RaunakSeksaria/anlp-project-datasets

ANLP project datasets

Data for the IIIT Hyderabad ANLP project Event-Aware Representation Learning for Neutral Multi-Source News. All sets are built from BigNewsAlign events that have a full left / centre / right trio of articles. Every neutral text was generated by nvidia/nemotron-3-ultra-550b-a55b at temperature 1.0.

config	task	train	validation	test
reframing_left	one left article → the same article without framing	8,247	885	868
reframing_right	one right article → the same article without framing	8,200	893	907
neutralised_across_articles	left + centre + right articles on one event → one neutral article	8,000	1,000	1,000
reframing_left_long	one left article (up to 1,024 tokens) → the same article without framing	7,989	964	1,008
reframing_right_long	one right article (up to 1,024 tokens) → the same article without framing	7,989	964	1,008

from datasets import load_dataset
left = load_dataset("RaunakSeksaria/anlp-project-datasets", "reframing_left")
pairs = [(r["source"], r["left_neutralised"]) for r in left["train"]]

Records are copied verbatim from the project's data branch (manifest.json records the commit and the SHA-256 of each source file), so the split field inside a record says val where the dataset split is called validation.

reframing_left / reframing_right

One file per stance; every input article appears exactly once (duplicates are judged ignoring case and punctuation; a merged record's event_ids lists every event the article appeared in).

field	meaning
source	model input: TITLE=> … . ARTICLE=> …
left_neutralised / right_neutralised	model target: the input article rewritten without framing
left_original / right_original	article_id, publisher, title, body of the input
edits	the generator's own change log (diagnostics only)
event_ids	every BigNewsAlign event the article belongs to
story_id, split, generator	event id, split, generation settings

• Sources. Every unique article in the events behind neutralised_across_articles (5,695 left, 6,158 right, in those events' splits), plus extra articles from the rest of BigNewsAlign added to train only (left 3,742 + 563, right 3,842), shortest first, screened so that no near-copy (word 5-gram Jaccard ≥ 0.8) of any validation or test article entered train. Validation and test are exactly the held-out events of neutralised_across_articles.
• Publishers. Left: hpo, nyt, dailykos, wpo, cnn; right: breitbart, fox, wat. No publisher appears on both sides.
• Quality (full audit in reports/reframing_{left,right}.eda.json): no parse errors, duplicate ids or split leakage. Loaded attribution verbs per 1,000 words drop 0.81 → 0.48 (left) and 0.95 → 0.46 (right); about 93% of named entities are kept; median output/input length is 0.88 (left) and 0.92 (right).
• Known issues. About 12% of left outputs (5.5% right) contain "The article reports / notes …"; about 22% add a title or first name the input lacks; about 1% add a date or year from world knowledge. A handful of records are truncated or carry JSON fragments; their ids are in the reports.
• Triplet training. 3,963 train articles share no event with an article from the other side, so they cannot have a same-event opposite-side partner.

neutralised_across_articles

One record per event (10,000 events, story_id unique).

field	meaning
left, center, right	the event's three input articles: article_id, publisher, title, body
source	the three articles concatenated in TITLE=> … ARTICLE=> … form
neutral	one neutral article carrying the union of the facts
story_id, split, generator	event id, split, generation settings (prompt v3)

The neutral article has a median of 215 words (90th percentile 285) against a median of 438 words across the three inputs. Centre publishers are hill, usatoday and ap.

reframing_left_long / reframing_right_long (data/long/)

The same task and record format as reframing_left / reframing_right, but longer articles, drawn at random from the whole BigNewsAlign export rather than its shortest events. Every input and every target is at most 1,024 LED tokens (allenai/led-base-16384, title and body together). Over-length articles were dropped, never truncated.

• Paired by event. Both configs cover the same 9,961 events in the same order (story_id = event id). Every article has an opposite-side partner about the same event, so every record can serve as a triplet anchor.

• Sampling. Events were visited in a seeded random order (seed 42), and the first 10,000 that passed deduplication were taken. The sample therefore keeps the pool's length distribution: median input 583 tokens (left) / 531 (right), and 59% / 63% of targets are under 512 tokens. The deduplicated pool of eligible events has 72,370 events.

• Deduplication. No text appears twice, and no left/right pair is the same wire copy. No article is a near-copy (word 5-gram Jaccard ≥ 0.8) of another in the set, or of any side_neutral_full validation/test article.

• Release filtering. 37 events were dropped because a target exceeded 1,024 tokens, and 2 as near-copies. The two sides of an event are always dropped together.

• Splits. By event, from a hash of the event id (about 80/10/10). No input text appears in two splits. The split field says validation here, not val.

• Not disjoint from the other configs. These events were sampled independently; only the E2 validation/test articles were excluded as text. 330 events also appear in a short config (reframing_* or neutralised_across_articles):

  • 260 are train in both;
  • 33 are long-validation and short-train;
  • 33 are long-test and short-train;
  • 3 are long-train and short-validation;
  • 1 is long-validation and short-test.

  If you train on the short configs and test on the long ones (or the reverse), drop those events first.

• Quality (reports/long/reframing_{left,right}.eda.json):

  • loaded attribution verbs per 1,000 words: 0.79 → 0.55 (left), 0.99 → 0.59 (right);
  • intensifiers per 1,000 words: about 0.22 → 0.07;
  • 89–90% of quotes verbatim;
  • 92–94% of named entities kept;
  • median similarity of target to input: 0.81 (left), 0.83 (right).

• Known issues. The same as the short configs, a little stronger on long inputs:

  • "The article reports …": 12.9% (left), 7.4% (right);
  • adds a title or first name the input lacks: about 28%;
  • adds a date or year: about 2%;
  • edits is not a reliable measure of how much changed.