---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: text
    dtype: string
  - name: label
    dtype: string
  - name: source
    dtype: string
  - name: length
    dtype: int64
  - name: length_bucket
    dtype: string
  - name: source_doc_id
    dtype: string
  splits:
  - name: train
    num_bytes: 19135114
    num_examples: 41996
  - name: validation
    num_bytes: 4126385
    num_examples: 8997
  - name: test
    num_bytes: 4127151
    num_examples: 9007
  download_size: 16528594
  dataset_size: 27388650
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
license: cc-by-4.0
language:
- sr
- bs
- hr
pretty_name: SR/BS/HR Language Identification Dataset
size_categories:
- 10K<n<100K
---

# SR/BS/HR Language Identification Dataset

<div align="center">

![Languages](https://img.shields.io/badge/languages-sr%20|%20bs%20|%20hr-blue)
![License](https://img.shields.io/badge/license-CC--BY--SA--4.0-green)
![Size](https://img.shields.io/badge/examples-60K-orange)
![Task](https://img.shields.io/badge/task-language%20identification-purple)

**Fine-grained language identification dataset for Serbian, Bosnian, and Croatian**

[RSA Team](https://huggingface.co/rsateam) • [GitHub](https://github.com/rsadevteam) • [Website](https://rsateam.com)

</div>

## Overview

This dataset enables fine-grained language identification between Serbian, Bosnian, and Croatian — three closely related South Slavic languages that are often misclassified by general-purpose language identification tools.

Derived from [rsateam/sr-bs-hr-clean-text](https://huggingface.co/datasets/rsateam/sr-bs-hr-clean-text), this dataset uses **source-based labeling** to provide ground-truth labels without circular dependencies on automatic language detection.

### The Challenge

Existing language identification tools often struggle with sr/bs/hr:

| Tool | Problem |
|------|---------|
| FastText lid.176 | Often treats them as one language |
| langid.py | Low precision between sr/bs/hr |
| CLD3 | Inconsistent across text lengths |

### This Dataset Enables

- Training specialized sr/bs/hr classifiers
- Benchmarking and evaluating LID systems
- Preprocessing pipeline routing for multilingual NLP
- Research on closely related language varieties

## Dataset Statistics

| Metric | Value |
|--------|-------|
| Total examples | **60,000** |
| Dataset size | 27.4 MB |
| Download size | 16.5 MB |
| Labels | 3 (sr, bs, hr) |

### Splits

| Split | Examples | Size |
|-------|----------|------|
| `train` | 41,996 | 19.1 MB |
| `validation` | 8,997 | 4.1 MB |
| `test` | 9,007 | 4.1 MB |

### Label Distribution

Each split is balanced across languages (~33% each):

| Label | Language |
|-------|----------|
| `sr` | Serbian |
| `bs` | Bosnian |
| `hr` | Croatian |

### Length Distribution

| Bucket | Description |
|--------|-------------|
| `short` | <100 characters |
| `medium` | 100-300 characters |
| `long` | >300 characters |

## Dataset Structure

### Data Format

```json
{
  "id": "550e8400-e29b-41d4-a716-446655440000",
  "text": "Kratak pasus teksta za identifikaciju...",
  "label": "bs",
  "source": "bs.wikipedia.org",
  "length": 156,
  "length_bucket": "medium",
  "source_doc_id": "original-document-uuid"
}
```

### Fields

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | Unique sample identifier |
| `text` | string | Text sample for classification |
| `label` | string | Language label (sr/bs/hr) |
| `source` | string | Source domain |
| `length` | int | Length in characters |
| `length_bucket` | string | Length category (short/medium/long) |
| `source_doc_id` | string | Reference to source document in clean-text corpus |

## Labeling Methodology

### Source-Based Labeling

Labels are assigned based on the **source of publication**, not automatic detection:

| Source | Label |
|--------|-------|
| sr.wikipedia.org | `sr` |
| bs.wikipedia.org | `bs` |
| hr.wikipedia.org | `hr` |

### Why Source-Based?

This approach is academically accepted and avoids common pitfalls:

| Advantage | Explanation |
|-----------|-------------|
| Ground truth | Labels reflect author/publisher intent |
| No circular dependency | Not trained on LID output |
| Reproducible | Deterministic labeling process |
| Transparent | Source field enables verification |

## Usage

### Loading the Dataset

```python
from datasets import load_dataset

# Load full dataset
dataset = load_dataset("rsateam/sr-bs-hr-language-id")

# Load specific split
train = load_dataset("rsateam/sr-bs-hr-language-id", split="train")
test = load_dataset("rsateam/sr-bs-hr-language-id", split="test")
```

### Training a Classifier

```python
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from transformers import TrainingArguments, Trainer

dataset = load_dataset("rsateam/sr-bs-hr-language-id")

# Create label mapping
label2id = {"sr": 0, "bs": 1, "hr": 2}
id2label = {v: k for k, v in label2id.items()}

# Load model and tokenizer
model_name = "xlm-roberta-base"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(
    model_name,
    num_labels=3,
    label2id=label2id,
    id2label=id2label
)

# Tokenize
def tokenize(examples):
    tokens = tokenizer(examples["text"], truncation=True, max_length=256)
    tokens["labels"] = [label2id[label] for label in examples["label"]]
    return tokens

tokenized = dataset.map(tokenize, batched=True)

# Train
trainer = Trainer(
    model=model,
    args=TrainingArguments(
        output_dir="./sr-bs-hr-lid",
        num_train_epochs=3,
        per_device_train_batch_size=32,
        evaluation_strategy="epoch"
    ),
    train_dataset=tokenized["train"],
    eval_dataset=tokenized["validation"]
)

trainer.train()
```

### Evaluating Existing LID Systems

```python
from datasets import load_dataset
import fasttext

# Load dataset and FastText model
dataset = load_dataset("rsateam/sr-bs-hr-language-id", split="test")
lid_model = fasttext.load_model("lid.176.bin")

# Evaluate
correct = 0
for example in dataset:
    pred = lid_model.predict(example["text"].replace("\n", " "))[0][0]
    pred_lang = pred.replace("__label__", "")
    
    # Map FastText predictions to our labels
    if pred_lang in ["sr", "bs", "hr"]:
        if pred_lang == example["label"]:
            correct += 1

accuracy = correct / len(dataset)
print(f"FastText accuracy on sr/bs/hr: {accuracy:.2%}")
```

### Filtering by Length

```python
# Get only short texts for challenging evaluation
short_texts = dataset["test"].filter(lambda x: x["length_bucket"] == "short")

# Get long texts for easier classification
long_texts = dataset["test"].filter(lambda x: x["length_bucket"] == "long")
```

## Relationship to Source Dataset

This dataset is derived from [rsateam/sr-bs-hr-clean-text](https://huggingface.co/datasets/rsateam/sr-bs-hr-clean-text):

```
sr-bs-hr-clean-text (641K docs)
        │
        ▼
   [Sampling & Segmentation]
        │
        ▼
sr-bs-hr-language-id (60K samples)
```

The `source_doc_id` field links each sample back to its source document.

## Considerations

### Use Cases

- **Primary**: Training and evaluating language identification models
- **Secondary**: Preprocessing pipelines, dialect research, cross-lingual studies

### Limitations

- Source-based labels may not capture individual variation
- Wikipedia style may differ from conversational text
- Some lexical overlap between languages is inherent

### Ethical Considerations

- Derived from Wikipedia (CC-BY-SA)
- No personally identifiable information
- Language labels reflect publication source, not speaker identity

### License

[CC-BY-SA-4.0](https://creativecommons.org/licenses/by-sa/4.0/) — consistent with source dataset.

## Citation

```bibtex
@dataset{rsateam_language_id_2026,
  title={SR/BS/HR Language Identification Dataset},
  author={RSA Team},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/rsateam/sr-bs-hr-language-id},
  note={Fine-grained language identification for Serbian, Bosnian, and Croatian}
}
```

## Related Datasets

| Dataset | Description |
|---------|-------------|
| [rsateam/sr-bs-hr-clean-text](https://huggingface.co/datasets/rsateam/sr-bs-hr-clean-text) | Source corpus (641K documents) |

## Contributing

For suggestions, bug reports, or improvements:

- Open an issue on [GitHub](https://github.com/rsadevteam/balkan-nlp)
- Email: office@rsateam.com

---

<div align="center">

**[RSA Team](https://huggingface.co/rsateam)** — *Building bridges between languages and AI, one dataset at a time.*

</div>