aeroo11/kanal-berita

KANAL — Indonesian news headlines, labelled by section

Indonesian news headlines and RSS summaries, labelled with the kanal (section) their publisher filed them under. Collected hourly from public RSS feeds.

1,295 articles · 4 publishers · 8 classes Published between 2025-09-30 09:21:14 and 2026-07-30 11:55:49.

Built as the data layer of KANAL, a model-lifecycle platform. Every figure below is computed from the file in this repository, not asserted by hand.

What is here

One row per article. Headline and summary only — no article bodies. Nothing is stored beyond what each publisher chose to place in a syndication feed, and every row links back to the original.

column	notes
title, summary	the only fields a model should see
kanal	the label: the section the publisher filed it under
source, channel, canonical_url	provenance — see the warning below
published_at, fetched_at	UTC
is_evergreen	over 30 days old when fetched
url_leaks_label	the URL gives the label away
cluster_id, cluster_size	articles sharing a normalised headline
is_cross_source_duplicate	the same story republished by another publisher
has_label_disagreement	that story was filed under different sections
label_is_judgement_call	the channel→kanal mapping was not a direct match

Read this before training on it

The URL gives the label away on most rows

source	share of URLs containing the label
antara	2.5%
cnn	100.0%
liputan6	98.6%
republika	35.6%

Publishers put the section in the article path, or in a subdomain. Any model shown canonical_url, channel or source will score near-perfectly and have learnt nothing. Those columns are here so the leakage can be measured and reproduced, not used.

These are per-source averages, and they hide structure: Republika's rate is the mean over its three feeds, where ekonomi leaks through its subdomain at 100% while nasional and internasional are served from a neutral host and leak almost nothing. So prefer the per-row column over the source:

clean = df[~df.url_leaks_label]      # 229 of 1,295 rows

That clean subset is the control group. Training on it versus on everything and comparing macro-F1 measures what leakage manufactures, rather than assuming it. At source level the low-leakage publisher is antara.

Some feeds mix in evergreen content

source	share over 30 days old at fetch
antara	59.0%
cnn	0.0%
liputan6	11.4%
republika	0.0%

ANTARA files explainers, profiles and fixture lists into its live news feeds, so a naive temporal split pushes most of its rows into train and leaves a test set dominated by the publishers that leak. Split on published_at, and check the source mix on both sides afterwards.

Split by cluster, never by row

ANTARA is a wire agency; its stories reappear near-verbatim elsewhere hours later, with different URLs and correctly different article_keys. A random split puts one copy in train and another in test, and then measures memorisation. cluster_id groups them — partition on it.

The labels are editorial decisions, not ground truth

has_label_disagreement marks stories that different publishers filed under different sections. That disagreement rate is an empirical ceiling: no model can beat the labels' own inconsistency. label_is_judgement_call marks rows whose channel→kanal mapping was a judgement rather than a direct match — the mapping, with its reasoning, is in the repository.

Distribution

kanal	articles
gaya-hidup-kesehatan	240
ekonomi	185
internasional	185
teknologi	170
olahraga	170
hiburan	170
politik	135
hukum-kriminal	40

source	articles
antara	200
cnn	700
liputan6	350
republika	45

Collection

Hourly polling with conditional requests, one request at a time, exponential backoff, and a per-source circuit breaker. Full policy: DATA.md.

RSS is a sliding window of the last 25–100 items with no archive endpoint, so an hour not captured cannot be recovered. The dataset therefore grows forward only; it cannot be backfilled.

One publisher (CNN Indonesia) sits behind Cloudflare and answers HTTP 403 to datacentre IPs, so it is absent from data collected in CI while being present in local runs. That is a property of the collection environment, and it is recorded rather than smoothed over.

Licence and attribution

Metadata in this dataset: CC-BY-4.0.

Article headlines and summaries remain the property of their publishers — ANTARA, Liputan6, Republika and CNN Indonesia. This dataset stores syndicated metadata and links back to the originals; no article text is republished. If you are a publisher and would prefer your feeds not be collected, open an issue on the repository and the source will be removed.