deka-bulk

Judgments and orders of the Supreme Court of Thailand (คำพิพากษาศาลฎีกา), as published on the court's search site deka.supremecourt.or.th. Working dataset of the korakot/deka project.

• Records: 105,591 across 69 years (B.E. 2501 to 2569).
• Collected: 10 November 2026, by searching each year with an empty keyword and saving every results page. Each year's record count was checked against the count the site itself reports.
• Status: years 2501 to 2569 are complete; 2463 to 2500 (about 27,700 records) are still to be added, see below for where to get them meanwhile.

Loading

from datasets import load_dataset
ds = load_dataset("korakot/deka-bulk", split="train")          # everything
ds = load_dataset("korakot/deka-bulk", data_files="data/2567.parquet", split="train")   # one year

or with pandas:

import pandas as pd
df = pd.read_parquet("hf://datasets/korakot/deka-bulk/data/2567.parquet")

If the dataset is set to private, sign in first (hf auth login) with an account that has access.

Files

Path	Content
data/<year>.parquet	one row per record, the fields below
raw/<year>.tar	the results pages of that year exactly as fetched (gzipped HTML, 20 records a page)
ids/<year>.tsv	docid, label, page and position in the search results

Fields

Field	Type	Meaning
docid	int	the site's document id; the key of a record
year	int	Buddhist-era year of the case number
label	string	as shown on the site, e.g. "คำพิพากษาศาลฎีกาที่ 6511/2567"
doc_type	string	คำพิพากษาศาลฎีกา, คำสั่งคำร้อง, คำวินิจฉัย, ...
case_no	string	e.g. "6511/2567"; can be a range such as "8122 - 8123/2567"
short_text	string	the headnote (ย่อสั้น); paragraphs separated by newlines
long_text	string	the full text (ย่อยาว); empty for many older records, which have only a headnote
laws	list	sections the site links the record to: law_id, law_name, law_abbr, section, paragraph, sub, display
litigants	list	role (โจทก์, จำเลย, ...) and name
judges	list	names of the panel
lower_courts	list	first-instance and appeal courts with their judges, as shown
source	string	แหล่งที่มา, e.g. กองผู้ช่วยผู้พิพากษาศาลฎีกา
department	string	แผนก, e.g. ภาษีอากร, แรงงาน; mostly empty
black_no	list	Supreme Court case-file numbers
lower_court_nos	list	first-instance case-file numbers
remark	string	หมายเหตุ
page, pos	int	where the record sat in the year's search results

Things to know

• One case number can have several records. They come from different sources and differ in content. Use docid, not case_no, as the key.
• laws uses the site's own ids (ป05-01 is ประมวลกฎหมายวิธีพิจารณาความอาญา). They are not the Council of State's law codes.
• Names of private persons are mostly reduced to initials by the court in recent years, but not in older records. Treat the texts as containing personal data.
• Licence: public domain (CC0 1.0). Judicial decisions are not copyright works under Section 7 of the Thai Copyright Act B.E. 2537, and whatever rights exist in this compilation and its added structure are waived under CC0 1.0. This says nothing about personal data, which stays the user's responsibility.

Years before B.E. 2501: where to get them meanwhile

This dataset does not yet hold the oldest judgments: about 27,700 records from B.E. 2463 to 2500, roughly a fifth of the court's collection. Until they are added here, they can be taken from another open dataset, phoneee/thai-legal-corpus (CC-BY 4.0, built in March 2026), which holds the Supreme Court's judgments of all years as plain text.

import re
from datasets import load_dataset

LABEL = re.compile(r"^(คำพิพากษาศาลฎีกา|คำสั่งคำร้อง|คำวินิจฉัย(?:\(คำสั่ง\))?|คำสั่งศาลฎีกา)ที่\s*(.+?)/(\d{4})")

def before_2501(row):
    if row["source"] != "supreme_court":
        return False
    m = LABEL.match(row["text"])
    return bool(m) and int(m.group(3)) <= 2500

old = load_dataset("phoneee/thai-legal-corpus", split="train", streaming=True).filter(before_2501)
for row in old.take(3):
    m = LABEL.match(row["text"])
    print(m.group(1), m.group(2) + "/" + m.group(3), len(row["text"]))

Its val and test splits hold a further tenth of the records, so load those too for full coverage.

What differs from the rows here, so the two can be used together with care:

	This dataset	thai-legal-corpus
A record	18 separate fields	one text string: the label, "หลักกฎหมาย:", the headnote, the full text, the judges and the trailing fields run together
Key	docid, the court's document id	doc_id made from the case number; records sharing a case number share it
Cited laws	the court's own links, with law id, section and paragraph	found by pattern matching in the text (metadata/deka_citations.jsonl)
Year	year, Buddhist era	year, Christian era (B.E. minus 543), best effort; the label at the start of text is more reliable
Up to	the collection date above	March 2026
Licence	CC0	CC-BY 4.0: credit it when you use it

To line the two up, parse the label at the start of text into doc_type and case_no, as the code above does. Do not mix rows of the same year from both sources, or records will be counted twice.

How it was collected

The court's search site returns every record of a year when searched with an empty keyword and that year as both start and end, 20 records a page. Each results page already carries every field of its records, so the pages were saved as fetched (raw/) and parsed into the rows in data/. The number of records found for each year equals the number the site itself reports for that year. Collection ran from a Thai address, because the site does not answer from outside Thailand.

To re-parse or check a record against the source, untar raw/<year>.tar: each pNNNN.html.gz is one results page, and a record starts at <li class="clear result">.

Known gaps and quirks

• Years before B.E. 2501 may be missing; see Status above, the table below for what is present, and the section above for where to get them meanwhile.
• long_text is empty for many older records. The site has only a headnote for them.
• remark is empty in every record so far; the site may load it separately.
• department, black_no and lower_court_nos are empty in most records.
• A few labels are not judgments: คำสั่งคำร้อง, คำวินิจฉัย and คำสั่งศาลฎีกา are orders and rulings.
• The same judgment text can appear under more than one docid when two sources published it.

Records per year

Year (B.E.)	Records
2569	120
2568	374
2567	675
2566	451
2565	528
2564	459
2563	1,187
2562	442
2561	626
2560	690
2559	856
2558	1,574
2557	977
2556	832
2555	1,067
2554	739
2553	1,330
2552	1,119
2551	1,373
2550	1,303
2549	1,335
2548	1,664
2547	1,317
2546	1,042
2545	1,958
2544	1,828
2543	2,214
2542	2,503
2541	2,280
2540	3,532
2539	3,051
2538	4,817
2537	3,840
2536	2,967
2535	2,908
2534	2,182
2533	4,313
2532	3,985
2531	3,444
2530	3,362
2529	2,825
2528	1,218
2527	1,600
2526	1,760
2525	1,843
2524	1,656
2523	1,625
2522	1,705
2521	1,030
2520	1,221
2519	1,291
2518	1,211
2517	1,039
2516	913
2515	782
2514	910
2513	858
2512	1,067
2511	1,519
2510	1,007
2509	1,061
2508	995
2507	544
2506	1,178
2505	720
2504	502
2503	837
2502	676
2501	734