finbert-multilingual-v2-large

Financial Sentiment, 7 Languages — Large

Sentiment of financial news, in seven languages, from one model. Feed it a headline or a sentence in English, Chinese, Japanese, Spanish, German, French or Arabic — get back negative, neutral or positive.

from transformers import pipeline

clf = pipeline("text-classification", model="Kenpache/finbert-multilingual-v2-large")

clf("The company reported record quarterly earnings, driven by strong demand.")
# [{'label': 'positive', 'score': 0.9445}]

clf("Die Aktie verlor nach der Gewinnwarnung deutlich an Wert.")
# [{'label': 'negative', 'score': 0.9435}]

clf("该公司宣布大规模裁员计划，股价应声下跌。")
# [{'label': 'negative', 'score': 0.9436}]

One model covers all seven languages — no per-language checkpoints, no translation step, no language ID in front of it. Mixed-language pipelines just work.

Smaller sibling: Kenpache/finbert-multilingual-v2 — 307M parameters, 87.2% on the same evaluation set. Take this one for accuracy, that one for footprint.

---

Accuracy

Measured on a held-out test set of 4,993 financial news sentences across the seven languages, published as Kenpache/financial-sentiment-eval-7lang.

Per language

This is the table to read before adopting the model — it tells you whether your language is covered properly, not just the average.

Every one of the seven languages is above 84%, and English is not at the top — Spanish and German are. That matters more than it looks: most "multilingual" financial models are English models with a multilingual tokenizer, and they collapse on CJK and right-to-left text. This one holds its level across scripts — Latin, Chinese, Japanese and Arabic alike.

Arabic is measured on 73 items, so treat its number as indicative rather than precise.

Reproducing these numbers

The evaluation set is public, and so is the protocol — max_length=192, fp32, raw text with no normalisation:

import pandas as pd, torch
from datasets import load_dataset
from transformers import AutoModelForSequenceClassification, AutoTokenizer

ds = load_dataset("Kenpache/financial-sentiment-eval-7lang", split="test").to_pandas()

REPO = "Kenpache/finbert-multilingual-v2-large"
tok = AutoTokenizer.from_pretrained(REPO)
model = AutoModelForSequenceClassification.from_pretrained(REPO).eval()

preds = []
with torch.no_grad():
    for i in range(0, len(ds), 64):
        enc = tok(ds.sentence[i:i + 64].tolist(), return_tensors="pt",
                  padding=True, truncation=True, max_length=192)
        preds += [model.config.id2label[j].lower()
                  for j in model(**enc).logits.argmax(-1).tolist()]

print((pd.Series(preds) == ds.label).mean())   # 0.8892

Per class

No class collapse: the three F1 scores sit within 2.2 points of each other, and neutral — the majority class, and the usual dumping ground for models that learned to hedge — has the lowest F1 of the three rather than the highest.

Polarity errors are rare. Across all 4,993 items, negative is called positive 14 times and positive is called negative 24 times — 38 cases, 0.8% of the set. Practically all remaining error sits on the boundary with neutral. The model may fail to register a weak signal; it very seldom reverses one.

Comparison on the English subset

Both models were run on the English portion — 780 items — of Kenpache/financial-sentiment-eval-7lang, under one identical protocol: max_length=192, fp32, raw text, argmax over the three classes, no tuning or threshold fitting for either model.

Two things belong next to those numbers. ProsusAI/finbert is an English-only model, so the comparison is confined to the English subset — which is, as the table above shows, this model's weakest language of the seven. And it was trained under a different annotation convention: most of its errors on this set are neutral items assigned a direction, so part of the gap reflects differing label conventions rather than capability.

These figures describe behaviour on this evaluation set only, under the protocol stated above. They are not a general claim about either model.

Cross-benchmark check: Financial PhraseBank

A model tends to look good on the benchmark its own authors picked, so here is the mirror image of the table above — the same two models on Financial PhraseBank (Malo et al., 2014), the long-standing English benchmark in this field, all 4,846 sentences, under the identical protocol.

Read that gap with one fact next to it: ProsusAI/finbert was fine-tuned on Financial PhraseBank — its model card states that "Financial PhraseBank by Malo et al. (2014) is used for fine-tuning". This model has never seen the corpus. All 4,846 sentences were checked against this model's training, validation and test data after normalising case and punctuation; the overlap is zero.

So neither benchmark is neutral ground. The first favours this model, the second favours ProsusAI/finbert. Together they bracket the answer:

This model gives up under two points when moved onto a foreign benchmark. The English-only model moves by nearly seventeen between the two.

Polarity holds. On its own evaluation set this model reverses polarity on 0.8% of items; on Financial PhraseBank — a corpus a decade older, in a different register, under a different annotation convention — the rate is 0.8% (37 of 4,846), unchanged. The core judgement of direction does not degrade off home ground.

95% of the remaining error sits on the boundary with neutral, which is where the two conventions genuinely disagree rather than where the model fails. Financial PhraseBank labels a signed contract or a reported sales increase as positive; this model treats a bare corporate fact as neutral unless the text carries an evaluative charge. Neither reading is wrong — they are two conventions, and each model follows the one it was built for.

These figures describe behaviour on these two evaluation sets only. They are not a general claim about either model.

---

Usage

pip install transformers torch

Pipeline

from transformers import pipeline

clf = pipeline("text-classification", model="Kenpache/finbert-multilingual-v2-large")

clf("Les bénéfices du groupe ont augmenté de 15% au premier trimestre.")
# [{'label': 'positive', 'score': 0.9443}]

Batch a whole list in one call:

texts = ["株価は決算発表後に急落した。",
         "La compañía anunció un despido masivo y sus acciones se desplomaron.",
         "Quarterly revenue beat analyst expectations by a wide margin."]

clf(texts, batch_size=32)
# [{'label': 'negative', 'score': 0.9458},
#  {'label': 'negative', 'score': 0.9462},
#  {'label': 'positive', 'score': 0.9467}]

Add top_k=None to get the full probability distribution over all three classes instead of the winner only — useful when you want to threshold on confidence rather than take the argmax.

Direct loading

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

REPO = "Kenpache/finbert-multilingual-v2-large"
tokenizer = AutoTokenizer.from_pretrained(REPO)
model = AutoModelForSequenceClassification.from_pretrained(REPO).eval()

text = "Der Umsatz blieb im Vergleich zum Vorjahr unverändert."
enc = tokenizer(text, return_tensors="pt", truncation=True, max_length=192)

with torch.no_grad():
    probs = torch.softmax(model(**enc).logits, dim=-1)[0]

for i, p in enumerate(probs):
    print(f"{model.config.id2label[i]:8} {p:.4f}")

# negative 0.0234
# neutral  0.9491
# positive 0.0275

On GPU

clf = pipeline("text-classification", model=REPO, device=0)          # CUDA
clf = pipeline("text-classification", model=REPO, device="mps")      # Apple Silicon

CUDA, Apple Silicon and plain CPU all work. At 560M parameters inference is comfortable on a laptop, though noticeably heavier than the 307M sibling.

Use max_length=192 to reproduce the numbers above. The hard ceiling of the backbone is 512 tokens — enough for headlines and single sentences, which is what this model is for, but do not expect it to take long documents.

---

Limitations

1. Sentence-level, not document-level. The model is built for headlines and single sentences. Feeding a full article gives you one label for the whole thing, which is rarely what you want — split it first. The 512-token ceiling enforces this anyway.
2. Financial sentiment is not general sentiment. "Shares fell 3% on the news" is negative in a market sense with no emotional language at all. On product reviews or social media this model is the wrong tool.
3. neutral is a convention, not a fact. The boundary between neutral and mildly positive/negative is where human annotators disagree most, and the model inherits that ambiguity. If a decision hinges on that boundary, use the probabilities and a threshold instead of the argmax.
4. Arabic coverage is thin in evaluation (73 items). The other six languages are measured on 499–1,063 items each.
5. Seven languages. The backbone is pretrained on a hundred, but this classifier was tuned for these seven. Other languages will produce output, but it is untested.
6. Not investment advice. The output is a sentiment label on a text, not a signal to trade on.

---

Intended use

Good fits:

• tagging multilingual financial news feeds in real time
• market-sentiment dashboards and indices across regions
• pre-screening research corpora before human analysis
• backtesting sentiment-based features on multilingual sources

Poor fits: general-purpose sentiment, long documents, languages outside the seven, anything where the neutral boundary carries legal or financial weight on its own.

---

Files

License

Apache 2.0.

Built on FacebookAI/xlm-roberta-large, which is MIT-licensed; that attribution is preserved here.

Citation

@misc{finbert_multilingual_v2_large,
  title  = {Financial Sentiment, 7 Languages — Large},
  author = {Kenpache},
  year   = {2026},
  url    = {https://huggingface.co/Kenpache/finbert-multilingual-v2-large}
}