Dataset cards
← ARI3 overview
ARI3 learns from, and is tested on, the items it collects and the editor's labels on samples of them. Counts are dated because the corpus grows every 4 hours.
ARI3 item corpus
Live, counted 2026-10-03 05:21 UTC
- What it is
- Dated news items collected from public RSS and Atom feeds about fashion and style.
- Sources
- 97 feeds, each mapped to a sector. 94 outlets have at least one stored item. The outlets are listed on the Sources page.
- Collection
- Every 4 hours. robots.txt is checked, requests to the same host wait one second, and items without a publish date are skipped.
- What is stored
- URL, title, publish time, fetch time, a content hash, feed language, and the feed's own summary capped at 500 characters. Full article text is never stored.
- Size
- 10,642 items, 10,101 with an excerpt. Publish dates run from 2014-07-24 to 2026-10-03. The first item was fetched 2026-09-23.
- First seen
- Since 2026-09-30 08:30 UTC, the time ARI3 first stores an item is recorded as it happens (2,687 items). For the 7,955 items stored before then, that time was reconstructed afterwards from the earliest surviving record of the item (its fetch time, a backup, a prediction row or a label). A reconstructed time can be later than the real first sighting, and it is labeled as reconstructed.
- Sectors
- Editorial 88.6%, trade 4.2%, retail 2.6%, independent criticism 1.5%, visual archive 1.3%, institutional 0.6%, runway 0.6%, designer origin 0.5%, resale 0.1%, social none.
- Languages
- By feed metadata, 78% English-tagged and 13% untagged. The rest are tagged Japanese, Italian, Portuguese or French. Some tags are wrong: a few items tagged English are in French or Italian.
- Excluded
- Social platforms are never scraped. Runway images are never stored.
- Known biases
- Editorial outlets dominate. Feeds keep only recent items, so older coverage is thin. Outlets with no feed are missing.
- License and access
- The items belong to their publishers. The corpus is kept locally for research and is not redistributed. The site shows links and short metadata only.
ARI3 style labels (is_style_signal)
350 labels, fingerprint sha256 20ca8e4a…, 2026-09-26
- Question
- Is this item about style: what people wear, how, and what it is called?
- Scope rule
- Yes: garments, silhouettes, materials, colors, styling, aesthetic terms, collections, clothing collaborations, nails. No: beauty (makeup, skincare, fragrance, hair, packaging), earnings, deals, executive moves, store openings, celebrity news with no clothing content.
- Annotator
- One: the editor. Labeled one item at a time with the ARI3 review app.
- Sampling
- 300 items drawn in turn from each sector (seed 7), so small sectors appear. 50 more chosen by active learning from items v0.0.1 was least sure about.
- Balance
- Random labels: 166 yes, 134 no. Active-learning labels: 14 yes, 36 no.
- Splits
- Fixed by a SHA-256 hash of each item at labeling time: 65% train, 15% calibration, 20% test. Active-learning labels are kept out of calibration and testing.
- Versions
- 145 labels (2026-09-23, used by v0.0.1). 350 labels (2026-09-26, used by v0.0.2).
- Known biases
- One annotator. 300 of 350 items are English-tagged. 276 of 350 were published in September 2026.
- Access
- Not published yet. The fingerprints let later releases prove which labels were used.
ARI3 time-holdout labels (EXP-003A)
150 labels, gold fingerprint sha256 7e15f2a1…, frozen 2026-10-02
- Question
- The same is_style_signal question and scope rule as the style labels.
- Items
- 150 items published after v0.0.2 was frozen (2026-09-26 19:25:40 UTC), from 2026-09-26 20:00 to 2026-09-30 04:03 UTC, drawn in turn from each sector (seed 13). An item was eligible only if its publish time and first-seen time were after the freeze and no earlier record showed it.
- Annotator
- One: the editor, 2026-10-02.
- Balance
- 82 yes, 68 no.
- Use
- Testing only. Never used for training or calibration. Frozen and committed before the evaluation ran.
- Known biases
- One annotator. 39 outlets, about 3.3 days of publication. 47% come from the editorial group (editorial, trade and independent criticism outlets), because the draw rotates through sectors.
ARI3 forecast labels (is_forecast, EXP-003B)
250 labels, fingerprint sha256 ff36351a…, 2026-09-26
- Question
- Does this item predict what will happen, instead of reporting what happened?
- Sampling
- 150 items drawn in turn from each sector and 100 items whose headline or excerpt contains a forecast word (seed 11). The forecast-word items are not a random sample.
- Balance
- 35 forecasts: 14 among the random items, 21 among the forecast-word items.
- Splits
- Fixed by the same hash rule: 161 train, 38 calibration, 51 test (5 forecasts in the test split).
- Annotator
- One: the editor.