-cited articles of Law 4054: 7 (merger control, 5,679), 4 (restrictive agreements, 3,177), 41 (investigation, 2,476), 6 (abuse of dominance, 1,849), 5 (exemption, 1,361), 16 (fines, 1,226).

Usage

from datasets import load_dataset

ds = load_dataset("emirms/turkish-competition-authority-decisions")

# what the Board actually held
violations = ds.filter(lambda r: r["outcome"] == "violation_found")
cleared    = ds.filter(lambda r: r["outcome"] == "cleared")

# by the provision it turns on
abuse   = ds.filter(lambda r: "6" in r["cited_law_articles"])   # Art. 6, abuse of dominance
mergers = ds.filter(lambda r: "7" in r["cited_law_articles"])   # Art. 7, merger control

# penalties
fined = ds.filter(lambda r: r["fine_imposed"])

# the precedent graph
edges = [(r["decision_number"], cited)
         for r in ds["train"] for cited in r["cites_decisions"]]

# decisions to treat with care
annulled = ds.filter(lambda r: r["appeal_status"] == "annulled")

Check appeal_status before relying on a decision. A decision annulled on judicial review is no longer good authority, and decision_type/outcome alone will not tell you.

Collection method

Fetched directly from rekabet.gov.tr between 2026-08-25 and 2026-08-26.

Judicial-review records fetched from the same host on 2026-08-26: 2,391 DavaDetay ids from the sitemap, of which 2,118 resolved and 273 were dead at source (that route soft-404s at HTTP 200). fetched + dead_at_source == discovered holds exactly.

Politeness: one connection, 2.0-second delay between requests, ~11,400 requests over 8h15m, a contact email in the User-Agent, Retry-After honoured. The site publishes no robots.txt, so the rate was self-imposed and deliberately conservative. Zero failed requests, zero unreachable documents.

Text extracted with pypdf. No OCR. Source PDFs retained with SHA-256 so every record is traceable to bytes.

Updates

This dataset is refreshed from the source, not a one-time snapshot. The Authority publishes new decisions continuously, and its search results are sorted publication-date descending, so a refresh only has to walk the first few index pages until it reaches one with nothing new — not re-crawl the whole corpus. Refreshes are keyed on publication order, not decision date, because publication can lag a decision by several months; keying on decision date would permanently miss late-published decisions. Each refresh also cross-checks the site's sitemap.xml id set against what is already collected, so a page-walk miss or a late backfill surfaces instead of silently disappearing. There is no version history in this repository — each publish replaces the dataset with the current complete state.

Licence and reuse

The underlying decisions are public regulatory material published by the Turkish Competition Authority. This dataset is a faithful compilation of that public material with derived structural metadata; no editorial content has been added to the decision texts.

Board members' and rapporteurs' names have been pseudonymised throughout, including in full_text (see above). Party representatives' names and some addresses remain in the decision texts and were not redacted. Users processing this data are responsible for their own compliance with applicable data protection law, including Turkey's KVKK (Law No. 6698) and, where relevant, the GDPR.

Note on Turkish copyright: FSEK Art. 31 exempts kazai kararlar (judicial decisions). Competition Board decisions are administrative, not judicial, so that provision does not straightforwardly apply. This compilation is published on the basis that the material is official public-sector information made freely available by the authority.

Versions

• v1.2 (2026-08-27) — personal names pseudonymised: board_chair/board_members/ rapporteurs replaced by stable salted ids plus counts, and the same names removed from full_text. Verified zero remaining occurrences corpus-wide. 35 fields.
• v1.1 (2026-08-26) — added outcome, cites_decisions (precedent graph), and appeal_status + companions. Tightened fine_amounts_try to the operative penalty holding and corrected fine_imposed to recognise the dative holding form (cezalandırılmasına), 730 → 795. 33 fields.
• v1.0 (2026-08-26) — initial release, 27 fields.

This is an archive of public regulator material. It is not legal advice.

Citation

@dataset{turkish_competition_authority_decisions_2026,
  title  = {Turkish Competition Authority Decisions (Rekabet Kurulu Kararları), 1997--2026},
  author = {Sezer, Emir Murat},
  year   = {2026},
  note   = {10,367 decisions, full text with derived structural metadata},
  url    = {https://huggingface.co/datasets/emirms/turkish-competition-authority-decisions}
}