File size: 5,031 Bytes
3cab8b5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 | ---
language:
- de
license:
- cc-by-nc-4.0
task_categories:
- question-answering
tags:
- rag
- german
- the-dark-eye
- aventurien
- benchmark
- monolingual
- knowledge-base
pretty_name: Reasoning-Language Alignment QA Benchmark
size_categories:
- 10<n<100
- 1k<n<10k
format: json
modality:
- text
---
# Reasoning-Language Alignment QA Benchmark
A **fully monolingual German** retrieval-augmented question-answering (RAG) benchmark over the fictional world of the tabletop role-playing game *The Dark Eye* (*Das Schwarze Auge*), set on the continent of **Aventurien**. The world is richly documented in German yet far too niche for large language models to answer from memory, so correct answers require reading and integrating retrieved evidence, making it a clean testbed for reasoning over retrieved context.
This dataset accompanies the paper:
> **Investigating the Role of Reasoning-Language Alignment in Monolingual Retrieval-Augmented Generation** — Oliver Hauck, Mario Sanz-Guerrero, Katharina von der Wense (2026)
>
> - Paper: <https://huggingface.co/papers/2610.03136>
> - Code (testbed + dataset creation pipeline): <https://github.com/soliver1/reason-in-the-language-you-retrieve>
## ⚖️ Copyright & what is (not) included
- This benchmark covers the fictional world of *The Dark Eye* (© **Ulisses Spiele**). It is **not** an official Ulisses Spiele product and is published solely for research and reproducibility of the paper above.
- The raw source text of the region-description books is not included, as it is copyrighted. We publish only the *question*, a *reference answer*, and *derived metadata*.
- The reference answers are paraphrases written by the authors; in `qax` the gold key-facts are highlighted with `*…*`.
- If you own the books, the full pipeline (raw text, chunking, and RAG context) can be regenerated with the code in the linked repository.
## Configurations (subsets)
| Config | Description | Rows | Book(s) |
|--------|-----------------------------------------------------------------------------------------------------------------------|-----:|---------|
| `qa2` | Curated QA set derived from book **G01** (*In den Dschungeln Meridianas*), with quality rating + human approval flags | 1,931 | G01 |
| `qax` | More complex, human-authored, multi-faceted questions | 30 | G01, G05, G08 |
### `qa2` fields
| field | type | description |
|-------|------|-------------|
| `id` | str | Stable unique id (`qa2-<chapter>-<n>`) |
| `book` | str | Book code, e.g. `G01` |
| `chapter` | str | Source chapter code, e.g. `G01-012` |
| `question` | str | German question |
| `answer` | str | Reference answer (author paraphrase of the gold answer) |
| `rating` | int | LLM self-rating of the question's quality (1–5) |
| `reason` | str | LLM rationale for the rating |
| `user_rating` | bool | **Human-approval flag — `true` marks the final benchmark subset** |
### `qax` fields
| field | type | description |
|-------|------|-------------|
| `id` | str | Stable unique id (`qax-<book>-<n>`) |
| `book` | str | Book code, e.g. `G01` |
| `question` | str | Complex German question |
| `answer` | str | Reference answer; gold key-facts highlighted with `*…*` |
## Reproducing the paper's benchmark
The experiments use the **human-approved** subset of the curated set (`user_rating == true`), together with the complex set:
```python
from datasets import load_dataset
curated = load_dataset("soliver1/reasoning-language-alignment", "qa2", split="test")
benchmark = curated.filter(lambda x: x["user_rating"]) # 585 approved questions
complex_set = load_dataset("soliver1/reasoning-language-alignment", "qax", split="test") # 30 complex questions
```
## Usage
```python
from datasets import load_dataset
qa2 = load_dataset("soliver1/reasoning-language-alignment", "qa2", split="test")
print(qa2[0])
qax = load_dataset("soliver1/reasoning-language-alignment", "qax", split="test")
print(qax[0])
```
## Citation
If you use this dataset, please cite the paper:
```bibtex
@misc{hauck2026reasoninglanguage,
title = {Investigating the Role of Reasoning-Language Alignment in Monolingual Retrieval-Augmented Generation},
author = {Hauck, Oliver and Sanz-Guerrero, Mario and von der Wense, Katharina},
year = {2026},
eprint = {2610.03136},
archivePrefix = {arXiv},
url = {https://huggingface.co/papers/2610.03136}
}
```
## Trademark note
> ⓘ **DAS SCHWARZE AUGE**, **AVENTURIEN**, **DERE** und **THE DARK EYE** sind eingetragene Marken der Ulisses Spiele GmbH. Ohne vorherige schriftliche Genehmigung der Ulisses Medien und Spiel Distribution GmbH ist eine Verwendung der genannten Markenzeichen nicht gestattet.
## License
`cc-by-nc-4.0` (Creative Commons Attribution-NonCommercial 4.0).
Derived from the world of *The Dark Eye*; not affiliated with or endorsed by Ulisses Spiele.
|