openalphalab/gdelt-news

GDELT News Reconstructions

• Content: multilingual news text reconstructed from GDELT Web News NGrams 3.0, including Type 1 and Type 2.
• Format: Zstandard-compressed Parquet only, with small manifests and a coverage checkpoint.
• Columns: date, language, source_url, text, observation_id, type, metadata.
• Metadata: source-file timestamp, source checksum and reconstruction diagnostics. Observation IDs are stable for the same source group. No country, publisher, author or external enrichment.
• Coverage: recent files get priority while historical backfill moves backward toward 2020-01-01 00:01 UTC. Previously uploaded history stays. Coverage is incomplete: see progress.json for the separate live and backfill cursors and row counts.
• Updates: recent news does not wait for backfill. The worker checks new files between short historical chunks, polls every 30 seconds when idle, and uses a one-minute safety margin. No 48-hour enrichment delay; upstream publication, processing and uploading add latency.
• Empty checks: no rows means no Parquet or manifest file. Only the shared progress.json checkpoint changes; missing recent files are retried for up to 24 hours. Polling does not create duplicate observations.
• Retrieve a timestamp: for 20200101014700 (2020-01-01 01:47 UTC), select Parquet filenames whose START-END range covers it, then filter metadata.source_minute == "20200101014700". Python example below.
• Dates: date is GDELT's observation time; metadata.source_minute identifies the input file. Neither is a verified publication date.
• Download: browse Parquet files, or use the streaming and download examples. Public downloads need no token.
• Quality: reconstruction is best effort; text may be incomplete, reordered or repeated. Both types retain diagnostics. Publisher content rights still apply.
• Retention: keep published history; stop new uploads before the configured 7 TB guard. No automatic oldest-data deletion.
• Code and details: GitHub repository · Full dataset guide · GDELT source specification.

Retrieve a timestamp with Python

• Change stamp to the UTC source timestamp you need.
• Downloads only matching shards and includes both reconstruction types. Only already-published timestamps are available.

pip install huggingface_hub pyarrow

from pathlib import PurePosixPath
from huggingface_hub import HfApi, hf_hub_download
import pyarrow as pa
import pyarrow.dataset as ds
import pyarrow.parquet as pq

repo = "openalphalab/gdelt-news"
stamp = "20200101014700"  # Change this UTC source timestamp
api = HfApi(token=False)
revision = api.repo_info(repo, repo_type="dataset").sha
tables = []
for file in api.list_repo_files(repo, repo_type="dataset", revision=revision):
    if not (file.startswith("data/") and file.endswith(".parquet")):
        continue
    start, end, *_ = PurePosixPath(file).stem.split("-")
    if start <= stamp <= end:
        local = hf_hub_download(repo, file, repo_type="dataset", revision=revision)
        tables.append(pq.read_table(
            local, filters=ds.field(("metadata", "source_minute")) == stamp
        ))
if not tables:
    raise LookupError("No published shard covers this source timestamp.")
result = pa.concat_tables(tables)
print(result.num_rows)
print(result.slice(0, 1).to_pylist())

Filter by datetime and language

• Run the timestamp example above first. This filters its downloaded observations by date and language.
• Times are UTC: start is inclusive and end is exclusive. Change stamp in the first example to retrieve a different source minute.
• Use the exact language code, for example en, fr, zh or zh-TW. This example filters the selected source minute, not the entire archive.

from datetime import datetime, timezone
import pyarrow as pa
import pyarrow.dataset as ds

start = datetime(2020, 1, 1, 1, 47, tzinfo=timezone.utc)
end = datetime(2020, 1, 1, 1, 48, tzinfo=timezone.utc)
language = "en"

observed_at = ds.field("date").cast(pa.timestamp("us", tz="UTC"))
filtered = ds.dataset(result).to_table(filter=(
    (observed_at >= start)
    & (observed_at < end)
    & (ds.field("language") == language)
))
print(filtered.num_rows)
print(filtered.slice(0, 1).to_pylist())