alexneakameni/language_detection

Language Detection Model

A BERT-based language detection model trained on hac541309/open-lid-dataset, which includes 121 million sentences across 200 languages. This model is optimized for fast and accurate language identification in text classification tasks.

Model Details

• Architecture: BertForSequenceClassification
• Hidden Size: 384
• Number of Layers: 4
• Attention Heads: 6
• Max Sequence Length: 512
• Dropout: 0.1
• Vocabulary Size: 50,257

Training Process

• Dataset:
  • Used the open-lid-dataset
  • Split into train (90%) and test (10%)
• Tokenizer: A custom BertTokenizerFast with special tokens for [UNK], [CLS], [SEP], [PAD], [MASK]
• Hyperparameters:
  • Learning Rate: 2e-5
  • Batch Size: 256 (training) / 512 (testing)
  • Epochs: 1
  • Scheduler: Cosine
• Trainer: Leveraged the Hugging Face Trainer API with Weights & Biases for logging

Evaluation

The model was evaluated on the test split. Below are the overall metrics:

• Accuracy: 0.969466
• Precision: 0.969586
• Recall: 0.969466
• F1 Score: 0.969417

Detailled evaluation (Size is the number of languages supported)

Script	Support	Precision	Recall	F1 Score	Size
Arab	819219	0.9038	0.9014	0.9023	21
Latn	7924704	0.9678	0.9663	0.9670	125
Ethi	144403	0.9967	0.9964	0.9966	2
Beng	163983	0.9949	0.9935	0.9942	3
Deva	423895	0.9495	0.9326	0.9405	10
Cyrl	831949	0.9899	0.9883	0.9891	12
Tibt	35683	0.9925	0.9930	0.9927	2
Grek	131155	0.9984	0.9990	0.9987	1
Gujr	86912	0.99999	0.9999	0.99995	1
Hebr	100530	0.9966	0.9995	0.9981	2
Armn	67203	0.9999	0.9998	0.9998	1
Jpan	88004	0.9983	0.9987	0.9985	1
Knda	67170	0.9999	0.9998	0.9999	1
Geor	70769	0.99997	0.9998	0.9999	1
Khmr	39708	1.0000	0.9997	0.9999	1
Hang	108509	0.9997	0.9999	0.9998	1
Laoo	29389	0.9999	0.9999	0.9999	1
Mlym	68418	0.99996	0.9999	0.9999	1
Mymr	100857	0.9999	0.9992	0.9995	2
Orya	44976	0.9995	0.9998	0.9996	1
Guru	67106	0.99999	0.9999	0.9999	1
Olck	22279	1.0000	0.9991	0.9995	1
Sinh	67492	1.0000	0.9998	0.9999	1
Taml	76373	0.99997	0.9999	0.9999	1
Tfng	41325	0.8512	0.8246	0.8247	2
Telu	62387	0.99997	0.9999	0.9999	1
Thai	83820	0.99995	0.9998	0.9999	1
Hant	152723	0.9945	0.9954	0.9949	2
Hans	92689	0.9893	0.9870	0.9882	1

A detailed per-script classification report is also provided in the repository for further analysis.

---

How to Use

You can quickly load and run inference with this model using the Transformers pipeline:

from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline

tokenizer = AutoTokenizer.from_pretrained("alexneakameni/language_detection")
model = AutoModelForSequenceClassification.from_pretrained("alexneakameni/language_detection")

language_detection = pipeline("text-classification", model=model, tokenizer=tokenizer)

text = "Hello world!"
predictions = language_detection(text)
print(predictions)

This will output the predicted language code or label with the corresponding confidence score.

---

Note: The model’s performance may vary depending on text length, language variety, and domain-specific vocabulary. Always validate results against your own datasets for critical applications.

For more information, see the repository documentation.

Thank you for using this model—feedback and contributions are welcome!