Task Details

The dataset consists of 5000 inheritance problem examples, distributed as follows, split into 3500 training, 500 validation and 1000 test examples.

📈  A second data release is planned, which will expand the dataset to approximately 10,000 examples, enabling larger-scale training and more robust evaluation.

Data Structure

The dataset follows a consistent and well-defined structure, where each data instance includes the following fields: 

• id: a unique identifier for the example.

• question: a natural-language description (in Arabic) of an Islamic inheritance problem, describing the deceased and the potential heirs.

• answer, composed of two parts:

  • • • <think>: a detailed, step-by-step reasoning process that explains how the inheritance problem is solved according to Islamic inheritance rules (e.g., identification of heirs, blocking rules, fixed shares, and adjustments).

      • <answer>: a concise final response that provides the inheritance distribution without detailed reasoning.

• output (required format): a structured representation that reports the results of each reasoning stage <think>, including:

  • • • identified heirs,

      • blocked (excluded) heirs,

      • assigned shares,

      • adjustment mechanisms (such as ʿawl and radd),

      • and the final distribution of the estate.

Sample Example

Output Example

Output Description (JSON Format) 

The system must output one JSON object containing all steps of the inheritance reasoning process.

The output represents: 

1. the list of heirs who actually inherit, 

2. the list of present but blocked heirs, 

3. the initial shares before adjustments, 

4. the type of global adjustment (ʿawl or radd), 

5. the state of the problem after adjustment, 

6. the final normalized distribution (taṣīl). 

Model predictions are evaluated by comparing the structured output of each reasoning stage with the gold reference. The evaluation measures not only the correctness of the final answer, but also the accuracy of the intermediate results produced at each step.

Each reasoning stage receives a score between 0 and 1, and the final score is obtained by combining the stage-level scores.

The evaluation is performed independently for the following stages:

• heirs: identified heirs and their counts

• blocked: blocked or excluded heirs and their counts

• shares: assigned inheritance fractions for each heir

• awl_or_radd: detected adjustment type (ʿawl, radd, or none)

• awl_stage: adjustment details, when applicable

• post_tasil: final distribution, including per-heir shares and percentages