Horizon-Labs/multilingual-sentiment-base

Multilingual Sentiment (base, 308M): negative / neutral / positive

A multilingual sentiment classifier built on mmBERT-base. It returns negative, neutral or positive for reviews, social posts, comments, messages and support tickets in many languages. It is Apache-2.0 and trained only on openly licensed text with labels from an Apache-2.0 LLM, so it can be used commercially. It includes ONNX files for CPU and the browser (transformers.js). A smaller, faster version is available as multilingual-sentiment-small. Try it in the browser.

• Mixed or balanced opinions ("good quality but too expensive") are neutral. Plain facts, questions and requests are neutral too.
• The scores are probabilities, e.g. positive - negative gives a -1 to 1 polarity score.
• onnx/model_quantized.onnx (int8 embeddings, 641 MB) gives the same label as fp32 on 99.5% of 430 benchmark texts.

Usage

from transformers import pipeline

clf = pipeline("text-classification", model="Horizon-Labs/multilingual-sentiment-base")
clf(["I love this phone, the camera is amazing!", "Le colis est arrivé cassé.", "Der Termin ist am Montag."])
# [{'label': 'positive', ...}, {'label': 'negative', ...}, {'label': 'neutral', ...}]

transformers.js:

import { pipeline } from "@huggingface/transformers";
const clf = await pipeline("text-classification", "Horizon-Labs/multilingual-sentiment-base", { dtype: "q8" });
console.log(await clf("Obrigado pela ajuda, vocês são incríveis!"));

Evaluation

These are public benchmarks, used only for evaluation. We never trained on them, and any training text that also appears in them was removed. The metric is macro-F1, averaged over the languages of each benchmark. On the two-class MTEB sets, each model's negative-vs-positive choice is scored, ignoring neutral. Every model is run with the same script on the same texts (code/), and each model's labels are mapped to negative/neutral/positive (1-2 stars = negative, 3 = neutral, 4-5 = positive; "very negative" = negative).

model	licence	Tweets (8 languages, 3-class)	Amazon reviews (6 languages, 3-class)	MTEB MultilingualSentiment (29 languages, pos/neg)	note
this model (308M)	Apache-2.0	0.660	0.703	0.857
multilingual-sentiment-small (141M)	Apache-2.0	0.635	0.695	0.822
cardiffnlp/twitter-xlm-roberta-base-sentiment (278M)	none given	0.700	0.533	0.772	trained on the train split of the tweet benchmark
cardiffnlp/twitter-xlm-roberta-base-sentiment-multilingual (278M)	none given	0.692	0.555	0.760	same tweet training data
lxyuan/distilbert-base-multilingual-cased-sentiments-student (135M)	Apache-2.0	0.414	0.484	0.650
nlptown/bert-base-multilingual-uncased-sentiment (167M)	MIT	0.390	0.645	0.683	trained on product reviews (1-5 stars)
tabularisai/multilingual-sentiment-analysis (135M)	CC-BY-NC-4.0	0.499	0.494	0.670
clapAI/modernBERT-base-multilingual-sentiment (150M)	Apache-2.0	0.784	0.775	0.653	trained on an aggregate of public sentiment sets that may include these benchmarks' train splits
Qwen3.8-27B (our teacher, zero-shot prompt)	Apache-2.0	0.692	0.724	0.892	27B LLM, for reference

• Numbers are for the released checkpoint (seed 0). Mean of two training seeds: small tweets .633 / Amazon .690 / MTEB .824; base tweets .658 / Amazon .703 / MTEB .855.
• Tweets: cardiffnlp/tweet_sentiment_multilingual test (870 per language). Amazon: amazon_reviews_multi test (900 per language, balanced; 3 stars = neutral). MTEB: mteb/multilingual-sentiment-classification test (up to 600 per language; several languages are machine-translated, e.g. Welsh = translated IMDB).
• Models ahead of this one: tweets: twitter-xlm-roberta-base-sentiment, twitter-xlm-roberta-base-sentiment-multilingual, modernBERT-base-multilingual-sentiment; Amazon: modernBERT-base-multilingual-sentiment; MTEB: none. Models trained on a benchmark's own training split have an in-domain advantage on it; this model has seen no benchmark data.

Per benchmark language:

set	this model	cardiffnlp xlm-r	teacher
amazon German	0.744	0.539	0.754
amazon English	0.705	0.548	0.721
amazon Spanish	0.723	0.575	0.746
amazon French	0.728	0.499	0.736
amazon Japanese	0.703	0.543	0.768
amazon Chinese	0.613	0.497	0.623
mteb Arabic	0.853	0.858	0.897
mteb Bambara	0.577	0.591	0.595
mteb Bulgarian	0.893	0.853	0.927
mteb Chinese (cmn set)	0.927	0.789	0.967
mteb Welsh	0.814	0.527	0.938
mteb German	0.885	0.722	0.906
mteb Algerian Arabic	0.927	0.820	0.860
mteb Greek	0.860	0.720	0.913
mteb English	0.893	0.847	0.952
mteb Basque	0.851	0.573	0.867
mteb Persian	0.804	0.797	0.829
mteb Finnish	0.880	0.905	0.917
mteb Hebrew	0.833	0.843	0.873
mteb Croatian	0.943	0.855	0.957
mteb Indonesian	0.945	0.935	0.965
mteb Japanese	0.933	0.827	0.960
mteb Korean	0.806	0.778	0.865
mteb Maltese	0.788	0.489	0.824
mteb Norwegian	0.838	0.744	0.880
mteb Polish	0.985	0.874	1.000
mteb Russian	0.894	0.726	0.862
mteb Slovak	0.944	0.845	0.947
mteb Spanish	0.945	0.916	0.947
mteb Thai	0.775	0.766	0.798
mteb Turkish	0.895	0.907	0.961
mteb Uyghur	0.694	0.580	0.887
mteb Urdu	0.780	0.764	0.786
mteb Vietnamese	0.860	0.808	0.923
mteb Chinese (zho set)	0.845	0.731	0.875
tweets Arabic	0.655	0.670	0.714
tweets English	0.722	0.725	0.704
tweets French	0.594	0.735	0.599
tweets German	0.696	0.749	0.740
tweets Hindi (romanized)	0.590	0.572	0.634
tweets Italian	0.724	0.692	0.791
tweets Portuguese	0.627	0.765	0.674
tweets Spanish	0.669	0.691	0.684

Training

• Text (about 433k training examples): snippets from FineWeb-2 and FineWeb (ODC-BY), in 67 languages, half of them from review, forum, comment and blog pages. We added synthetic reviews, posts, comments, messages and complaints in about 90 languages and varieties (including Hinglish and Arabic dialects), written by Qwen3.8-27B (Apache-2.0) in many genres, topics and lengths.
• Labels: soft labels (probabilities for negative, neutral and positive) from Qwen3.8-27B with a fixed zero-shot instruction (code/sentiment/teacher_sent.py). The model is trained to match these probabilities. The teacher's scores on the benchmarks are in the table above.
• Model: mmBERT-base with a 3-way head, 3 epochs, max length 512 tokens. The checkpoint was chosen by agreement with the teacher on held-out teacher-labelled data, never on the benchmarks.
• Code: code/ in this repository.

Limitations

• Labels come from an LLM, not from human annotators, so the model inherits the teacher's judgement. Its idea of "neutral" can differ from a given dataset's convention: tweet benchmarks mark many mildly opinionated posts as neutral.
• Accuracy is lower on sarcasm, on very short or context-dependent messages (tweets), and on some low-resource languages. Benchmark sets below 0.70 macro-F1: amazon_zh, mteb_bam, mteb_uig, tweets_ar, tweets_fr, tweets_ge, tweets_hi, tweets_po, tweets_sp.
• Trained on texts up to 512 tokens and evaluated on texts up to 2,000 characters. The model accepts up to 8,192 tokens, but longer documents are untested; for those, pass truncation=True, max_length=512, or score paragraphs separately.
• Sentiment is not stance, emotion or toxicity. Don't use it to make decisions about individuals.