algerian-realestate-intent

A 3-class intent classification dataset for Algerian real-estate posts scraped from public Facebook groups — labels each post as BUYER, SELLER, or IRRELEVANT.

Posts are written in Algerian Darja (dialect), Arabizi (Arabic in Latin script), Modern Standard Arabic, and French, often heavily code-switched. This dataset was used to train 81melody/algerianDeBERTa-realestate-intent, the intent-classification stage of the DZ Pulse real-estate data-curation pipeline.

---

Label Schema

id	label	Meaning
0	BUYER	The author is looking to buy or rent a property
1	SELLER	The author is offering a property for sale or rent
2	IRRELEVANT	Unrelated post — chit-chat, spam, off-topic content

---

Dataset Statistics

Metric	Value
Total examples	13,130
BUYER	6,565 (50%)
SELLER	3,070 (23%)
IRRELEVANT	3,495 (27%)
Typical train split	11,160 (85%)
Typical val split	1,970 (15%)
Languages	Darja, Arabizi, French, MSA
Domain	Real-estate (apartments, villas, land — Algeria)

---

File Format

intents.json — a single JSON array. Each element:

{"text": "Salam les amis nhawes villa jwayeh El Biar max 1 milliard", "label": 0}
{"text": "نبيع شقة F3 في درارية واتساب فقط", "label": 1}
{"text": "Bonjour tout le monde, j'espère que vous allez bien", "label": 2}

Labels are integers: 0=BUYER, 1=SELLER, 2=IRRELEVANT.

---

Quick Start

import json
from sklearn.model_selection import train_test_split

with open("intents.json") as f:
    data = json.load(f)

# data is a list of {"text": str, "label": int}
texts  = [d["text"]  for d in data]
labels = [d["label"] for d in data]

# Reproduce the 85/15 split used in training
train_texts, val_texts, train_labels, val_labels = train_test_split(
    texts, labels, test_size=0.15, random_state=42, stratify=labels
)

# Or load directly with the model
from transformers import pipeline
clf = pipeline("text-classification", model="81melody/algerianDeBERTa-realestate-intent", top_k=None)
print(clf("نبيع شقة F3 في درارية"))

---

Notes on Collection and Preprocessing

• Source: Public Algerian Facebook real-estate groups, collected via the Facebook Graph API.
• PII removed: phone numbers and email addresses were masked with [PHONE] and [EMAIL] placeholders before release. The rest of each post is left intact.
• Short posts (<3 tokens) and pure gibberish were filtered prior to annotation.
• Labels were assigned by rule-based classifiers followed by manual review and correction.

---

Associated Model

81melody/algerianDeBERTa-realestate-intent — DeBERTa-v2 classifier fine-tuned on this dataset, achieving strong macro-F1 on the validation split.

---

Citation

@dataset{himeur2026algerian_realestate_intent,
  title        = {algerian-realestate-intent: Buyer/Seller/Irrelevant Intent Classification
                  Dataset for Algerian Real-Estate Facebook Posts},
  author       = {Himeur, Ayoub},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/81melody/algerian-realestate-intent},
  note         = {13,130 labelled posts in Darja, Arabizi, French, and MSA}
}

---

License

Apache 2.0