WebSentiment ML-7: multilingual 7-point sentiment for web text

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 taken from FineWeb and FineWeb-2, each segment labeled in the context of its full passage.86 % of the passages are real web text; the remaining 14 % are synthetic passages, flagged in a `source` column, written to cover formats the web rarely provides (customer emails, ad scripts, support replies).This is a silver dataset: no human labeled these segments. Agreement of the labeler with human annotators is measured on nine existing research datasets and reported in the README.Two parts: `segments` (one row per sentence, with score, label, format, topic and source URL) and `pieces` (one row per passage, with the full text, to rebuild the context). Two splits: `train`, enriched in strong sentiment, and `eval_natural`, a random sample of 300 passages per language that shows the distribution found on the open web.The scale and the taxonomy of 26 text formats follow Dejan AI's sentiment classifier, which trains on synthetic English data; this dataset replaces most of the synthetic text with real web passages and extends it to six languages.Also on Hugging Face: https://huggingface.co/datasets/SDeaure/websentiment-ml-7Models trained on it, licences and how to credit:https://sylvain-deaure.fr/en/websentiment/