Datasets:

SDeaure
/

websentiment-ml-7

segment_id string	piece_id string	segment_index int64	language string	text string	score int64	label string	source string	sampling string	round int64	format string	topic string	synthetic_target string	url string
cv-0396-0:0	cv-0396-0	0	de	Drei Wochen ohne Waschmaschine – ein Erfahrungsbericht	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:1	cv-0396-0	1	de	Am 3.	0	neutral	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:2	cv-0396-0	2	de	März ging meine Waschmaschine kaputt.	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:3	cv-0396-0	3	de	Ich rief den Kundendienst an, den ich über die Herstellerseite gefunden hatte.	0	neutral	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:4	cv-0396-0	4	de	Nach 47 Minuten in der Warteschleife (ja, ich habe mitgestoppt) wurde mir ein Termin „zwischen 8 und 18 Uhr" angeboten.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:5	cv-0396-0	5	de	Ein Zeitfenster von zehn Stunden.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:6	cv-0396-0	6	de	Ich nahm mir also einen Tag frei.	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:7	cv-0396-0	7	de	Der Techniker kam nicht.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:8	cv-0396-0	8	de	Um 17:40 Uhr rief ich an, und man sagte mir, der Termin sei „leider storniert worden".	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:9	cv-0396-0	9	de	Von wem?	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:10	cv-0396-0	10	de	Das wusste niemand.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:11	cv-0396-0	11	de	Beim zweiten Termin erschien ein freundlicher junger Mann, der kurz in die Trommel schaute, „Lager vermutlich" murmelte und mir einen Kostenvoranschlag über 389 Euro in die Hand drückte.	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:12	cv-0396-0	12	de	Für eine Maschine, die neu 450 kostet.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:13	cv-0396-0	13	de	Ich lehnte ab und bezahlte trotzdem 79 Euro „Anfahrt und Diagnose".	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:14	cv-0396-0	14	de	Als ich widersprach, hieß es, das stehe in den AGB.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:15	cv-0396-0	15	de	Natürlich steht es dort, auf Seite 11, in Schriftgröße 7.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:16	cv-0396-0	16	de	Seitdem trage ich meine Wäsche zum Waschsalon.	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:17	cv-0396-0	17	de	Ich habe gelernt: Wer repariert, verdient am Problem.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:18	cv-0396-0	18	de	Wer nicht repariert, verdient auch.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:19	cv-0396-0	19	de	Nur der Kunde verliert immer.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:20	cv-0396-0	20	de	Mein Fazit: Finger weg.	-3	very negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-0:21	cv-0396-0	21	de	Und falls Sie jemanden kennen, der Waschmaschinen wirklich reparieren kann: Bitte melden Sie sich bei mir.	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-1:0	cv-0396-1	0	de	Der Paketbote und ich: eine Liebesgeschichte, die keine war	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-1:1	cv-0396-1	1	de	Sie ahnen es schon: Mein Paket ist wieder nicht angekommen. „Zustellung gescheitert – Empfänger nicht angetroffen", steht in der App.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-1:2	cv-0396-1	2	de	Ich saß seit sieben Uhr morgens im Wohnzimmer.	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-1:3	cv-0396-1	3	de	Ich habe sogar das Treppenhaus gewischt, aus Nervosität.	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-1:4	cv-0396-1	4	de	Es ist das fünfte Mal in diesem Jahr.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-1:5	cv-0396-1	5	de	Mal liegt die Benachrichtigungskarte im Briefkasten des Nachbarn, mal wurde das Paket „an einen Nachbarn abgegeben", der angeblich keiner meiner Nachbarn ist, mal steht es laut Foto vor einer Tür, die definitiv nicht meine ist.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-1:6	cv-0396-1	6	de	Die Hotline?	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-1:7	cv-0396-1	7	de	Ein Chatbot, der mir höflich „Ich habe Sie nicht verstanden" ausgibt, bis ich selbst aufgebe.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-1:8	cv-0396-1	8	de	Man fragt sich, wie ein Unternehmen mit Milliardenumsatz es nicht schafft, einen Karton von A nach B zu bringen.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-1:9	cv-0396-1	9	de	Die Antwort ist wohl: Es muss nicht.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-1:10	cv-0396-1	10	de	Es gibt keine Alternative, also keine Konsequenzen.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-1:11	cv-0396-1	11	de	Mein Tipp: Bestellen Sie nichts, was Sie wirklich brauchen.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:0	cv-0396-2	0	de	Notaufnahme, Sonntag, 23:15 Uhr.	0	neutral	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:1	cv-0396-2	1	de	Ein Protokoll.	0	neutral	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:2	cv-0396-2	2	de	23:15 – Ankunft mit starken Bauchschmerzen.	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:3	cv-0396-2	3	de	Am Empfang: ein Zettel, „Bitte Nummer ziehen".	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:4	cv-0396-2	4	de	Der Automat ist kaputt.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:5	cv-0396-2	5	de	23:40 – Eine Schwester nimmt meine Daten auf.	0	neutral	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:6	cv-0396-2	6	de	Dreimal fragt sie nach meiner Versichertennummer, weil das System hängt.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:7	cv-0396-2	7	de	00:30 – Warteraum.	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:8	cv-0396-2	8	de	Ein Mann hustet unentwegt.	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:9	cv-0396-2	9	de	Ein Kind weint.	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:10	cv-0396-2	10	de	Der Fernseher zeigt eine Teleshopping-Sendung für Messerblöcke.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:11	cv-0396-2	11	de	02:10 – Ich frage nach, wie lange es noch dauert. „Wir sind unterbesetzt." Das weiß ich inzwischen auswendig.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:12	cv-0396-2	12	de	03:45 – Endlich ein Arzt, sichtlich übermüdet, der nach vier Minuten „vermutlich Magen-Darm, Tee und Schonkost" sagt.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:13	cv-0396-2	13	de	Keine Untersuchung des Bauches.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:14	cv-0396-2	14	de	04:20 – Entlassung ohne Entlassungsbrief.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:15	cv-0396-2	15	de	Zwei Wochen später bekam ich eine Rechnung über 60 Euro, weil „der Überweisungsschein fehlte".	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:16	cv-0396-2	16	de	Die Schmerzen?	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:17	cv-0396-2	17	de	Es war der Blinddarm, den ein anderer Arzt am Dienstag feststellte.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:18	cv-0396-2	18	de	Ich schreibe das nicht aus Wut, sondern aus Müdigkeit.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-2:19	cv-0396-2	19	de	Wer in diesem Land krank wird, braucht vor allem Geduld und Glück.	-3	very negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:0	cv-0396-3	0	de	An meine Bank (ein offener Brief in Reimen, weil mir sonst die Worte fehlen)	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:1	cv-0396-3	1	de	Sehr geehrte Sparkasse, liebes Haus,	0	neutral	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:2	cv-0396-3	2	de	Sie sehen mich bei Ihnen ein und aus –	0	neutral	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:3	cv-0396-3	3	de	Doch ein und aus wohl nur im Netz,	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:4	cv-0396-3	4	de	denn die Filiale schloss, Sie sagten: „Jetzt	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:5	cv-0396-3	5	de	spart man in Zukunft Zeit und Geld!"	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:6	cv-0396-3	6	de	Ich spar nur Nerven, die ich nicht mehr hab, in dieser Welt.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:7	cv-0396-3	7	de	Das TAN-Verfahren?	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:8	cv-0396-3	8	de	Dreimal neu.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:9	cv-0396-3	9	de	Die App?	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:10	cv-0396-3	10	de	Verlangt, dass ich mich scheu	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:11	cv-0396-3	11	de	per Foto, Video, Ausweis ausweise –	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:12	cv-0396-3	12	de	und stürzt dann ab, ganz leise.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:13	cv-0396-3	13	de	Die Gebühr für Konto, Karte, Bar	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:14	cv-0396-3	14	de	stieg zum Jahresbeginn, wie jedes Jahr.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:15	cv-0396-3	15	de	Der Kredit, den ich dringend brauchte? „Leider nein."	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:16	cv-0396-3	16	de	Der Dispo hingegen?	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:17	cv-0396-3	17	de	Teuer, aber fein.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:18	cv-0396-3	18	de	Ich schließ den Brief mit Zorn und Spott:	-3	very negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:19	cv-0396-3	19	de	Sie nennen sich Partner.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:20	cv-0396-3	20	de	Gott,	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:21	cv-0396-3	21	de	wenn Partnerschaft so aussieht, dann	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:22	cv-0396-3	22	de	leb ich lieber allein – ich fang gleich an.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:23	cv-0396-3	23	de	Mit freundlichen Grüßen, nicht mehr Ihr Kunde,	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-3:24	cv-0396-3	24	de	Brandt	0	neutral	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-4:0	cv-0396-4	0	de	Deutsche Bahn: Warum ich das Zugfahren aufgegeben habe	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-4:1	cv-0396-4	1	de	Es gab keinen Skandal, keinen Unfall.	0	neutral	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-4:2	cv-0396-4	2	de	Es war die Summe der Dinge.	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-4:3	cv-0396-4	3	de	Der ICE, der 40 Minuten verspätet war, weil „Verzögerungen im Betriebsablauf" (was soll das sein?).	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-4:4	cv-0396-4	4	de	Die reservierte Sitzplatznummer, die es im Wagen nicht gab.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-4:5	cv-0396-4	5	de	Die Klimaanlage, die im Juli ausfiel, im Januar die Heizung.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-4:6	cv-0396-4	6	de	Die Durchsage, die niemand verstand, weil das Mikrofon rauschte, und die Anschlusszüge, die immer genau dann abfahren, wenn man den Bahnsteig erreicht.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-4:7	cv-0396-4	7	de	Das Schlimmste ist die Freundlichkeit der Verspätung: Alles wird entschuldigt, nichts gelöst.	-3	very negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-4:8	cv-0396-4	8	de	Die Erstattung?	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-4:9	cv-0396-4	9	de	Ein Formular, drei Wochen Wartezeit, 4,50 Euro Gutschein.	-2	negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-4:10	cv-0396-4	10	de	Seit Mai fahre ich Auto.	0	neutral	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-4:11	cv-0396-4	11	de	Teurer, stauanfälliger, schlecht für die Umwelt, ich weiß.	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0396-4:12	cv-0396-4	12	de	Aber ich komme an.	1	somewhat positive	synthetic	synthetic_complement	2	blog post	services	-3	null
cv-0397-0:0	cv-0397-0	0	de	Wie ein Hamburger Fahrradladen mein Verhältnis zu Montagen veränderte	1	somewhat positive	synthetic	synthetic_complement	2	blog post	brands	+3	null
cv-0397-0:1	cv-0397-0	1	de	Ehrlich gesagt hatte ich nie viel für Fahrräder übrig.	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	brands	+3	null
cv-0397-0:2	cv-0397-0	2	de	Dann ging mein altes Rad kaputt, und eine Kollegin schickte mich zu „Kettenreaktion" in Hamburg-Ottensen.	-1	somewhat negative	synthetic	synthetic_complement	2	blog post	brands	+3	null
cv-0397-0:3	cv-0397-0	3	de	Schon beim Eintreten roch es nach Öl, Kaffee und frisch gebackenem Zimtgebäck.	2	positive	synthetic	synthetic_complement	2	blog post	brands	+3	null
cv-0397-0:4	cv-0397-0	4	de	Jemand hatte mir tatsächlich ein Stück hingestellt, bevor ich überhaupt etwas gefragt hatte.	2	positive	synthetic	synthetic_complement	2	blog post	brands	+3	null
cv-0397-0:5	cv-0397-0	5	de	Inhaberin Merle Brandt hörte mir zehn Minuten zu, ohne auf die Uhr zu sehen.	2	positive	synthetic	synthetic_complement	2	blog post	brands	+3	null
cv-0397-0:6	cv-0397-0	6	de	Dann schob sie ein schlichtes Stadtrad nach vorn und sagte: „Das ist dein Fahrrad.	1	somewhat positive	synthetic	synthetic_complement	2	blog post	brands	+3	null
cv-0397-0:7	cv-0397-0	7	de	Ich weiß es einfach."	1	somewhat positive	synthetic	synthetic_complement	2	blog post	brands	+3	null

End of preview. Expand in Data Studio

WebSentiment ML-7

Sentence-level sentiment on a 7-point scale (-3 to +3) for web text in six languages: English, French, Spanish, German, Italian and Portuguese.

462,272 segments from 27,964 web passages (FineWeb and FineWeb-2), labeled in context. 86 % of the pieces are real web text; the remaining 14 % are synthetic pieces, flagged in a source column, written to cover formats the web rarely provides (customer emails, ad scripts, support replies).

It is a silver dataset: no human labeled these segments. Agreement of the labeler with human annotators is measured on existing research datasets and reported below. A human gold subset is planned (see "Help build a gold subset").

Why this dataset

Most sentiment datasets are tweets, product reviews or film reviews, and most are English only. SEO and content work needs something else: the tone of a paragraph in a blog post, a product page, a forum thread or a press release, in the language of the site.

The 7-point scale, labeling each segment in the context of its full text and the taxonomy of 26 text formats follow the approach of Dejan AI's sentiment classifier, which trains on fully synthetic English data. This dataset keeps the scale and the formats, replaces most of the synthetic text with real web passages and extends to more languages.

Content

Language	Pieces	Segments	of which web	of which synthetic
English	5,674	96,346	87,089	9,257
French	6,608	109,788	90,338	19,450
Spanish	3,787	52,527	44,973	7,554
German	4,074	80,133	69,016	11,117
Italian	3,995	62,370	51,207	11,163
Portuguese	3,826	61,108	53,358	7,750
Total	27,964	462,272	395,981	66,291

Score distribution (segments):

	-3	-2	-1	0	+1	+2	+3
train, web	10,123	40,229	45,945	142,719	65,736	47,775	15,140
train, synthetic	2,137	13,822	9,446	17,129	9,653	11,077	3,027
eval_natural	191	1,498	2,913	14,114	6,054	3,095	449

The training split is deliberately enriched in strong sentiment. The eval_natural split is not: it is a random sample of 300 passages per language, drawn before any preselection, so it shows the distribution you meet on the open web (about half neutral, -3 under 1 %).

Configs and fields

Segments come in one config per language (en, fr, es, de, it, pt) and in all (default). The pieces config holds the passages.

from datasets import load_dataset
fr = load_dataset("SDeaure/websentiment-ml-7", "fr")          # French segments
pieces = load_dataset("SDeaure/websentiment-ml-7", "pieces")  # full passages

Segment configs, one row per segment:

Field	Description
`segment_id`	`<piece_id>:<segment_index>`
`piece_id`	the passage it comes from
`segment_index`	position in the passage, from 0
`language`	`en`, `fr`, `es`, `de`, `it`, `pt`
`text`	the segment (one sentence, or one short line such as a heading)
`score`	-3 to +3
`label`	`very negative`, `negative`, `somewhat negative`, `neutral`, `somewhat positive`, `positive`, `very positive`
`source`	`web` or `synthetic`
`sampling`	how the piece was selected (see below)
`round`	collection round: 0 (French only), 1, 2
`format`	one of 26 text formats, or `other`
`topic`	`products`, `services`, `brands`, `people`, `everyday life`, `other`
`synthetic_target`	for synthetic pieces, the tone the generator was asked for
`url`	source URL (web pieces)

The pieces config, one row per passage, with the full text (text), the same metadata and the FineWeb document id (fineweb_id). Use it to rebuild the context of a segment.

How it was built

Passages. Random documents from FineWeb (sample-10BT, English) and FineWeb-2 (French, Spanish, German, Italian, Portuguese), keeping documents with a MinHash cluster size of 2 or less where the column exists (FineWeb-2). From each document, one window of whole paragraphs, up to 350 words. Domains that host the reviews behind SST (Rotten Tomatoes, IMDb reviews) are excluded.

Formats and topics. Claude Sonnet 5.5 classifies every published piece into one of the 26 formats (or other) and one topic. An in-house classifier was used only to route candidates toward rare formats; the published values come from Sonnet.

Sampling (sampling column):

random: no selection on sentiment;
preselected_negative, preselected_positive, preselected_mixed: chosen because a 3-class sentiment head (trained on English human data and earlier silver data) predicted strong sentiment in at least one segment;
learned_preselection_very_negative, learned_preselection_very_positive: chosen by two logistic regressions trained on the round-1 labels to predict whether a passage contains a -3 or a +3 segment, applied to the rest of the pool, at most 5 passages per host;
synthetic_complement: written by Claude Sonnet 5.5 for one format, one topic and one target tone, in the target language, using verbalized sampling (5 texts per call with self-estimated probabilities, which raised diversity in our tests).

Segmentation. One line is one block; Markdown markers are removed; blocks are split into sentences.

Labeling. Claude Sonnet 5.5 (adaptive thinking, low effort) receives the full passage and its numbered segments, and scores each segment from -3 to +3 following a written rubric. The labeler never sees the sampling arm or the synthetic target. The labeler is stricter than the generator: a piece written as "very negative" yields about 11 % of -3 segments and 50 % of -2.

Leak check. Any piece containing a segment of 3 words or more that also appears in a test or dev split of the benchmarks below was removed.

Personal data. Email addresses and phone numbers are replaced by [email] and [telephone] (462 and 276 segments). Person names in reviews and forum posts are left as they appear on the public web.

Label quality

The same labeler, with the same rubric, scored the test items of existing human-annotated datasets. Macro-F1 after folding its score to 3 classes (negative < 0, neutral = 0, positive > 0), and Spearman correlation with the human ordinal labels:

Benchmark (human labels)	Items	Macro-F1, 3 classes	Spearman
SST-5 (film review sentences, EN)	2,210	0.687	0.873
Financial PhraseBank, all agree (EN)	2,259	0.945	0.935
SemEval-2007 Affective Text (headlines, EN)	1,250	0.674	0.822
Tweet Sentiment Multilingual (EN)	870	0.688	0.768
Tweet Sentiment Multilingual (FR)	870	0.682	0.593
Tweet Sentiment Multilingual (ES)	870	0.677	0.762
Tweet Sentiment Multilingual (DE)	870	0.689	0.748
Tweet Sentiment Multilingual (IT)	870	0.774	0.815
Tweet Sentiment Multilingual (PT)	870	0.721	0.837

Read these numbers as a ceiling: a model trained on this dataset learns the labeler, and cannot agree with humans more than the labeler does. No human benchmark covers 7 levels at sentence level in these languages, so the 7-level scores themselves are not validated against humans.

Baseline

A logistic regression on frozen EmbeddingGemma-300m embeddings (classification prompt, mean-centered per corpus), trained on the web part of the training split only, with no human label:

Test	Metric	Baseline
9 human benchmarks above (mean)	macro-F1, 3 classes	0.673
Tweet Sentiment Multilingual (FR)	macro-F1, 3 classes	0.632
eval_natural (mean of 6 languages)	macro-F1, 7 levels	0.539
eval_natural	exact within ±1 (always predicting 0: 0.815)	0.947

The eval_natural labels come from the same labeler as the training data, which favors models trained on this dataset; the human benchmarks are the fair reference. Adding the synthetic complement to training did not improve results on eval_natural, including on -3 (per-level F1 about 0.33). Its value on the rare formats it targets is not measured.

This baseline is released as SDeaure/websentiment-egemma-ml-7, with Python and transformers.js usage.

Known limitations

Silver labels from a single LLM, with its own biases; strong sentiment is the least reliable (-3 has the lowest per-level F1).
Preselection makes the training split unrepresentative of the web; use eval_natural, or filter on sampling = random, for a natural mix.
Formats come from an LLM classifier too, and other remains the largest format in every language.
News and blog posts dominate the web part. Rare formats are covered by synthetic text in several languages.
Segments are short (median 12 words); some are headings or list items.

Help build a gold subset

We want to release a human-annotated subset of 2,000 to 3,000 segments per language, each labeled by three people, with agreement reported (Krippendorff's alpha). If you can annotate in one of the six languages, open a discussion on this page.

Licence

The web text comes from FineWeb and FineWeb-2, released under the ODC-By 1.0 licence and subject to the Common Crawl terms of use. This dataset is released under ODC-By 1.0 as well. Labels, formats, topics and synthetic texts were produced with Claude Sonnet 5.5 (Anthropic). The dataset is meant for training and evaluating sentiment classifiers, not text generators.

Attribution. ODC-By 1.0 (section 4.3) requires a notice whenever you publicly use a work produced from this dataset, such as a model trained on it, a published analysis or a tool. The following notice satisfies it:

Contains information from WebSentiment ML-7 by Sylvain Deauré (https://sylvain-deaure.fr/en/websentiment/), which is made available under the ODC Attribution License.

On a web page, please make it a plain link to https://sylvain-deaure.fr/en/websentiment/, without rel="nofollow". This is a request, not a condition of the licence.

If you find your content in this dataset and want it removed, open a discussion with the URL.

Citation

Author: Sylvain Deauré, ORCID 0009-0009-1013-0873. DOI 10.5281/zenodo.23964284 (Zenodo).

@misc{websentiment_ml_7,
  author = {Sylvain Deaur{\'e}},
  title  = {WebSentiment ML-7: multilingual 7-point sentiment for web text},
  year   = {2026},
  publisher = {Zenodo},
  doi    = {10.5281/zenodo.23964284},
  url    = {https://huggingface.co/datasets/SDeaure/websentiment-ml-7}
}

Downloads last month: 51

Models trained or fine-tuned on SDeaure/websentiment-ml-7

SDeaure/websentiment-egemma-ml-7

Text Classification • Updated about 3 hours ago

SDeaure/websentiment-egemma-ml-7-ctx

Text Classification • Updated about 3 hours ago

Collection including SDeaure/websentiment-ml-7

WebSentiment

Collection

Multilingual sentiment for web text: datasets and models. • 3 items • Updated about 19 hours ago