GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
DarjaCore
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
DarjaCore
Model Tree
Qwen/Qwen3-4B-Base
Input Modalities
Output Modalities
Supported Functionality
| Property | Value |
|---|---|
| Base model | Qwen/Qwen3-4B-Base |
| Model size | ~4B parameters |
| Architecture | Qwen3 |
| Model type | Causal language model |
| Primary language | Algerian Darja / Algerian Arabic |
| Training approach | Continued pretraining / language-model adaptation |
| Training data | DarjaCore Algerian Darja Dataset |
| Dataset fraction | ~5–10% |
| Repository | DarjaCore/Qwen3-4B-Algerian-Darja |
The source dataset contains approximately:
The dataset is structured as conversational message pairs containing user and assistant roles.
Example:
json
{"messages": [{"role": "user","content": "واش راك اليوم؟"},{"role": "assistant","content": "راني مليح الحمد لله، ونتا؟"}]}
The dataset contains natural Algerian online language, including:
Dataset:
https://huggingface.co/datasets/DarjaCore/algerian-darja-forum-posts
The purpose of this training run was to investigate how effectively a relatively small portion of Algerian Darja data can adapt a general-purpose language model toward Algerian linguistic patterns.
The model was not trained on the complete DarjaCore corpus. Approximately 5–10% of the available dataset was used for this experiment.
text
Full dataset:Samples: 3,209,157Estimated tokens: 1,193,257,847Approx. tokens: 1.193BTraining subset:Approximately 5–10% of the full dataset
The exact number of samples and tokens used in the training run may differ depending on dataset selection, preprocessing, filtering, and tokenization.
Token counts are tokenizer-dependent.
This model is primarily intended for research and experimentation involving Algerian Darja and Algerian Arabic NLP.
Potential use cases include:
This release should not be considered:
This is an experimental language adaptation release based on a subset of the available training data.
Because only a subset of the available corpus was used, the model may have limited coverage of:
The model may also retain linguistic patterns, behaviors, and capabilities inherited from the original Qwen3-4B base model.
Training on a larger portion of the DarjaCore corpus may improve Darja coverage, but this has not yet been established through a controlled benchmark.
No standardized Algerian Darja benchmark is currently provided with this release.
Initial qualitative evaluation can focus on:
Future DarjaCore releases may include dedicated evaluation datasets and quantitative benchmarks for Algerian Darja.
python
from transformers import AutoTokenizer, AutoModelForCausalLMmodel_id = "DarjaCore/Qwen3-4B-Algerian-Darja"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(model_id,torch_dtype="auto",device_map="auto")messages = [{"role": "user","content": "واش راك اليوم؟"}]text = tokenizer.apply_chat_template(messages,tokenize=False,add_generation_prompt=True)inputs = tokenizer(text,return_tensors="pt").to(model.device)outputs = model.generate(**inputs,max_new_tokens=256)response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:],skip_special_tokens=True)print(response)
The model is based on the Qwen3 4B architecture and retains the architecture and tokenizer of the base model.
Base model:
Qwen/Qwen3-4B-Base
The repository contains the model weights and configuration required for inference.
DarjaCore is an open research project focused on Algerian Darja, Algerian Arabic NLP, datasets, language models, and language technologies.
This model is one experimental step toward developing language models with stronger representation of Algerian linguistic patterns.
Planned or potential future work includes:
DarjaCore — Algerian Darja Dataset
https://huggingface.co/datasets/DarjaCore/algerian-darja-forum-posts
Current dataset scale:
text
Samples: 3,209,157Estimated tokens: 1,193,257,847Approx. tokens: 1.193B
Token counts are estimates and depend on the tokenizer used.
Qwen3-4B-Base
https://huggingface.co/Qwen/Qwen3-4B-Base
This model is an adaptation of the original Qwen3-4B-Base model and should be evaluated independently from the base model.
This repository uses the license specified for this model release.
Users should also review the licensing terms of the base model and the DarjaCore dataset before using the model or dataset in a particular application.
Thanks to the Qwen team for the Qwen3 model family and to contributors to the open-source Algerian Arabic and Darja NLP ecosystem.
If you use this model in your research or project, please cite:
bibtex
@misc{darjacore_qwen3_4b_algerian_darja,author = {DarjaCore},title = {Qwen3-4B-Algerian-Darja},year = {2026},publisher = {Hugging Face},url = {https://huggingface.co/DarjaCore/Qwen3-4B-Algerian-Darja}}
For the training dataset:
bibtex
@dataset{darjacore_algerian_darja,author = {DarjaCore},title = {Algerian Darja Dataset},year = {2026},publisher = {Hugging Face},url = {https://huggingface.co/datasets/DarjaCore/algerian-darja-forum-posts}}
DarjaCore — Algerian Language and AI Research