algerian-cars-realestate-search-dataset

A multilingual query/listing relevance dataset for Algerian marketplace search, built from algerian search listings (vehicles and real estate) with user queries in Darja, Arabizi, French, and Arabic.

Each example pairs a search query with a listing document and a relevance score in [0, 1]. Positive pairs come from true query-listing matches; negatives were mined using the 81melody/algerianME5 dense retriever (hard negatives: top-retrieved but irrelevant listings). This dataset was used to train the cross-encoder reranker, which reaches a mean NDCG@10 of 0.967 on the full retrieval pipeline across 7 multilingual query categories.

---

Dataset Statistics

Split	Examples	File
Train	14,292	train.jsonl
Validation	1,588	val.jsonl
Total	15,880

Category	Train examples
immo (real estate)	9,263
auto (vehicles)	5,029

Strategy	Examples
mined_negative (hard negatives from bi-encoder)	10,246
positive (true matches)	4,046

Positive/negative breakdown (label > 0.5): positive 5,937 · negative 8,355

---

File Schema

Each line in train.jsonl and val.jsonl is a JSON object:

{
  "query":    "appart Souk Ahras",
  "document": "بيع شقة 4 غرف سوق أهراس ... Appartement f4 en plein centre ville ...",
  "label":    0.7,
  "weight":   1.0,
  "category": "immo",
  "strategy": "mined_negative"
}

Field	Type	Description
query	string	User search query — Darja, Arabizi, French, MSA or mixed
document	string	Listing text (title + description concatenated)
label	float [0, 1]	Relevance score: 1.0 = exact match, 0.7 = partial match, 0.0 = irrelevant
weight	float	Loss weight for training (positive examples may be upweighted)
category	string	"immo" (real estate) or "auto" (vehicles)
strategy	string	How the pair was generated: "positive" or "mined_negative"

---

Quick Start

from datasets import load_dataset
import json

# Load locally
def load_jsonl(path):
    with open(path) as f:
        return [json.loads(line) for line in f]

train = load_jsonl("train.jsonl")
val   = load_jsonl("val.jsonl")

# Inspect a positive pair
pos = [ex for ex in train if ex["label"] > 0.5 and ex["strategy"] == "positive"][0]
print("Query:   ", pos["query"])
print("Document:", pos["document"][:200])
print("Label:   ", pos["label"])

---

Query Language Distribution

Queries span the full Algerian multilingual space:

Language / Script	Examples
French	"appart Souk Ahras", "villa avec piscine"
Algerian Darja (Arabic)	"نحوس على فيلا في العاصمة"
Arabizi (Arabic in Latin)	"nhawes 3la dar"
Mixed French-Arabic	"appart f3 في الجزائر العاصمة"
MSA	"شقة للبيع في وهران"

---

Performance of the Trained Cross-Encoder

The model trained on this data achieves the following NDCG@10 on an internal 46-query eval set spanning both categories and all query languages:

Category	Description	NDCG@10
A	Exact-match vehicle queries	0.929
B	Semantic French queries	0.951
C	Darja queries	0.984
D	Arabizi queries	0.998
E	Real-estate (AR/FR)	0.956
F	Semantic real-estate	0.999
G	Mixed-language	0.999
Mean		0.967

These scores measure the full two-stage pipeline (bi-encoder retrieval + cross-encoder reranking), not the cross-encoder in isolation.

---

Source

Listings collected from the largest Algerian online marketplaces (vehicles and real estate). Queries are synthetic or paraphrased from real user search patterns. Listing text is public-facing ad content. Phone numbers and email addresses in listing text were masked with [PHONE] and [EMAIL] placeholders before release.

---

Citation

@dataset{himeur2026algerian_search,
  title        = {algerian-cars-realestate-search-dataset: Multilingual Query/Listing Relevance Dataset
                  for Algerian Marketplace Search},
  author       = {Himeur, Ayoub},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/81melody/algerian-cars-realestate-search-dataset},
  note         = {15,880 query/listing pairs from algerian Marketplaces (vehicles + real estate),
                  Darja/Arabizi/French/MSA queries, hard-negative mined}
}

---

License

Apache 2.0