ReGKG-2025

An archive-derived sample of 281,620 article records from 4,301 sources for the missing GDELT Global Knowledge Graph interval. Records occupy all 1,665 target batches from 14 June 2025, 18:00 through 2 July 2025, 02:00 UTC, inclusive. The publication window ends at 2 July 2025, 02:15 UTC, exclusive.

This README covers collection, validation, uses, and loading. See SCHEMA.md for field definitions and missingness. Exact counts, processing identities, and hashes are recorded in manifest.json.

The record count is approximately 14% of the volume estimated from native GKG source frequencies before and after the gap. The coverage section below explains the reference sources and calculation. Batch labels are publication times rounded down to 15 minutes. Processing and first-availability times are unknown; no artificial delay is applied.

Uses

Use for retrospective analysis of recovered reporting, daily content summaries with date-precision checks, comparisons among represented sources, and testing GKG-reading pipelines. Keep native and reconstructed records distinguishable. Absolute news counts, complete event coverage, and historical real-time availability cannot be inferred from this sample.

Files

File	Contents
data.zip	The 1,665 per-batch .gkg.csv.zip files plus data/masterfilelist.txt.
record_provenance.parquet	One row per released GKG record: URL, source, archive/capture, publication/batch times, extraction precision/method, language, translation/manual-review flags, and compact processing-profile linkage.
archive_provenance.parquet	Used archive observations: Wayback replay addresses or Common Crawl WARC paths and byte ranges, with locator evidence.
sources.csv	All sources observed in the native reference, with retrieval quotas, candidate and output counts, estimated volume, attained fractions, and quota fill.
source_day_weights.parquet	Source × UTC date counts, modeled native volume, expansion and composition weights, including unsupported cells.
review_journal.parquet	All 1,648 recorded manual decision events, including reversals, superseded decisions, and excluded articles.
SCHEMA.md	All 27 GKG columns, four unavailable media fields, metadata definitions, and weighting formulas.
README.md	Dataset description, collection, validation, limitations, and loading examples.
LICENSE.txt, NOTICE.md	CC BY 4.0 legal text and attribution/third-party rights notices.
manifest.json	Release inventory, SHA-256 hashes, batch/member hashes, processing and translation profiles, exact period, and counts.
checksums.sha256	SHA-256 of every other release file, including the manifest; excludes itself.

data.zip stores the already compressed batch ZIPs without recompressing them. Extract it once to obtain data/gkg/*.gkg.csv.zip. Each inner ZIP contains exactly one UTF-8 tab-separated .gkg.csv member with 27 columns and no header. GKG ZIPs total 1,182,954,020 bytes before the outer container overhead.

The inner masterfilelist.txt lines are ZIP_bytes ZIP_MD5 relative_path, with paths relative to the release root after extraction. MD5 is retained for GKG-style compatibility; use SHA-256 in the manifest for integrity. Batch member hashes refer to uncompressed CSV bytes separately.

Read one GKG batch and join provenance

Run the examples from the parent directory containing regkg-2025-v1.0.0. They require Python with pandas and a Parquet engine such as pyarrow (pip install pandas pyarrow).

from io import BytesIO
from pathlib import Path
from zipfile import ZipFile
import csv
import pandas as pd

release = Path('regkg-2025-v1.0.0')
with ZipFile(release / 'data.zip') as outer:
    name = 'data/gkg/20250614180000.gkg.csv.zip'
    with ZipFile(BytesIO(outer.read(name))) as batch:
        with batch.open('20250614180000.gkg.csv') as stream:
            gkg = pd.read_csv(stream, sep='\t', header=None, dtype=str,
                              keep_default_na=False, quoting=csv.QUOTE_NONE)
gkg = gkg.rename(columns={0: 'record_id'})
provenance = pd.read_parquet(release / 'record_provenance.parquet')
joined = gkg.merge(provenance, on='record_id', validate='one_to_one')

Archive observations join on archive_id. Processing configurations join through processing_profile_id to the manifest. translation_profiles records the M2M100 model revision and generation settings.

Source adjustment

For each source/day, expansion_weight = estimated_native_records / reconstructed_records when both counts are positive. Multiply record contributions by this weight to obtain aggregates expanded to modeled volume on supported cells. These factors calibrate counts to the reference-based volume estimates. Article inclusion probabilities are unknown.

composition_weight instead reproduces the modeled source mixture among cells with both observed and expected records, while preserving the observed article total on that daily panel. Missing source/day cells have null weights and cannot be reconstructed by weighting other articles. See SCHEMA.md for the exact denominators; always report retained panel coverage and weight dispersion.

weights = pd.read_parquet(release / 'source_day_weights.parquet')
provenance['day'] = provenance['publication_timestamp'].dt.date
weights['day'] = pd.to_datetime(weights['day']).dt.date
weighted = provenance.merge(weights, on=['source', 'day'], validate='many_to_one')
supported = weighted[weighted['weight_status'].eq('estimable')]
# For a per-record variable x: sum(x * expansion_weight) within each day.
# For a composition-adjusted mean: sum(x * composition_weight) / sum(composition_weight).

For intraday analysis, retain only time_precision == 'minute'; day-only labels do not identify midnight publication. The limitations section below describes timestamp uncertainty and capture/publication conflicts.

Coverage and construction

Native GKG records from 1–14 June and 2–8 July 2025 provide the reference for estimating missing-period volume. The reference contains 12,459 sources. Archive retrieval targeted sources with at least 50 reference records and at least one dated record on each side of the gap, or 5,548 sources. For each source, observed record rates by weekday and 15-minute batch position were projected onto the missing intervals. The reconstructed record count equals 13.63% of the estimated volume across all reference sources and 14.28% of the estimated volume for the 5,548 retrieval sources. These percentages compare reconstruction volume with modeled expectations. The fraction of actual missing native records recovered remains unknown. The 4,301 represented sources comprise 77.52% of the retrieval sources and account for 83.94% of their reference records.

Dataset structure

The unit of observation is a reconstructed article record. GKG batches join one-to-one to record provenance on record_id; archive observations join on archive_id. Source summaries and source/day weights describe the sampling frame and supported calibration cells. The manual-review journal includes excluded entries and superseded decisions. There are no predefined training, validation, or test splits. See the file inventory above and SCHEMA.md for types, null values, field availability, and weight definitions.

Dataset creation

Curation rationale and sources

The collection supplies archive-accessible reporting for a missing native GKG interval in a compatible wire format. Surrounding-week source frequencies guide the sample allocation.

URL discovery and text retrieval used existing Wayback and Common Crawl captures. The release includes 256,617 Wayback and 25,003 Common Crawl records. Native surrounding-week source frequencies guided quotas. No source-site article scraping was used for this collection.

Processing and languages

The preparation catalog contains 285,136 retrieved entries: 281,620 eligible records, 3,408 rejections, and 108 translation failures/unsupported-language deferrals. Screening rejected nonarticle payloads, boilerplate, and corruption; cleaning normalized Unicode and whitespace. Language normalization and detection preceded translation of eligible non-English bodies. There are 8,512 translated records, using the M2M100 418M model; the full revision/settings are in the manifest. The accepted source-language labels include 65 non-English labels plus English. Text analysis uses English input or English translations; original-language GCAM can also use the original body.

Manual review

All 1,648 journal events concern 1,545 manually checked entries. The released human_verified flag is true for 619 articles accepted through manual review. Other events include rejections, deferrals, superseded decisions, and undo/reset operations. Review targeted flagged entries. Its coverage therefore supports eligibility decisions on those entries; collection-wide accuracy and per-field accuracy require separate evaluation.

Field construction and time assignment

The output uses the 27-column GKG 2.1 wire layout. Images and social-media embeds (columns 19–22) are unavailable. Text-derived themes, entities, tone, counts, GCAM, dates, quotations, and amounts use gdelt-regkg configurations recorded in the manifest.

There are 237,465 minute-precision and 44,155 day-only publication labels. Day-only timestamps represent an extracted date with an assigned clock time; source-local midnight can appear at another UTC time after conversion. Stored publication times determine the output batches. Native publication-to-batch lag in tagged neighboring data has median 129 minutes and IQR 57–240 minutes. Native processing time, reconstructed publication time, capture time, and historical upload/first availability are different clocks. No artificial delay is applied in this release. Capture/publication conflicts are preserved, and archive timestamps do not prove historical feature availability.

Validation

Intact-week backtest

The reconstruction process was tested on 2–8 June 2025, when all 672 native GKG batches are available. Source selection and quotas used surrounding weeks; native test-week URLs were excluded from archive discovery. Preparation and manual review yielded 14,437 eligible articles from 3,132 sources, separate from this dataset.

Exact source–URL matching found 5,918 unambiguous pairs: 40.99% of recovered articles and 0.700% of 845,948 native pairs in the batch-time frame. Publication time and native batch time define different windows; unmatched URLs are not automatically errors. A source-balanced subset of 5,000 articles reduced total variation distance to the surrounding-week target from 0.543 to 0.343; target-to-native-week distance was 0.074.

On matched documents, the gdelt-regkg benchmark gave tone mean absolute error of 1.690 points, concordance correlation of 0.806, and reconstructed-minus-native tone bias of +0.108 points. Micro F1 was 0.638 for themes, 0.673 for persons, 0.456 for organizations, and 0.648 for location identities. One malformed native location field was excluded.

The experiment measures agreement on recovered documents. Accuracy and recall inside the actual outage remain unknown. The experiment covers one publication week, with archive captures searched through 15 July 2025 and bounded candidate evaluation. Acquisition used adaptive budgets and an aggregate URL-overlap diagnostic. Final evaluation includes explicit manual decisions on flagged articles. The analysis of the recovered subset is exploratory. Mean tone was +0.140 in the recovered pool and −0.629 in the native week: paired field agreement does not establish population equivalence. Native GKG provides the observed comparison records; its coverage of all publications is unknown.

Biases, risks, and limitations

Archive accessibility and unequal recovery across publishers and articles affect selection. Source/day weights calibrate observed cells to a surrounding-week volume model. Article inclusion probabilities are unknown. The native reference is English-stream data; 3.02% of reconstructed records are translated. Adjacent-week content differences combine changing news events, source selection, and processing differences. Language, translation, article eligibility, and publication-date accuracy have not been independently estimated on a probability sample of the release.

Use the 237,465 minute-precision labels for intraday placement. The 44,155 day-only labels require date-interval treatment: 38,568 occupy 00:00 UTC buckets and 5,587 occupy other times after timezone conversion. The original timezone is absent from the sidecar, so exact UTC-day membership can be uncertain. Minute-precision records alone occupy all 1,665 target buckets. There are 8,062 records whose archive capture precedes the stored publication timestamp. The stored dates retain these conflicts. Archive capture and native publication-to-batch lag do not establish historical upload or first availability.

Rights and sensitive information

Article bodies and archive payloads are not distributed. URLs, personal names, geographic mentions, titles, and automatically extracted quotations are present and can identify people or reproduce excerpts. The dataset does not anonymize these fields or verify allegations in the reporting. CC BY 4.0 covers only rights held by the dataset creator; publishers and other rights holders retain rights in their material. No endorsement by GDELT, the Internet Archive, Common Crawl, or publishers is implied.

See LICENSE.txt and NOTICE.md for licensing and attribution.

Contact: Nikita A. Klimin, naklimin@edu.hse.ru, for questions, corrections, or removal requests. Include the affected record IDs or URLs.