DarjaCore/algerian-darja-forum-posts

Algerian Darja Dataset

A large-scale Algerian Darja conversational dataset prepared for NLP, language-model training, instruction tuning, and conversational AI research.

3,209,157 samples · 1.193B tokens · 371.83 tokens/sample on average

Dataset at a Glance

Property	Value
Samples	3,209,157
Total tokens	1,193,257,847
Approx. tokens	1.193B
Average tokens / sample	371.83
Language	Algerian Darja
Format	Conversational JSON
Storage format	Parquet
License	MIT

Dataset Format

The dataset uses a JSON-based conversational format, with each sample represented as a sequence of messages between a user and an assistant.

A typical sample follows this structure:

{
  "messages": [
    {
      "role": "user",
      "content": "واش راك اليوم؟"
    },
    {
      "role": "assistant",
      "content": "راني مليح الحمد لله، ونتا؟"
    }
  ]
}

Message Roles

Role	Description
user	User input or question
assistant	Assistant response

The conversational structure makes the dataset suitable for training models to understand and generate natural Algerian Darja conversations.

Storage

The dataset is distributed in Parquet format and split into multiple shards for efficient loading and processing.

Example shard structure:

train-00000-of-00005.parquet
train-00001-of-00005.parquet
train-00002-of-00005.parquet
train-00003-of-00005.parquet
train-00004-of-00005.parquet

Dataset Scale

The dataset contains more than 3.2 million conversational samples and approximately 1.193 billion tokens.

This makes it suitable for large-scale experiments involving Algerian Darja language modeling and conversational model adaptation.

Key Statistics

Metric	Value
Samples	3,209,157
Total tokens	1,193,257,847
Total tokens	~1.193B
Average tokens/sample	371.83

Token counts are dependent on the tokenizer used during analysis. The reported total should therefore be interpreted as an estimate tied to the tokenizer used for this dataset analysis.

Language

The primary language of the dataset is Algerian Darja (Algerian Arabic).

The dataset reflects the natural diversity of Algerian online language and may contain:

• Arabic-script Algerian Darja
• Latin-script Algerian Darja
• Arabizi / Franco-Arabic
• French expressions
• English expressions
• Code-switching
• Informal vocabulary
• Internet abbreviations
• Emojis
• Non-standard spelling

The dataset does not attempt to impose a single standardized orthography on Algerian Darja.

Conversational Data

The dataset is designed around user-assistant interactions rather than isolated text samples.

This conversational structure provides training examples containing:

1. A user message
2. An assistant response

This makes the dataset particularly relevant for:

• Conversational AI
• Chat models
• Instruction following
• Supervised fine-tuning
• Assistant response generation
• Algerian Darja adaptation

Token Statistics

The corpus contains approximately 1.193 billion tokens.

Statistic	Value
Total samples	3,209,157
Total tokens	1,193,257,847
Average tokens/sample	371.83
Approximate scale	1.193B

Tokenization efficiency can vary significantly between models because Algerian Darja contains Arabic script, Latin script, Arabizi, French, and mixed-language text.

For this reason, token counts should always be reported together with the tokenizer used.

Data Quality

The dataset has been processed and prepared for language-modeling experiments.

The preprocessing pipeline includes data cleaning and dataset preparation intended to make the corpus more suitable for training.

However, conversational user-generated data can naturally contain:

• Spelling variations
• Informal expressions
• Dialectal variations
• Arabizi
• Code-switching
• Emojis
• Abbreviations
• Noisy text
• Context-dependent statements
• Repetitive patterns

The dataset should therefore not be considered a perfectly standardized representation of Algerian Darja.

Preprocessing

The dataset preparation process may include:

• Empty-sample filtering
• Text cleaning
• Duplicate detection
• Conversation processing
• Dataset validation
• Tokenization analysis
• Length analysis
• Data formatting

The objective is to preserve authentic Algerian Darja usage while producing data suitable for language-model training.

Intended Use

The dataset is intended for research and development in:

• Algerian Darja NLP
• Conversational AI
• Language-model pretraining
• Continued pretraining
• Supervised Fine-Tuning (SFT)
• Instruction tuning
• Chat model development
• Dialect adaptation
• Text generation
• Tokenizer evaluation
• Linguistic research

Training Applications

Supervised Fine-Tuning

The conversational format can be used to fine-tune existing language models to produce Algerian Darja responses.

Continued Pretraining

The corpus can be used as part of continued-pretraining experiments for adapting multilingual or Arabic language models to Algerian Darja.

Conversational Models

The user-assistant structure makes the dataset suitable for developing Algerian Darja conversational capabilities.

Dialect Adaptation

Researchers can use the dataset to study how existing language models can be adapted to Algerian linguistic patterns and vocabulary.

Limitations

This dataset should not be considered a complete representation of the Algerian language or all Algerian speakers.

Potential limitations include:

• Online conversations may not represent spoken Algerian Darja.
• Writing conventions vary between users.
• Arabizi introduces significant spelling variation.
• French and English code-switching may occur.
• User-generated content may contain noise.
• Some responses may depend on conversational context.
• Some content may be sensitive or context-dependent.
• Token statistics depend on the tokenizer used.

Researchers should evaluate and filter the dataset according to their specific use case.

Tokenizer Considerations

Algerian Darja presents several challenges for modern subword tokenizers.

Tokenization can vary depending on:

• Arabic vocabulary coverage
• Latin-script vocabulary coverage
• Arabizi representation
• French vocabulary
• Subword segmentation
• Mixed-script text

Consequently, the 1.193B token figure should not be treated as a tokenizer-independent measurement.

Dataset Structure

Conceptually, each sample follows a conversational message structure:

{
  "messages": [
    {
      "role": "user",
      "content": "..."
    },
    {
      "role": "assistant",
      "content": "..."
    }
  ]
}

This structure follows the common conversational format used by modern chat-oriented language models.

Recommended Uses

Recommended research applications include:

• Algerian Darja language-model adaptation
• Conversational model fine-tuning
• SFT experiments
• Language understanding research
• Text generation
• Dialect-specific evaluation
• Tokenizer benchmarking
• Algerian Arabic linguistic analysis

Dataset Statistics

Current confirmed statistics:

Metric	Value
Samples	3,209,157
Total tokens	1,193,257,847
Approximate tokens	1.193B
Average tokens/sample	371.83

Additional statistics may be added as the dataset analysis pipeline is expanded.

Citation

@dataset{darjacore_algerian_darja,
  author = {DarjaCore},
  title = {Algerian Darja Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/DarjaCore}
}

License

MIT

See the repository license for the applicable terms.

Organization

DarjaCore

Open-source research focused on Algerian Darja, NLP, language models, and language technologies.