A multilingual dataset and models to categorize the sentiment of real-world text.
Sentence by sentence, from -3 to +3, in six languages. Small enough to run on your laptop, or in a browser.
At a glance
WebSentiment is an open dataset, and the models trained on it.
- The training dataset: 462,272 sentences and passages, 86% from the real web, each labeled by an LLM on a 7-level sentiment scale, from -3 (very negative) to +3 (very positive).
- Six languages: English, French, Spanish, German, Italian, Portuguese.
- The model: 119 KB, on top of an open embedding model (EmbeddingGemma). It gives the sentiment level of a sentence, on your machine or in the browser.
What’s inside
The dataset. WebSentiment ML-7: real web text from FineWeb, plus a small synthetic part for rare formats, scored in context.
→ huggingface.co/datasets/SDeaure/websentiment-ml-7
The model. websentiment-egemma-ml-7: a sentiment head on EmbeddingGemma, for Python or the browser.
→ huggingface.co/SDeaure/websentiment-egemma-ml-7
The context model (experimental). websentiment-egemma-ml-7-ctx: reads each sentence with the page it comes from.
→ huggingface.co/SDeaure/websentiment-egemma-ml-7-ctx
Which one?
- Isolated texts (reviews, tweets, titles): the model.
- Sentences of a web page, with the page at hand: the context model.
- Your own classifier, your own scale: the dataset.
Licences and credit
The dataset is under ODC-By 1.0, the models under CC BY 4.0. Free to use, commercially too. A credit is required when you share them or publish work built from them; copy one of these:
Contains information from WebSentiment ML-7 by Sylvain Deauré (https://sylvain-deaure.fr/en/websentiment/), which is made available under the ODC Attribution License.
Sentiment by WebSentiment EmbeddingGemma ML-7 by Sylvain Deauré (https://sylvain-deaure.fr/en/websentiment/), CC BY 4.0.
Using a model in a hosted tool without sharing it? Please credit it in the tool’s help or about page. And on the web, a plain link to this page (without nofollow) is the best thank-you.
The models need EmbeddingGemma-300m (Gemma Terms of Use). Details on each Hugging Face card.
Cite
Sylvain Deauré (2026). WebSentiment ML-7. DOI 10.5281/zenodo.23964284. ORCID 0009-0009-1013-0873. BibTeX on the Hugging Face cards.
How good is it?
0.673 macro-F1 on 9 human-labeled benchmarks in 6 languages, close to the 0.726 of the LLM that labeled the training data. On real web text, the exact level 67% of the time, within one level 95% of the time. Irony and mixed sentences remain the hard part.
Help build the human gold set: if you can annotate in one of the six languages, I’d like to hear from you (address at the bottom of the page).
Who’s behind it
I’m Sylvain Deauré. I’ve been tinkering with computers since 1980 and building for the web since 1997. I created Cocon.se and co-founded 1492.vision, and I work on graphs, embeddings, recommendation and small models that classify text. I presented this approach at TeknSEO 2026: “Google classifies everything, even you. Do the same.”
Substack · X · LinkedIn · Hugging Face · ORCID
Working together
Need a classifier trained on your own data or a workshop for your team?
To reach me by email: syl@sylvain-deaure.fr
Version 1, November 2026. Inspired by DEJAN’s 7-level sentiment model.