[!NOTE] This is a copy of papluca/xlm-roberta-base-language-detection at commit 9865598, made on 2026-10-30. vLLM CI uses this checkpoint and the original is on a personal account, so we keep a copy here to reduce supply-chain risk.

xlm-roberta-base-language-detection

This model is a fine-tuned version of xlm-roberta-base on the Language Identification dataset.

Model description

This model is an XLM-RoBERTa transformer model with a classification head on top (i.e. a linear layer on top of the pooled output). For additional information please refer to the xlm-roberta-base model card or to the paper Unsupervised Cross-lingual Representation Learning at Scale by Conneau et al.

Intended uses & limitations

You can directly use this model as a language detector, i.e. for sequence classification tasks. Currently, it supports the following 20 languages:

arabic (ar), bulgarian (bg), german (de), modern greek (el), english (en), spanish (es), french (fr), hindi (hi), italian (it), japanese (ja), dutch (nl), polish (pl), portuguese (pt), russian (ru), swahili (sw), thai (th), turkish (tr), urdu (ur), vietnamese (vi), and chinese (zh)

Training and evaluation data

The model was fine-tuned on the Language Identification dataset, which consists of text sequences in 20 languages. The training set contains 70k samples, while the validation and test sets 10k each. The average accuracy on the test set is 99.6% (this matches the average macro/weighted F1-score being the test set perfectly balanced). A more detailed evaluation is provided by the following table.

Benchmarks

As a baseline to compare xlm-roberta-base-language-detection against, we have used the Python langid library. Since it comes pre-trained on 97 languages, we have used its .set_languages() method to constrain the language set to our 20 languages. The average accuracy of langid on the test set is 98.5%. More details are provided by the table below.

How to get started with the model

The easiest way to use the model is via the high-level pipeline API:

from transformers import pipeline

text = [
    "Brevity is the soul of wit.",
    "Amor, ch'a nullo amato amar perdona."
]

model_ckpt = "papluca/xlm-roberta-base-language-detection"
pipe = pipeline("text-classification", model=model_ckpt)
pipe(text, top_k=1, truncation=True)

Or one can proceed with the tokenizer and model separately:

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

text = [
    "Brevity is the soul of wit.",
    "Amor, ch'a nullo amato amar perdona."
]

model_ckpt = "papluca/xlm-roberta-base-language-detection"
tokenizer = AutoTokenizer.from_pretrained(model_ckpt)
model = AutoModelForSequenceClassification.from_pretrained(model_ckpt)

inputs = tokenizer(text, padding=True, truncation=True, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits

preds = torch.softmax(logits, dim=-1)

# Map raw predictions to languages
id2lang = model.config.id2label
vals, idxs = torch.max(preds, dim=1)
{id2lang[k.item()]: v.item() for k, v in zip(idxs, vals)}

Training procedure

Fine-tuning was done via the Trainer API. Here is the Colab notebook with the training code.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 64
• eval_batch_size: 128
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 2
• mixed_precision_training: Native AMP

Training results

The validation results on the valid split of the Language Identification dataset are summarised here below.

In short, it achieves the following results on the validation set:

• Loss: 0.0101
• Accuracy: 0.9977
• F1: 0.9977

Framework versions

• Transformers 4.12.5
• Pytorch 1.10.0+cu111
• Datasets 1.15.1
• Tokenizers 0.10.3