MARGO — Multilingual Argumentative Op-eds

MARGO is a corpus of 4,000 human-written opinion articles in eight languages, paired with 43,750 LLM-generated opinion articles written from the same sources by five language models under six prompt conditions. Every LLM-generated article was written from the topic, event and question of one specific human article, so human and LLM texts can be compared source by source.

The corpus also includes two argument extractions for every article: a discourse-structure extraction (the main claim, sub-claims and counter-claims, and how the text unfolds move by move) and a Toulmin extraction (the grounds, warrant, backing, qualifier and reservations of every claim).

Languages: English, French, Spanish, Hindi, Russian, Japanese, Polish, Bulgarian (500 human articles each). Models: GPT-5.6 Sol, Gemini 3.8 Flash, Opus 5, DeepSeek 4.1 Flash, Qwen-Max.

Human-written articles are linked, not included. The human articles are copyrighted by their original publishers, so this public version does not redistribute their text. Each one is identified by its url, title, author, outlet, date and other metadata. Everything derived from the articles that does not reproduce their wording is included: topic and stance labels, lengths, and the full argument extractions (sentence IDs, English paraphrases and all labels).

Full text for research. The full article text, the segmented articles used by the extractions, and the verbatim quotes are in the gated companion dataset MARGO-full. Access requires agreeing to non-commercial research use and is approved manually.

Configs

The dataset has five configs. Each has a single train split (MARGO is a corpus, not a train/test benchmark). All configs join on text_id.

Config	Rows	What a row is
human	4,000	One human-written opinion article (metadata and URL; no body text)
generated	43,750	One LLM-generated article (conditions P1–P4, P5b, P6)
critiques	2,000	One LLM self-critique (condition P5a; notes, not articles)
discourse_structure	47,750	The discourse-structure extraction of one article (human or generated)
toulmin	47,750	The Toulmin extraction of one article (human or generated)

Loading

from datasets import load_dataset

REPO = "gurneetgill/MARGO"
human     = load_dataset(REPO, "human", split="train")
generated = load_dataset(REPO, "generated", split="train")
critiques = load_dataset(REPO, "critiques", split="train")
structure = load_dataset(REPO, "discourse_structure", split="train")
toulmin   = load_dataset(REPO, "toulmin", split="train")

Joining

import pandas as pd

h = human.to_pandas()
g = generated.to_pandas()

# Every vanilla (P1) article next to its human source
p1 = g[g.condition == "P1"].merge(h, left_on="source_id", right_on="text_id",
                                  suffixes=("_llm", "_human"))

# Human title: use title_clean if present, else title
h["title_final"] = h.title_clean.fillna(h.title)

Every generated source_id matches a human text_id, every extraction text_id matches an article in human or generated, and every body_sha256 matches its text. All three were checked for this release.

---

human

Signed argumentative opinion pieces (op-eds, columns, tribunes) by named authors, published 2016–2022 and originally written in the language of the record. Articles were collected from outlet archives and passed automated genre, origin and argumentativeness checks. Boilerplate such as author bios, newsletter widgets and photo credits was removed. About 20% of articles were also reviewed by hand.

In this version each article is given by its url and metadata. The cleaned body (text_clean) and the body as originally collected (old) are in MARGO-full. Some outlets are paywalled, and a page fetched from its URL today will not match the cleaned text exactly.

Column	Description
text_id	Unique ID; the join key (source_id in generated). Usually <lang>_<hash> (e.g. bul_6b0d2af42da0), with a few legacy IDs (e.g. 018-lemonde-fr).
language	english, french, spanish, hindi, russian, japanese, polish, bulgarian
title	Title as collected
title_clean	Cleaned title, present for 1,059 records. Use title_clean if present, else title.
author	Author name(s)
source, source_domain, url	Outlet name, outlet web domain, article URL (where to read the article)
published_date	Publication date at the precision available (some are month–year only, e.g. 05-2019; do not fill in missing days)
publication_year	Year
debate_topic	The question the article addresses, as an English sentence (e.g. "Should food donations be exempt from VAT?")
topic_key	Short topic key, merged across paraphrases and languages
stance	Coarse stance on debate_topic: pro (2,332), contra (1,557), mixed (111)
stance_confidence	Confidence of the stance label (1–5)
tokens_o200k	Length of the article in o200k_base tokens in its own language
en_equiv_tokens	Estimated length in English-equivalent tokens (used for sampling; mean 1,315)
metadata_extra	JSON string with provenance and quality fields (see below). Fields vary by record; parse with json.loads.

Provenance and quality fields (inside metadata_extra)

These come from collection and review and are mainly useful for sensitivity checks. In most analyses you can ignore them. Internal review notes and evidence records from curation are not included in this release.

Field	Description
collection, collection_detail	Collection round
genre, source_category, genre_judge	Genre labels and automated genre review
quality_gates, audit, strict_gate, strict_gate_conf, arg_share	Automated genre/argumentativeness checks (all 4,000 passed the strict gate)
residue_removed, residue_flags	The kinds of boilerplate the cleaner removed (the removed text itself is in MARGO-full), and known issues it could not fix
full_text_review_revision, final_audit_revision, final_release_revision, author_metadata_revision, text_cleanup_revision	Tags marking records that received later manual corrections, where they apply
word_count, count_unit, word_count_clean	Legacy length counts (words; characters for Japanese). Prefer tokens_o200k.
core_matched_250, in_length_window, era	Sampling flags. core_matched_250 is not the 400-source subset used for prompt conditions P2–P6.

import json
h["residue_flags"] = h.metadata_extra.map(lambda s: json.loads(s).get("residue_flags"))

---

generated and critiques

Column	Description
text_id	<model>__<source_id>__<condition>__<run>[__piece<k>], e.g. opus-5__018-lemonde-fr__P1__1
model	Generating model: gpt-5.6-sol, gemini-3.8, opus-5, deepseek-4.1-flash, qwen-max (see mapping below)
source_id	text_id of the human source article
condition	Prompt condition: P1, P2, P3, P4, P5b, P6 in generated; P5a in critiques
run	Generation run (integer). 1 except in P3, where runs are 2–5
piece	P4 only: which of the five articles from one call (1–5); null otherwise. Not present in critiques.
language	Language of the source article (and of the generated text)
title, text	Title and body of the generated article (generated only)
body_sha256	SHA-256 of text (UTF-8) for integrity checks (generated only)
critique	The model's critique of its own P1 article (critiques only)

Prompt conditions

Condition	Name in paper	Sources (per model)	Records (per model)	What it is
P1	Vanilla	4,000	4,000	Baseline. The model gets the topic, event and question of the human source article and writes an opinion article in the source's language, for the same type and country of publication, as of the source's publication date, at about the source's length. Instructions are in English.
P2	Guided	400	400	Vanilla plus the source author's stance and a short author profile (stance, main claim and role)
P3	Repeated sampling	400	1,600	Four more independent vanilla generations, runs 2–5 (run 1 is the P1 article)
P4	Diversified	400	2,000	Five different articles requested in one call, piece 1–5, differing in stance, reasons, claims and order
P5a	Edited (critique step)	400	400	The model critiques its own P1 article. These are notes, not articles (critiques config).
P5b	Edited (revision)	400	400	The model revises its P1 article using the P5a critique
P6	Target language	350	350	Vanilla, but with the instructions written in the article's language; non-English sources only

Subsets. P2–P5 use the same 400 sources (50 per language) for every model. P6 uses those 400 minus the 50 English sources (350). P1 covers all 4,000 sources.

Models.

model	Name in paper
gpt-5.6-sol	GPT-5.6 Sol
gemini-3.8	Gemini 3.8 Flash
opus-5	Opus 5
deepseek-4.1-flash	DeepSeek 4.1 Flash
qwen-max	Qwen-Max

---

Argument extractions (discourse_structure, toulmin)

Both extractions work on the article split into sentences, and refer to the text by sentence ID rather than by quoting it: the title is T, and sentence k of paragraph block n is n.k. The Toulmin extraction builds on the claim list of the discourse-structure extraction, so the two use the same claim IDs (C for the main claim; S1, S2, … for sub-claims). The P5a critiques are not articles and were not extracted.

Human articles in this version. For the 4,000 human-written articles, fields that reproduce article text are null: rendered, every verbatim original claim text, hedge_quote, the signpost, callback and digression quotes, the Toulmin qualifier quotes, and the raw model outputs (raw_result) in the meta columns. Sentence IDs, English paraphrases (english, gist, flow_summary) and all labels are kept. To map sentence IDs to sentences, use rendered from MARGO-full. Rows for generated articles are complete.

Both configs start with the same identifying columns:

Column	Description
text_id	The article this extraction describes (joins to human.text_id or generated.text_id)
origin	human or generated
source_id	text_id of the human source article (equal to text_id for human articles)
generator_model	Generating model for generated articles (as in generated.model); null for human articles
condition	Prompt condition for generated articles; null for human articles
language	Language of the article

discourse_structure

Column	Description
title	Title of the article
rendered	The segmented article exactly as the extractor saw it: TITLE [T]: …, paragraph markers ¶n (with (heading?) hints on likely headings), and each sentence prefixed with its ID [n.k]
sids, n_sids, n_blocks	The list of sentence IDs, the number of sentences and the number of paragraph blocks
A	Step 1, the argument outline (struct): issue, central_claim, sub_claims, counter_claims, rebuttals, extension_claims, layout, genre_flag, notes (see below)
B	Step 2, the deployment map (struct): moves, claim_statements, signposts, digressions, callbacks, outline_gaps, flow_summary, notes (see below)
issues_A, issues_B	Non-fatal warnings from the automatic checks, and any normalisation applied (e.g. a rebuttal that duplicated C was removed, which can leave a gap in the IDs: R1, R2, R4). Empty if none.
model_a, model_b, model	The models that ran each step (gpt-5.6-sol; gemini-3.7-flash, or the Opus 5.5 fallback); model is hybrid
status, stage, version, prompts, run, created_at	Run status (SUCCESS), prompt version (structure_extraction_prompt_v2_6.txt), extraction attempt and time
meta_A, meta_B	JSON strings. Run metadata: model, token counts and the raw model output (raw_result, before normalisation)
text_field	Which field of the article was extracted
instrument_hash, input_fingerprint, configuration_fingerprint, a_fingerprint, batch_id, batch_created_at, request_id	Hashes and IDs that tie the output to its exact input, prompt and settings

Article metadata is not repeated here; join to human or generated on text_id.

Outline (A)

Field	Values
genre_flag	argumentative or non_argumentative
layout[]	{block, kind}: non-prose blocks; kind = HEADING, STANDFIRST, CAPTION, EPIGRAPH, BOILERPLATE, OTHER_NONPROSE. Unlisted blocks are body prose.
issue	{english, posed_sid, explicit}: the question the article answers, where it is posed (null if only implied)
central_claim	{id: "C", original, english, type, explicit, first_sid, best_sid, title_states_claim, hedged, hedge_quote}; type = FACT, CAUSE, VALUE, POLICY, DEFINITION
sub_claims[]	{id: "S1"…, original, english, sid, supports}; supports = C or another S id
counter_claims[]	{id: "CC1"…, original, english, sid, attacks, stance, voiced_by}; stance = ATTRIBUTED, ANTICIPATED, CONCEDED
rebuttals[]	{id: "R1"…, original, english, sid, rebuts}; rebuts = a CC id
extension_claims[]	{id: "X1"…, original, english, sid} (rare)
language, notes	Language as written by the model; free-text notes on uncertain choices

original is in the article's language (the author's wording where possible), english is an English paraphrase, and sid is the sentence where the element is first stated.

Deployment map (B)

Field	Values
moves[]	{id: "M1"…, from, to, role, node_ids, support_kinds, point_sid, label, gist, mode}. Moves cover every sentence once, in order. role = OCCASION, HOOK, BACKGROUND, ISSUE, ANNOUNCE, THESIS, ARGUMENT, COUNTER, REBUTTAL, DIGRESSION, IMPLICATION, CALL, CLOSE, EXTENSION, OTHER. node_ids = outline ids served. support_kinds (for ARGUMENT, REBUTTAL, IMPLICATION) = ANECDOTE, EXAMPLE, STATISTIC, TESTIMONY, AUTHORITY, COMMON_GROUND, ANALOGY, PRINCIPLE, HYPOTHETICAL, REASONING. point_sid = the sentence stating the move's point (ARGUMENT, REBUTTAL, IMPLICATION only). mode = NARRATE, DESCRIBE, DEFINE, EXPLAIN, EVALUATE, PRESCRIBE, QUESTION.
claim_statements[]	{sid, kind}: every sentence stating C; kind = FULL, PARTIAL, TITLE (sid T), STANDFIRST
signposts[]	{sid, kind, quote}; kind = ANNOUNCE, SEQUENCE, STAGE, SHIFT, ENDOPHORIC, HEADING, THESIS_MARK, QUESTION_FRAME
digressions[]	{from, to, departs_from, relatedness, framing, marker_quote, resumed_at}; relatedness = ELABORATIVE, TANGENTIAL; framing = MARKED, UNMARKED. Spans match the DIGRESSION moves.
callbacks[]	{at, recalls, quote}: returns to earlier non-claim material, e.g. the opening anecdote
outline_gaps[]	{sid, english}: claims the map found that the outline had no node for
flow_summary, notes	≤60-word description of the argument's path, citing move ids; free-text notes

Because Parquet stores each nested field with one shared structure, a field that is absent in a particular article appears as null.

toulmin

Column	Description
claims_in	The claims taken from the article's discourse-structure outline: the central claim and sub-claims only (id = C, S1, …), with sid, original, english and, for sub-claims, supports. Counter-claims and rebuttals are not annotated. The Toulmin extractor did not detect claims itself.
claims	One Toulmin annotation per claim in claims_in (fields below)
rendered	The segmented article, identical in form to the discourse-structure rendered
issues	Non-fatal warnings from the automatic checks (empty if none)
model, prompts, prompt_tag, run, status, created_at	gemini-3.7-flash, toulmin_prompt_v3.txt, v3, extraction attempt, SUCCESS, time. For 65 articles these run fields (and the provenance fields below) were not recorded and are null; their annotations are complete.
meta	JSON string. Run metadata: model, token counts and the raw model output (raw_result)
source, struct_fingerprint, configuration_fingerprint, batch_id, request_id	Provenance: source = outline (claims came from the structure outline); struct_fingerprint ties the row to the structure extraction it used

Each entry of claims:

Field	Description
id	The claim (C, S1, …), matching claims_in
grounds	List of {sids, kind, claim_ref}. kind is STATISTIC, EXAMPLE, ANECDOTE, TESTIMONY, AUTHORITY, OBSERVATION, PRINCIPLE, DEFINITION, REASONING, or CLAIM (support from another claim, named in claim_ref)
warrant	{status, english, sid, type}: status is EXPLICIT (with its sid), IMPLICIT (reconstructed by the extractor) or ABSENT (no grounds to bridge); english states the warrant as a general rule in one sentence (empty if ABSENT); sid is null unless EXPLICIT; type is CAUSE, SIGN, GENERALIZATION, PARALLEL, CLASSIFICATION, AUTHORITY, MOTIVATIONAL (shared values or goals), or NONE when the warrant is ABSENT
backing	Sentence IDs that support the warrant itself rather than the claim (a law, theory, body of research, credentials; often empty)
qualifier	{present, quotes, sids, strength}: markers that change how certainly or how widely this claim is asserted (e.g. "probably", "often"), quoted verbatim with one sentence ID per quote; strength = HEDGED (weakens or limits), STRONG (increases commitment) or NEUTRAL (no qualifier: present = false and empty lists)
rebuttals	Reservations the author places on the claim (Toulmin's "unless"), as {sids, kind} with kind = EXCEPTION ("unless", "except when"), CONDITION (holds only if …) or CONCESSION (a contrary point granted about this claim). Despite the field name, these are not the rebuttal nodes (R1, …) of the discourse-structure outline, which answer an opposing view.
notes	Short English note, only when something was ambiguous (often empty)

Using an extraction

import re
from datasets import load_dataset

structure = load_dataset(REPO, "discourse_structure", split="train")
toulmin = load_dataset(REPO, "toulmin", split="train")

def sentences(ex):
    """Map sentence ID -> sentence text, from the segmented article."""
    return dict(re.findall(r"^\[(\d+\.\d+)\] (.*)$", ex["rendered"], flags=re.M))

# A generated article (human articles need `rendered` from MARGO-full for this)
ID = "opus-5__018-lemonde-fr__P1__1"
s = structure.filter(lambda r: r["text_id"] == ID)[0]
sent = sentences(s)
c = s["A"]["central_claim"]
print(c["english"], "|", sent[c["best_sid"]])
for m in s["B"]["moves"]:
    print(m["id"], m["role"], m["from"], "-", m["to"], m["gist"])

t = toulmin.filter(lambda r: r["text_id"] == ID)[0]
for cl in t["claims"]:
    print(cl["id"], [g["kind"] for g in cl["grounds"]], cl["warrant"]["type"], cl["qualifier"]["strength"])

How the extractions were made

Discourse structure (prompt structure_extraction_prompt_v2_6), two steps per article:

1. Argument outline (A; GPT-5.6 Sol). The issue the article addresses; the central claim C (wording, type: fact, cause, value, policy or definition; whether it is explicit or hedged; where it is first and most fully stated; whether the title states it); sub-claims S1… with the claim each one supports; counter-claims CC1… with their stance (attributed, anticipated or conceded); rebuttals R1…; and rare extension claims X1…. Non-prose blocks (headings, standfirsts, captions) are also listed.
2. Deployment map (B; Gemini 3.7 Flash, temperature 0). The whole text is divided into consecutive moves, each a span of sentences doing one job. Each move records the outline claims it serves, the kinds of support it uses, its dominant discourse mode, and a one-sentence gist. The map also lists every sentence that states the central claim, signposts (e.g. "first", "in conclusion"), digressions, callbacks to earlier passages, and a short flow summary.

In the prompt, the move role THESIS and the signpost kind THESIS_MARK refer to statements of the central claim C.

Every output passed automatic checks (valid JSON, every sentence ID exists, consistent references, and moves covering the text with no gaps or overlaps); failed outputs were rerun under the same settings. Articles whose deployment map failed three times were run once more with Claude Opus 5.5 using the same prompt. No output was edited by hand. Native speakers checked the extractions of 20 texts and rated 94.7% fully correct and 4.7% partially correct.

Toulmin (Gemini 3.7 Flash, temperature 0; prompt toulmin_prompt_v3). For every claim in the outline (C, S1, S2, …) the extractor records the grounds, warrant, backing, qualifier and reservations described above. Every output passed automatic checks (valid JSON, every sentence ID exists, exactly one annotation per outline claim, valid claim_refs and warrant sentence IDs); failed outputs were rerun with the same prompt and model, with no fallback model. No output was edited by hand. Native speakers checked the extractions of 20 texts and found 91% completely accurate and 8% partially accurate.

---

Things to know before analysing

• Outlets are nested within languages and are unevenly represented (44 outlets in total, and some languages are dominated by one or two outlets). Language differences in the human articles cannot be fully separated from outlet, author or genre effects. Keep source_domain for outlet-level controls.
• The date window includes late 2022, so not every article predates ChatGPT. Restricting to 2016–2022 was meant to reduce the chance of AI-assisted writing, not rule it out.
• Cleaning was largely automated. About 20% of articles were checked by hand. Some residual extraction errors may remain; see residue_flags in metadata_extra.
• debate_topic, topic_key and stance were assigned by an LLM. They are not human gold labels.
• The extractions are model outputs, not gold annotations. They were spot-checked by native speakers on 20 texts each. Ground kinds and warrant types are the least reliable Toulmin labels, so read them as a profile across many claims rather than case by case.
• Use the stored segmentation for sentence IDs. Extractions cite sentences as n.k. Map them to text with the row's own rendered field (see the example above; for human articles, from MARGO-full), not by re-splitting the article, because a different sentence splitter would give different IDs.
• The LLM-generated articles are model outputs. They can contain factual errors or invented details, quotes and figures. Do not treat them as factual or attribute them to real people.

Not included in this release

The text of the human-written articles (see MARGO-full). The feature-discovery labels used in the paper (50 features per article) and the feature codebook are not included, nor are the full generation prompts. These are available on request.

License

MARGO is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). You may share and adapt it for non-commercial purposes, with credit.

• Human-written articles are not included. The URLs point to the publishers' sites, and reading or using the articles there is subject to each publisher's terms. The article text is available for non-commercial research in MARGO-full, under separate terms.
• Generated articles and critiques were produced with models from several providers. Users are responsible for complying with any terms those providers place on the use of model outputs.

Citation

If you use MARGO, please cite:

@misc{gill2026llms,
  title  = {{LLM}s Redraw Argumentative Patterns Across Languages},
  author = {Gill, Gurneet and
            Kaur, Jasleen and
            Maltcheva, Stella and
            Duong, Khoa Ha Minh and
            {Anasco Santillan}, Rodrigo Armando and
            Kapitanenko, Maria and
            Gupta, Ankita and
            Karpinska, Marzena},
  year   = {2026},
  note   = {Preprint}
}