---
language:
- ar
- fr
license: apache-2.0
task_categories:
- text-retrieval
- text-ranking
tags:
- algerian-arabic
- darja
- arabizi
- code-switching
- legal
- algerian-family-law
- information-retrieval
- question-answering
- low-resource-nlp
size_categories:
- 1K<10K
configs:
- config_name: biencoder
  data_files:
  - split: train
    path: biencoder_train.jsonl
- config_name: reranker
  data_files:
  - split: train
    path: reranker_train.jsonl
- config_name: golden
  data_files:
  - split: dev
    path: golden_dev.csv
  - split: test
    path: golden_test.csv
---

# algerian-family-law-qa

**A retrieval and reranking dataset for Algerian family-law question answering — real user questions scraped from Algerian Facebook legal-advice groups, paired with relevant articles from the Algerian Family Code (Law No. 84-11).**

Questions are written in Algerian Darja (dialect), Arabizi (Arabic in Latin script), Modern Standard Arabic, and French code-switched text. Documents are articles from the Algerian Family Code covering divorce, custody, alimony, and related topics. This dataset was used to train the two-stage legal retrieval system: [`81melody/algerian-law-biencoder-v2`](https://huggingface.co/81melody/algerian-law-biencoder-v2) (dense retriever) and [`81melody/algerian-law-reranker-v1`](https://huggingface.co/81melody/algerian-law-reranker-v1) (cross-encoder reranker).

---

## Dataset Contents

| File | Split | Examples | Format | Use |
|---|---|---:|---|---|
| `biencoder_train.jsonl` | Train | 389 | JSONL | Bi-encoder / dense retriever training |
| `reranker_train.jsonl` | Train | 9,123 | JSONL | Cross-encoder reranker training |
| `golden_dev.csv` | Dev | 67 | CSV | Human-annotated evaluation (retrieval) |
| `golden_test.csv` | Test | 150 | CSV | Human-annotated test set (retrieval) |

The golden dev and test sets were manually reviewed — each query has a primary article (gold label), optional secondary article, and a hard negative, with a labeler-confidence score.

---

## File Schemas

### `biencoder_train.jsonl`
Each line is a JSON object with hard negatives, used for training a bi-encoder with `MultipleNegativesRankingLoss`:

```json
{
  "query_id": "B24_6",
  "query": "17سنة زواج خيانات متكررة كي نواجهو يضربي...",
  "positive": "المادة 53",
  "hard_negatives": ["المادة 52", "المادة 112", "المادة 51"],
  "in_batch_eligible": true
}
```

### `reranker_train.jsonl`
Each line is a (query, article) pair with a relevance label and confidence weight, used for training a CrossEncoder:

```json
{
  "query": "17سنة زواج...",
  "article_id": "المادة 53",
  "article_structured": "[الموضوع: طلاق] [الباب: الباب الثاني: انحلال الزواج] يجوز للزوجة أن تطلب التطليق...",
  "article_text": "يجوز للزوجة أن تطلب التطليق...",
  "label": 1,
  "weight": 1.0
}
```

`label` is binary (1 = relevant, 0 = not relevant). `weight` in `[0, 1]` reflects annotator confidence (primary match = 1.0, secondary = 0.7, etc.). `article_structured` prefixes the article with its chapter and topic context.

### `golden_dev.csv` / `golden_test.csv`
Columns: `query_id, query_text, article_primary, article_secondary, hard_negative, labeler_confidence, is_oos, parse_warnings, stratum`

`is_oos=True` marks out-of-scope queries (the legal question cannot be answered by the Family Code corpus). `stratum` labels the legal topic (طلاق / OOS / etc.).

---

## Quick Start

```python
from datasets import load_dataset
import json

# Load reranker training pairs
with open("reranker_train.jsonl") as f:
    reranker_data = [json.loads(line) for line in f]

# Load biencoder triplets
with open("biencoder_train.jsonl") as f:
    biencoder_data = [json.loads(line) for line in f]

# Load golden dev set
import csv
with open("golden_dev.csv", newline='', encoding='utf-8-sig') as f:
    dev = list(csv.DictReader(f))

# Filter to in-scope only
in_scope = [row for row in dev if row['is_oos'] == 'False']
```

---

## Statistics

| Metric | Value |
|---|---|
| Unique queries in reranker set | ~1,014 |
| Positive pairs in reranker set | ~1,014 |
| Negative pairs in reranker set | ~8,109 |
| Unique queries in biencoder set | 389 |
| Family Code articles covered | ~100 (Law 84-11 articles 1–124+) |
| Golden dev queries (in-scope) | 48 |
| Golden test queries (in-scope) | ~100 |
| Languages | Darja, Arabizi, MSA, French |

---

## Source

Questions were scraped from public Algerian Facebook groups dedicated to family-law advice. All questions are already publicly posted by users seeking legal guidance. No personally identifiable information (names, phone numbers) appears in the query text as collected — questions are framed around legal situations, not personal contact details.

---

## Associated Models

| Model | HF Repo | Task |
|---|---|---|
| Dense bi-encoder | [81melody/algerian-law-biencoder-v2](https://huggingface.co/81melody/algerian-law-biencoder-v2) | First-stage retrieval |
| CrossEncoder reranker | [81melody/algerian-law-reranker-v1](https://huggingface.co/81melody/algerian-law-reranker-v1) | Second-stage reranking |
| Domain-adapted base | [81melody/algerian-law-marbert](https://huggingface.co/81melody/algerian-law-marbert) | Shared encoder backbone |

---

## Citation

```bibtex
@dataset{himeur2026algerian_family_law_qa,
  title        = {algerian-family-law-qa: Retrieval and Reranking Dataset
                  for Algerian Family-Law Question Answering},
  author       = {Himeur, Ayoub},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/81melody/algerian-family-law-qa},
  note         = {User questions from Algerian Facebook legal-advice groups
                  paired with Algerian Family Code articles (Law 84-11)}
}
```

---

## License

[Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0)