---
license: apache-2.0
language:
- en
- zh
- ja
- es
- de
- fr
- ar
task_categories:
- text-classification
task_ids:
- sentiment-classification
size_categories:
- 1K<n<10K
tags:
- finance
- financial-sentiment
- sentiment-analysis
- multilingual
- financial-news
- evaluation
- benchmark
configs:
- config_name: default
  data_files:
  - split: test
    path: test.csv
---

# Financial Sentiment Evaluation Set, 7 Languages

**The held-out test set behind the reported accuracy of
[`Kenpache/finbert-multilingual-v2`](https://huggingface.co/Kenpache/finbert-multilingual-v2),
published together with the protocol so the results are independently reproducible.**

4,993 financial news sentences in English, Chinese, Japanese, Spanish, German, French and
Arabic, each labelled `negative`, `neutral` or `positive`.

## Reproducing the reported result

The model card reports **accuracy 0.8724**. End to end:

```python
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"
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("accuracy:", (pd.Series(preds) == ds.label).mean())
# accuracy: 0.8724
```

Protocol that the published numbers use: **`max_length=192`, fp32, raw text with no
normalisation**. Deviating from any of the three shifts the result slightly.

Per-language accuracy, the way the model card reports it:

```python
ds["pred"] = preds
print(ds.groupby("language").apply(lambda g: (g.pred == g.label).mean()).sort_values())
```

## What is in the file

| Column | What it is |
|---|---|
| `row_id` | stable id, traceable back to [the source corpus](https://huggingface.co/datasets/Kenpache/multilingual-financial-sentiment) |
| `group_id` | near-duplicate cluster id (see below) |
| `sentence` | the text to classify |
| `label` | `negative` / `neutral` / `positive` |
| `source` | publication the sentence came from (82 outlets) |
| `language` | `en` · `zh` · `ja` · `es` · `de` · `fr` · `ar` |

### Size and balance

| Language | | Items | negative | neutral | positive |
|---|---|---:|---:|---:|---:|
| Japanese | `ja` | 1,063 | 244 | 423 | 396 |
| Chinese | `zh` | 1,023 | 252 | 379 | 392 |
| Spanish | `es` | 905 | 220 | 473 | 212 |
| English | `en` | 780 | 200 | 357 | 223 |
| German | `de` | 650 | 200 | 303 | 147 |
| French | `fr` | 499 | 123 | 189 | 187 |
| Arabic | `ar` | 73 | 21 | 34 | 18 |
| **Total** | | **4,993** | **1,260** | **2,158** | **1,575** |

Arabic is small — 73 items. Any per-language number computed on it carries a wide
confidence interval, and it should be read as indicative rather than precise.

### Near-duplicate grouping

Financial news is full of templated sentences — earnings-per-share lines, index closing
levels, boilerplate disclosures. Left alone, a near-identical sentence can land in
training while its twin lands in the test set, and the resulting score measures memory
rather than generalisation.

`group_id` marks those clusters, and **the split was made by group, not by row**: every
member of a cluster stays on the same side. The largest cluster here holds 18 sentences.
If you build your own splits from this file, split on `group_id` too.

## Intended use

This is an **evaluation set**, intended for measuring and comparing models.

It is not intended as training data. Including it in training invalidates it as a
benchmark — both for the resulting model and for any comparison against results reported
on it.

## License

Apache 2.0, following the source corpus. Sentences are short excerpts from public
financial news, attributed per row in the `source` column.

## Citation

```bibtex
@misc{fin_sentiment_eval_7lang,
  title  = {Financial Sentiment Evaluation Set, 7 Languages},
  author = {Kenpache},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Kenpache/financial-sentiment-eval-7lang}
}
```