strickvl/gtmo-military-commissions

GTMO Military Commissions Dataset

This dataset is a structured, research-oriented snapshot of public military commissions docket records from mc.mil. It combines public docket metadata collected from mc.mil, a locally hash-checked archive of public source PDFs, direct PDF text extraction, page-level quality metadata, and transcript structure derived from the PDF text and word positions.

The most important thing to know is that the raw PDFs are the source files for this dataset's derived tables. The tables are designed to help you search, filter, audit, and navigate those PDFs. Extracted text, OCR-candidate flags, visual-redaction candidates, speaker roles, and transcript-line rows are research aids. They are not official court annotations and should not be treated as legal findings.

This is the v0.2 public release, prepared on 2026-07-12. It adds a separately auditable OCR-attempt table while preserving every v0.1 table and the raw PDF archive. It is not an automatically updating mirror of the official docket.

At a glance

Public config / path	Rows / files	Unit of one row	Use this when you want to...
manifest	43,205 rows	Discovered docket entry	Search the broad docket inventory, including unavailable rows and categories outside v0.1 extraction.
download_ledger	40,441 rows	v0.1 download attempt	Audit which in-scope PDFs downloaded successfully and why some did not.
documents	40,273 rows	Successfully extracted PDF	Work at document level with title/date/page counts and quality summaries.
pages	711,348 rows	PDF page	Search extracted page text and find pages that likely need OCR or manual review.
transcript_turns	363,139 rows	Heuristic speaker-turn segment	Analyze transcript dialogue in larger speaker-like chunks.
transcript_lines	3,796,976 rows	Reconstructed visual transcript row	Work with line-like transcript rows derived from PDF word geometry.
ocr_pages	71,667 rows	OCR attempt on a candidate PDF page	Audit OCR text, attempts, routing decisions, and selected LightOn results for 69,660 candidate pages.
raw/	40,273 PDFs	Source PDF file	Inspect, cite, or verify the original public PDF.

All seven table configs use a single split named train. That split is just Hugging Face's table container here. No train/test/validation partition was created.

What is included

This release includes:

• public docket metadata discovered from the military commissions site;
• Court Filings and Transcripts in the v0.1 download and extraction scope;
• raw public PDFs for successfully downloaded in-scope files;
• document-level and page-level direct PDF text extraction;
• page-quality metadata for OCR triage, mostly black pages, blank-like pages, visual redaction candidates, and suspicious text layers;
• speaker-turn-like transcript segments;
• transcript rows reconstructed from PDF word geometry;
• byte counts and SHA-256 hashes that connect derived rows to exact source-PDF bytes;
• 71,667 OCR attempts covering 69,660 candidate pages, with attempt-level provenance and conservative policy-v2 routing.

What is not included

This release does not include:

• OCR text merged into or replacing pages.clean_text;
• reconstructed sealed, unavailable, or source-missing documents;
• extracted text tables for Trial Exhibits or Allied Papers;
• confirmed redaction-span annotations;
• court-certified transcript line citations;
• a canonical person, counsel, witness, or speaker table;
• data from non-public, hidden, or administrative endpoints.

OCR remains separate from direct PDF extraction. Use pages for the original extracted text and quality signals, and ocr_pages for OCR attempts and policy-v2 selections.

How the tables relate

The tables form a pipeline from broad inventory to source PDFs to derived text:

manifest
  -> download_ledger
      -> raw PDFs
      -> documents
          -> pages
          -> transcript_turns
          -> transcript_lines
          -> ocr_pages (candidate pages only)

The transitions matter:

1. manifest inventories all discovered docket records, including unavailable rows and categories outside this release's extracted scope.
2. download_ledger covers available v0.1 candidates: Court Filings and Transcripts.
3. Only ledger rows with status == "ok" have raw PDFs in raw/ and corresponding rows in documents.
4. pages expands each extracted document into page-level text and quality records.
5. transcript_turns and transcript_lines are derived only from transcript PDFs.
6. ocr_pages contains all retained OCR attempts for the 69,660 candidate pages. It does not overwrite pages.clean_text.

Important join rules:

• file_id is the primary cross-table join key.
• pdf_sha256 identifies the exact source PDF bytes used for extraction.
• page_number is 1-based in page and transcript-line tables.
• turn_id and line_id are release-derived identifiers, not official court identifiers.
• case_id is the source-system case identifier; it is not a chronology or legal classification.
• local_path, where present, is build-time local metadata. It is useful for provenance but is not a Hugging Face download path.

A row can disappear between stages for ordinary reasons. For example, a manifest row may have no download_ledger row because it is unavailable or outside the v0.1 categories. A download_ledger row may have no documents row or raw PDF because its status is failed or remote_not_pdf.

Quickstart

Install the packages you need:

pip install datasets huggingface_hub pandas pyarrow

Load a small document-level table:

from datasets import load_dataset

repo = "strickvl/gtmo-military-commissions"
documents = load_dataset(repo, "documents", split="train")
print(documents.num_rows)  # 40273
print(documents[0])

For large tables, prefer streaming, column projection, or direct Parquet reads. This example streams only a few columns from pages and uses a Parquet filter to find OCR candidates:

from itertools import islice
from datasets import load_dataset

repo = "strickvl/gtmo-military-commissions"
ocr_pages = load_dataset(
    repo,
    "pages",
    split="train",
    streaming=True,
    columns=["file_id", "case_id", "page_number", "ocr_decision", "clean_text"],
    filters=[("ocr_decision", "==", "candidate_for_ocr")],
)

for row in islice(ocr_pages, 3):
    print(row["file_id"], row["page_number"], row["clean_text"][:120])

Load selected policy-v2 OCR text separately from direct extraction:

ocr_attempts = load_dataset(repo, "ocr_pages", split="train")
selected_ocr = ocr_attempts.filter(lambda row: row["selected_for_public_text"])
print(selected_ocr.num_rows)  # 67931

A page may have no selected OCR result. This is intentional for 1,684 unverified fallback pages and 45 pages with no usable OCR attempt.

Direct Parquet loading also works well for analysis in pandas, Polars, or DuckDB. When reading a large table such as pages, use column projection and filters where possible:

import pandas as pd

pages = pd.read_parquet(
    "hf://datasets/strickvl/gtmo-military-commissions/pages/pages.parquet",
    columns=["file_id", "case_id", "page_number", "ocr_decision", "clean_text"],
    filters=[("file_id", "==", 4568)],
)

print(pages[["file_id", "page_number", "ocr_decision"]])

Common workflows

Join document metadata to page text

Use an inner join when you only want successfully extracted PDFs:

import pandas as pd

repo_path = "hf://datasets/strickvl/gtmo-military-commissions"

docs = pd.read_parquet(
    f"{repo_path}/documents/documents.parquet",
    columns=["file_id", "case_id", "designation", "description", "file_date", "pdf_sha256"],
    filters=[("file_id", "==", 4568)],
)

pages = pd.read_parquet(
    f"{repo_path}/pages/pages.parquet",
    columns=["file_id", "pdf_sha256", "page_number", "clean_text", "ocr_decision"],
    filters=[("file_id", "==", 4568)],
)

joined = pages.merge(
    docs,
    on=["file_id", "pdf_sha256"],
    how="inner",
)

print(joined[["file_id", "page_number", "designation", "ocr_decision"]].head())

Use a left join from manifest when missingness is part of your analysis. For example, start from manifest if you want to count unavailable records, Trial Exhibits, Allied Papers, or rows that were not part of the v0.1 extraction scope.

Retrieve and verify a source PDF

Raw PDFs are archive files, not a load_dataset(..., "raw") config. The path formula is:

def raw_pdf_path(case_id: int, file_id: int) -> str:
    bucket = file_id // 1000
    return f"raw/{case_id}/{bucket:03d}/{file_id}.pdf"

This example downloads one known OK PDF and verifies both byte count and SHA-256 against the public download ledger:

from hashlib import sha256
from pathlib import Path

import pandas as pd
from huggingface_hub import hf_hub_download

repo = "strickvl/gtmo-military-commissions"
file_id = 4568

ledger = pd.read_parquet(
    f"hf://datasets/{repo}/download_ledger/download_ledger.parquet",
    columns=["file_id", "case_id", "status", "n_bytes", "pdf_sha256"],
)

row = ledger.loc[ledger["file_id"].eq(file_id)].iloc[0]
assert row["status"] == "ok"

remote_path = raw_pdf_path(case_id=int(row["case_id"]), file_id=file_id)
assert remote_path == "raw/34/004/4568.pdf"

local_path = Path(hf_hub_download(repo_id=repo, repo_type="dataset", filename=remote_path))

hasher = sha256()
with local_path.open("rb") as f:
    for chunk in iter(lambda: f.read(1024 * 1024), b""):
        hasher.update(chunk)

assert local_path.stat().st_size == int(row["n_bytes"])
assert hasher.hexdigest() == row["pdf_sha256"]
print(local_path)

There is no raw PDF path for rows whose download_ledger.status is failed or remote_not_pdf.

Work with transcripts

Use transcript_turns for larger speaker-like chunks:

from datasets import load_dataset

repo = "strickvl/gtmo-military-commissions"
turns = load_dataset(repo, "transcript_turns", split="train")

for row in turns.select(range(3)):
    print(row["speaker_role"], row["page_start"], row["page_end"], row["text"][:120])

Use transcript_lines when visual line structure matters:

from datasets import load_dataset

repo = "strickvl/gtmo-military-commissions"
lines = load_dataset(
    repo,
    "transcript_lines",
    split="train",
    streaming=True,
    columns=[
        "file_id",
        "page_number",
        "transcript_line_number",
        "line_number_confidence",
        "speaker_label_candidate",
        "text",
    ],
    filters=[("line_number_confidence", "==", "high")],
)

first_line = next(iter(lines))
print(first_line)

For citation-like work, treat page_start / page_end and the source PDF as safer anchors than line spans. Line numbers in this release are derived from extraction and geometry, not certified by the court.

Table guide

manifest: broad docket inventory

Use manifest when you want to understand the discovered docket, including unavailable records and categories that v0.1 did not extract.

Rows: 43,205

Category counts:

Category	Rows
Court Filings	40,853
Transcripts	1,729
Trial Exhibits	524
Allied Papers	99

Availability counts:

Availability	Rows
available	41,061
unavailable	2,144

The source file_date values in the manifest range from 2002-02-07 to 2026-07-02. Treat these as docket metadata dates, not necessarily document-creation dates.

Common fields:

• file_id: source-system file identifier and primary join key.
• case_id, case_name, case_folder: source case identifiers/names.
• designation, description, file_date, filed_by: docket metadata.
• category: one of Court Filings, Transcripts, Trial Exhibits, Allied Papers.
• source_url: public URL recorded for the docket row.
• is_available, availability_reason: whether the source presented a downloadable public file.
• retrieved_at: when the manifest row was collected.

download_ledger: download provenance for v0.1 PDFs

Use download_ledger when you need to know whether an in-scope PDF downloaded, how many bytes it had, which hash was extracted, or why it is absent from the raw archive.

Rows: 40,441

Status counts:

Status	Rows	Meaning
ok	40,273	PDF downloaded, validated, hashed, and included in raw/archive extraction.
remote_not_pdf	111	Public URL returned content that was not a PDF, such as HTML or an external video embed.
failed	57	Download ended in a terminal source/network error after retry policy.

Common fields:

• file_id, case_id, category, source_url: source row identity.
• final_url, http_status, content_type: HTTP result details.
• status, error_type, error_message: terminal download outcome.
• n_bytes, n_pages, pdf_sha256: source-PDF identity for OK rows.
• manifest_snapshot_sha256, downloader_version: build provenance.
• local_path, part_path: local build paths; not portable Hugging Face paths.

documents: one row per extracted PDF

Use documents for document-level filtering, counting, and joining to page or raw-PDF records.

Rows: 40,273

Category counts:

Category	Rows
Court Filings	38,562
Transcripts	1,711

Common fields:

• file_id, case_id, case_name, category, designation, description, file_date, source_url: source metadata.
• pdf_sha256, n_bytes, n_pages: exact source-PDF identity and size.
• text_chars, clean_text_chars, word_count, clean_word_count: document-level text counts.
• avg_text_chars_per_page, min_page_text_chars, p10_page_text_chars: page-density summaries.
• likely_needs_ocr: compatibility summary based on page-level OCR decisions.
• ocr_candidate_page_count, mostly_black_page_count, blank_like_page_count, visual_redaction_candidate_page_count: page-quality aggregates.
• document_quality_status, quality_flags: coarse quality summary.
• local_path: local raw-PDF cache path used during the build; not a Hub path.

pages: one row per PDF page

Use pages to search page text, inspect page-level quality, and choose pages for OCR or manual review.

Rows: 711,348

OCR-decision counts:

ocr_decision	Pages	Interpretation
not_needed	536,455	Direct PDF text looked usable enough for this release's rules.
candidate_for_ocr	69,654	OCR may improve the page because text is absent or likely unreliable.
not_useful_likely_fully_redacted	64,383	Rendered page is mostly black/dark; OCR may not recover meaningful text.
not_useful_likely_blank	28,602	Rendered page appears mostly blank/white.
manual_review_before_ocr	12,254	Page has warning signs where human review is safer before scaling OCR.

Important text fields:

• raw_text: direct text returned by the PDF extractor.
• clean_text: conservative cleaned text. This does not include OCR text.
• text_chars, clean_text_chars, word_count, clean_word_count: page text counts.
• direct_text_status: ok, low_text, or zero_text.

Important quality fields:

• text_layer_quality_status: ok, not_assessed, suspect_garbled, or likely_bad_ocr_layer.
• ocr_decision, ocr_reason: page-level OCR triage metadata.
• mostly_black_page, blank_like_page: rendered-page appearance flags.
• visual_redaction_candidate_count: candidate visual black-bar/dark-band count.
• redaction_text_hit_count: extracted text hit count for words/markers such as REDACTED.
• quality_flags: pipe-delimited reasons that explain the page status.

transcript_turns: speaker-turn-like transcript chunks

Use transcript_turns to analyze transcript dialogue in larger chunks. The parser detects speaker labels and groups following text into turn-like records.

Rows: 363,139

Line-span status counts:

line_span_status	Turns	Meaning
none	348,433	No usable line span was parsed.
partial	14,706	Some line information was parsed, but not enough for citation-grade spans.

Common fields:

• turn_id: derived identifier in the form <file_id>:<sequence>.
• file_id, case_id, case_name, designation, file_date, pdf_sha256: provenance.
• speaker_raw: speaker label as parsed from transcript text.
• speaker_role: coarse inferred role, useful for exploration but not verified identity.
• page_start, page_end: safer location anchor for this release.
• line_start, line_end, line_span_status, line_number_confidence: non-citation-grade line metadata.
• citation_grade_line_span: currently false for this release.
• text: parsed turn text.

ocr_pages: auditable OCR attempts

Use ocr_pages when direct PDF text is missing or unreliable and you want OCR text with exact attempt and routing provenance.

Rows: 71,667 attempts across 69,660 candidate pages.

Policy-v2 page outcomes:

outcome	pages	public selection
select_lighton_clean	66,693	LightOn first attempt selected
select_lighton_guarded	1,189	LightOn selected with manual review recommended
select_lighton_retry	49	Clean 8,192-token LightOn retry selected
unverified_fallback_no_public_text	1,684	Tesseract attempt preserved, nothing selected
no_usable_ocr	45	Attempts preserved, nothing selected

One row is one attempt, so a page can have multiple rows. selected_for_public_text is true for at most one attempt per page. Every attempt remains available for auditing, including rejected LightOn output and Tesseract fallback text. No Tesseract fallback is automatically selected in policy v2 because final manual review found automatic fallback selection unreliable.

Join to pages using file_id, pdf_sha256, and 1-based page_number. Inspect the raw PDF before quoting or citing OCR text.

transcript_lines: geometry-derived visual rows

Use transcript_lines when you want line-like transcript records. These rows are reconstructed from PDF word positions, not OCR and not plain text alone.

Rows: 3,796,976

Quality counts:

Signal	Count
Transcript PDFs processed	1,711
Transcript pages processed	145,817
Rows with detected transcript line numbers	3,128,852
High-confidence numbered rows	3,030,593
Pages with strong or promising reconstruction	137,602 / 145,817

Common fields:

• line_id: derived identifier in the form <file_id>:<page_number>:<physical_line_index>.
• file_id, case_id, case_name, designation, file_date, pdf_sha256: provenance.
• page_number: 1-based PDF page number.
• physical_line_index: 1-based visual-row index on the page.
• transcript_line_number, line_number_text, line_number_confidence: detected line-number metadata.
• speaker_label_candidate: possible speaker label at the start of the row.
• text: row text after removing the detected left-margin line number.
• raw_row_text: reconstructed visual row before removing that number.
• bbox_x0, bbox_y0, bbox_x1, bbox_y1: row bounding box in PDF page coordinates.

These rows are useful for navigation and research, but they are not court-certified citations by default.

Field conventions and interpretation notes

Identifiers

• file_id is the safest join key across public tables.
• pdf_sha256 links derived text to exact source-PDF bytes.
• case_id comes from the source system.
• turn_id and line_id are generated by this release.

Text fields

• raw_text is the extractor's direct page text.
• clean_text applies conservative cleanup such as banner removal and simple hyphenated line-break joining.
• OCR text is available separately in ocr_pages; it is never silently substituted into pages.clean_text.

Nulls and optional values

• Missing dates, line numbers, speaker labels, and error fields should be treated as unknown, not as meaningful zeros.
• Empty strings in pipe-delimited fields such as quality_flags mean no flag was recorded.

Build-time paths

• local_path and part_path record local build paths from the extraction/download machine.
• Raw PDFs on Hugging Face use raw/<case_id>/<bucket>/<file_id>.pdf, not local_path.

Raw PDF archive

Raw PDFs are stored under:

raw/<case_id>/<bucket>/<file_id>.pdf

where:

bucket = file_id // 1000

zero-padded to three digits.

Examples:

raw/34/004/4568.pdf
raw/50/067/67777.pdf

The bucketed layout is required because Hugging Face rejects repositories that put more than 10,000 files in a single directory. Case 35 has 13,730 OK PDFs, so the archive cannot use a flat raw/35/<file_id>.pdf layout.

The raw archive contains only the 40,273 successfully downloaded PDFs. The 168 non-OK v0.1 rows remain visible in download_ledger.

Provenance and verification

The build used public docket pages and public PDF URLs from mc.mil. It did not use private, hidden, or administrative endpoints.

Before extraction, every local OK PDF was checked for:

• file existence;
• byte count match;
• basic PDF sanity;
• SHA-256 match against the ledger.

After raw upload, the remote archive was checked by listing and sample download-back:

• remote raw PDF count: 40,273;
• .part files: 0;
• flat full-archive raw paths: 0;
• bad bucket paths: 0;
• 50 sampled remote PDFs downloaded back and matched local byte counts and SHA-256 hashes.

The 50-file sample is evidence that the remote archive pathing and representative objects match the ledger. It does not mean every remote PDF was re-downloaded and rehashed after upload.

Quality metadata: what it can and cannot tell you

The quality fields help users decide which pages to trust, inspect, or OCR. They are extraction and page-appearance metadata, not legal findings.

Signal	What it supports	What it does not establish
ocr_decision == "candidate_for_ocr"	Prioritizing pages where OCR may improve text.	That every word in current text is unusable.
text_layer_quality_status == "likely_bad_ocr_layer"	Finding pages whose embedded text layer looks garbled.	The correct replacement text.
mostly_black_page == True	Finding pages that are mostly dark/black when rendered.	Confirmed redaction or recoverable hidden text.
blank_like_page == True	Finding pages that appear mostly blank.	Source intent or legal significance.
visual_redaction_candidate_count > 0	Finding pages worth visual review for black bars/dark bands.	Number, legal meaning, or certainty of redactions.
redaction_text_hit_count > 0	Finding extracted words/markers like REDACTED.	Redaction geometry or confirmed concealed content.
line_number_confidence	Ranking transcript-line reconstruction quality.	Court-certified line-citation accuracy.
speaker_role	Coarse transcript exploration.	Verified speaker identity.

For quotation, citation, or close reading, inspect the raw PDF.

OCR status

OCR was run for 69,660 candidate pages. LightOnOCR-2-1B was the first-pass engine, with a narrowly permitted longer LightOn retry and local Tesseract fallback attempts. The table preserves 71,667 total attempts.

Policy v2 selects 67,931 LightOn results. It deliberately selects no public text for 1,684 non-empty but unverified Tesseract fallback pages and 45 pages where no usable OCR remained. Tesseract attempts are still present for research and manual adjudication. OCR output is machine generated and may contain omissions, repetition, wrong reading order, or unsupported text, especially around scans, images, and redactions.

Known limitations

• This is a snapshot of public docket records discovered during the build. It is not a live mirror of mc.mil.
• The dataset does not reconstruct unavailable, sealed, withdrawn, or source-missing documents.
• Trial Exhibits and Allied Papers are present in manifest, but not extracted into documents, pages, transcript_turns, or transcript_lines in v0.1.
• Some PDFs have poor or garbled embedded text layers. Use page-quality fields before trusting page text at scale.
• Visual redaction fields are candidate signals only.
• Transcript speaker turns are heuristic parser output.
• Transcript line rows are geometry-derived and confidence-scored, not court-certified.
• The raw PDF archive contains only successfully downloaded PDFs. Non-OK rows remain documented in download_ledger.
• Source metadata such as file_date, designation, and description comes from the public docket and may reflect source-system inconsistencies.

Responsible use and public-record context

These records may contain allegations, testimony, personal information, descriptions of violence, institutional terminology, and disputed claims. Appearance in a filing or transcript does not establish that a claim is true.

The dataset reflects what the military commissions system published, including its categories, terminology, omissions, redactions, unavailable rows, and publication choices. Do not interpret an unavailable row as proof that a document was sealed or deliberately withheld unless the source metadata explicitly supports that conclusion.

Extracted text can lose layout, handwriting, signatures, images, stamps, marginalia, and redaction context. Machine-generated redaction, OCR, speaker, and line-number signals should not be used to infer concealed facts, verified identities, culpability, risk, or legal status.

For authoritative legal, historical, or journalistic use, compare the relevant derived rows with the raw PDF and, when necessary, the current official docket.

Suggested uses

This dataset may be useful for:

• searching public military commissions filings and transcripts;
• studying docket chronology and document availability;
• building retrieval systems over public legal records;
• identifying pages that likely need OCR;
• auditing extracted text against source PDFs;
• analyzing transcript structure at page, turn, or visual-row level;
• linking public-record text back to exact source-PDF hashes.

License and rights

The source records are public military commissions docket materials from mc.mil. No new copyright claim is made over the source government records.

The derived metadata and extraction outputs are provided for research and public-record access. Users should independently assess rights and obligations for their intended use, especially if redistributing source PDFs or large derived subsets.

The Hugging Face metadata field is set to other because this release combines public source records with derived extraction metadata rather than a single simple license grant.

Citation

If you use this dataset, please cite the Hugging Face dataset repository and the public military commissions docket source.

@dataset{gtmo_military_commissions_dataset,
  title = {GTMO Military Commissions Dataset},
  author = {Strick van Linschoten, Alex},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/strickvl/gtmo-military-commissions}
}

Source website:

https://www.mc.mil/

Version history

v0.2 public release — 2026-07-12

Added the ocr_pages config with 71,667 preserved OCR attempts across 69,660 candidate pages. Policy v2 selects 67,931 LightOn results and leaves 1,729 uncertain or unusable pages without selected public OCR text. Direct extraction tables and raw PDFs are unchanged.

The release is pinned as Hub tag v0.2.

v0.1 public release — 2026-07-10

Initial public release with manifest, download ledger, document rows, page rows, transcript turns, transcript lines, and raw PDFs. OCR has not yet been run.

This release is pinned on the Hugging Face Hub as tag v0.1, so examples can be made stable by passing revision="v0.1" to load_dataset(...) or hf_hub_download(...).

Contact

Questions or issues can be opened on the Hugging Face dataset page or directed to @strickvl.