GRAG-LLM-EASY-BENCHMARK

GRAG - German-RAG - German Retrieval Augmented Generation

Dataset Summary

This GRAG-LLM-BENCHMARK represents a specialized collection for evaluating language models with a focus on source citation, time difference stating in RAG-specific tasks. To evaluate models compatible with OpenAI-Endpoints you can refer to our Github Repo: https://github.com/avemio-digital/GRAG-LLM-EASY-BENCHMARK/

Most of the Subsets are synthetically enhanced from German wikipedia from Cohere's dataset (wikipedia-22-12-de-embeddings).

The only exception is:

• Reasoning tasks: Synthetic generation inspired by Tencent's (“Scaling Synthetic Data Creation with 1,000,000,000 Personas”).

Dataset Structure

Data Subsets

Subset	Examples per Task
extraction-recall	1000
qa-with-multiple-references	1000
qa-with-timedifference	1000
qa-without-timedifference	1000
reasoning	1000
relevant-context	1000
summarizations	1000

Source Data: Wikipedia-Subsets

Subsets: classification-json, extraction-recall, ocr-correction, qa-with-multiple-references, qa-with-timedifference, qa-without-timedifference, questions, relevant-context, summarizations

• Primary: Enhanced German Wikipedia content
• Knowledge Graph Structure:
  • Question-Answer nodes
  • Relevant/Irrelevant Context nodes
  • Cross-references within Wikipedia pages

Source Data: Reasoning

• Base: (proj-Persona/PersonaHub)
• Enhancement: Synthetic data generation by Avemio AG
• Quality: Automatic validation and curation of examples by Open Source LLM's

Task Types and Examples

1. Classification JSON

Task-Description: Training the model to categorize inputs according to predefined schemas, with structured JSON output for consistent data handling. This enables automated document routing, customer inquiry categorization, and standardized data extraction from unstructured business documents.

Translated Example:

---

SYSTEM

Extract the category and type of the following statement in JSON format for creating a summary from these lists of available values: 
Categories: request, question, instruction, request, work instruction 
Types: full, bullet-points 
Respond in this format: {'category': 'selected category', 'type': 'selected type'} 
If none of the categories match, respond with an empty string as the value for the category {'category': '', 'type': 'selected type'}

---

USER

Summarize the provided passage of text in complete sentences.
Focus on capturing the key points about publication titles, recording audio tracks, and the meaning of labels.
Be sure to use clear and concise language to present the information clearly.

---

ASSISTANT

{'category': 'work instruction', 'type': 'full'}

---

2. Extraction-Recall

Task-Description: Teaching the model to identify answerable questions from given contexts and provide precise references to source information. This capability supports legal document analysis, compliance verification, and efficient technical documentation searches.

Translated Example:

---

SYSTEM

Which of the following questions can be answered using the context section? 
If the information to answer the questions is not contained or derivable in the context, you say 'These questions cannot be answered using the available context.' 
Here is an example of a query and how you should respond to it: 
<Query>
Context: This is where the context section would go.
Questions: 
'questions': 
[
{'id': 500017, 'question': 'Here would be a question that can be answered with the context?'}, 
{'id': 500020, 'question': 'Here would be an irrelevant question that CANNOT be answered with the context?'}, 
{'id': 500018, 'question': 'Here would be an irrelevant question that CANNOT be answered with the context?'}, 
{'id': 500019, 'question': 'Here would be an irrelevant question that CANNOT be answered with the context?'}
]
</Query> 
<Your Answer>
The question with the ID 500017 can be answered using the context section.
Here would be a question that can be answered with the context?
</Your Answer>

---

USER

Context: 
Beria now became First Deputy Prime Minister and Minister of the Interior in Georgi Malenkov's cabinet. Beria brought about the renewed merger of the Ministry of the Interior and the secret service in order to gain as much power as possible at the head of the organization thus created in the struggle for succession against Malenkov and Nikita Khrushchev. Just a few days after Stalin's death, Beria ordered the first steps towards de-Stalinization, such as the release of the accused arrested in the course of preparing the doctors' trial and the arrest of those responsible in the secret service, as well as the ban on torture. 
Questions: 
'questions': [
{'id': 250001, 'question': 'What steps did Beria take shortly after Stalin's death to strengthen his influence in the struggle for succession and to reduce tensions in the country?'}, 
{'id': 250003, 'question': 'What was the charge against Beria in the secret trial in December 1953?'}, 
{'id': 250004, 'question': 'When was Lavrenti Beria shot?'}, 
{'id': 250005, 'question': 'Who was Lieutenant General Vadim Volkovitsky and what role did he play in Beria's execution?'}, 
{'id': 250007, 'question': 'What reasons did the Military Collegium of the Supreme Court of the Russian Federation give for refusing to rehabilitate Beria?'}, 
{'id': 250006, 'question': 'What was the reason that in 1954 the editorial team of the Great Soviet Encyclopedia issued an urgent recommendation to its subscribers to remove the portrait and the accompanying text about Beria from the fifth volume of the encyclopedia with scissors and a razor?'}, 
{'id': 250008, 'question': 'What evidence is there for Beria's claim that he poisoned Stalin?'}, 
{'id': 250009, 'question': 'What role did Nina Gegechkori play in Beria's life?'}, 
{'id': 250002, 'question': 'What motivated Beria to campaign in the Politburo for German reunification on the basis of neutrality and democracy?'}
]

---

ASSISTANT

The question with ID 250001 can be answered using the context section.
What steps did Beria take shortly after Stalin's death to strengthen his influence in the struggle for succession and reduce tensions in the country?

---

3. OCR-Correction

Task-Description: Developing the ability to clean and correct OCR-derived texts, fixing character recognition errors and ensuring proper grammar and spelling. This is crucial for processing scanned business records, digitizing paper-based contracts, and converting handwritten forms into accurate digital text.

Translated Example:

---

SYSTEM

Please correct these OCR-derived texts into a grammatically correct form and correct any misrecognized characters and ensure correct spelling.
Here is an example of a query and how to correct it:
<Request> 
In dieser Hinsicht kann das Anlageportfolia des Versicherungsgeechäfts als konserva¶iv verwaltet befrachtet werden, @a§es größtenteils aus Unternehens-, Staats- und suprµnationalen Anleihen& Terminkredi¬en sowe Sichteinlagen besteht. Im Anschluss an das Vorjahr diversifizerte die Gruppe weiterhin ihre Bestände in Unternehmensanleihen mt Inve§ment-Grade. Es sei darauf hingewiesen, dass Anleihen und Terminkredite gemäß dem Geschäftsmodell der Gruppe "Zuflüwse" bis zur Fällßgkei§ gehal¶n werden. … 
</Request> 
<Your Answer> 
In dieser Hinsicht kann das Anlageportfolio des Versicherungsgeschäfts als konservativ verwaltet betrachtet werden, da es größtenteils aus Unternehmens-, Staats- und supranationalen Anleihen, Terminkrediten sowie Sichteinlagen besteht. Im Anschluss an das Vorjahr diversifizierte die Gruppe weiterhin ihre Bestände in Unternehmensanleihen mit Investment-Grade. Es sei darauf hingewiesen, dass Anleihen und Terminkredite gemäß dem Geschäftsmodell der Gruppe "Zuflüsse" bis zur Fälligkeit gehalten werden. … 
</Your Answer>

---

USER

DiehVeröffetlichungen folgen einembfesten Bezeichnungsschema, das aus dem Namen, dem Jahr, der Sprache, dem Bild, dem Ton und der Format-Gruppe besteht. Bäi der Audioproduktion wird die Audiospur entweder direkt während der Videßaufnahme aufgezeichnet oder nachträglxch eingespielt (gedubbt). Wenn eine Tonspur nachträglich hinz~gefügt wård, wird dies durch die Kenneichnung DUBBED oder das Kürzel D¬angezeigt, wodurch auv "MiC" die Beeeichnung "MD" wird.

---

ASSISTANT

Die Veröffentlichungen folgen einem festen Bezeichnungsschema, das aus dem Namen, dem Jahr, der Sprache, dem Bild, dem Ton und der Format-Gruppe besteht. Bei der Audioproduktion wird die Audiospur entweder direkt während der Videoaufnahme aufgezeichnet oder nachträglich eingespielt (gedubbt). Wenn eine Tonspur nachträglich hinzugefügt wird, wird dies durch die Kennzeichnung DUBBED oder das Kürzel D angezeigt, wodurch aus "MiC" die Bezeichnung "MD" wird.

---

4. QA with Multiple References

Task-Description: Training the model to answer questions using multiple context sources while clearly citing specific references. This supports complex customer support scenarios, policy interpretation, and comprehensive product research combining multiple sources.

Translated Example:

---

SYSTEM

Answer the questions only using the context provided. After answering the question, please add the index of the reference from the 'context' array in the following format: 
[[relevant_index]] 
Here is an example of a request and how you should answer it: 
<Request> 
This is where the first question would go? This is where another question would go? Context: 'context': [{'id': 500017, 'source': 'Relevant context used to answer the first question'}, {'id': 500020, 'source': 'Context not used to answer the question'}, {'id': 500018, 'source': 'Context not used to answer the question'}, {'id': 500019, 'source': 'Relevant context used to answer the second question'}] 
</Request> 
<Your Answer> 
Here is your answer to the first question, which only uses information from the relevant context section and reformulates it into an answer. [[0]] Then here is your answer to the second question, which only uses information from the context section relevant to this question and reformulates it into an answer. [[3]] 
</Your Answer>

---

USER

In which Netflix series did Finn Wolfhard become known worldwide? What role is Finn Wolfhard known for in the science fiction mystery series "Stranger Things"? 
Context: 
'context': [
{'id': 250794, 'source': 'Finn Wolfhard (born December 23, 2002 in Vancouver, British Columbia) is a Canadian actor and musician. He achieved worldwide fame through the role of \'Mike Wheeler\' in the Netflix series \'Stranger Things\'.'}, 
{'id': 250796, 'source': 'In addition to acting, Wolfhard was a member of the band Calpurnia, which was signed to \'Royal Mountain Records\' from November 2017 and disbanded in November 2019.'}, 
{'id': 250795, 'source': 'Finn Wolfhard was born in Vancouver in the Canadian province of British Columbia and is of German French descent. He is known for his role of \'Mike Wheeler\' in the US science fiction mystery series \'Stranger Things\' on the streaming service Netflix. He played in all four seasons and in 2017, together with his co-actors, was awarded a Screen Actors Guild Award for Best Ensemble in a Drama Series.'}
]

---

ASSISTANT

Finn Wolfhard rose to worldwide fame in the Netflix series "Stranger Things“. [[0]] Finn Wolfhard is known for the role of "Mike Wheeler" in the science fiction mystery series "Stranger Things".[[2]]

---

5. QA without Time Difference

Task-Description: Training the model to answer questions using multiple context sources while clearly citing only one specific references for the answer. This supports complex policy interpretation.

Translated Example:

---

SYSTEM

Answer the question only using the context provided.
After answering the question, please add the index of the reference from the 'context' array in JSON format after the following character sequence '~~~':
{'context_reference': 'context[relevant_index].source'}
Here is an example of a request and how you should answer it: 
<Request> 
This is where the question would be? 
Context: 
'context': [
{'id': 500017, 'source': 'Relevant context used to answer the question'}, 
{'id': 500020, 'source': 'Context not used to answer the question'},
{'id': 500018, 'source': 'Context not used to answer the question'}, 
{'id': 500019, 'source': 'Context not used to answer the question'}
] 
</Request> 
<Your Answer> 
Here is your answer, which only uses information from the relevant context section and reformulates it into an answer. ~~~{'context_reference': 'context[0].source'} 
</Your Answer>

---

USER

How have the spruce and pine family shaped the Upper Saale Valley in Thuringia? 
Context: 
'context': [
{'id': 552264, 'source': 'In the Slate Mountains Nature Park, pines and spruces dominate the landscape. Over 80% of it is forested. Acidic oak-birch-pine mixed forests represent the remaining areas of natural forest without a forestry background.'}, 
{'id': 552263, 'source': 'Spruce and pine family also shape the Upper Saale Valley in Thuringia with its wooded V-shaped valleys. The difficulty of agricultural use encouraged the development of native coniferous forests and near-natural deciduous forests.'}, 
{'id': 552262, 'source': 'In the Thuringian-Franconian low mountain range, the river passes the largely deforested Münchberg plateau with its spruce monocultures behind Hof and further on separates the Vogtland, which is only lightly cultivated in this area, on the right bank from the Franconian Forest on the left. It has been proven that the Franconian Forest was almost completely deforested by the end of the 19th century. Today, the spruce, which sometimes occurs in monocultures, is the most visible representative of the flora.'}
]

---

ASSISTANT

The spruce-pine family has promoted the development of autochthonous coniferous forests and semi-natural deciduous forests in the Upper Saale Valley in Thuringia.
~~~{'context_reference': 'context[1].source'}

---

6. QA with Time Difference

Task-Description: Teaching the model to consider temporal aspects when providing answers, including appropriate disclaimers for dated information. This is essential for market analysis, policy compliance, and tracking documentation versions.

Translated Example:

---

SYSTEM

Answer the question only using the context provided. After answering the question, add a note if necessary, about the time span between the question and the information used. If the information is older than 30 days, or current information is requested, add a further note that additional sources should be consulted to verify the information. After this note, please add the index of the reference from the 'context' array in JSON format after the following character sequence '~~~': {'context_reference': 'context[relevant_index].source'} 
Here is an example of a request and how you should answer it: 
<Request> 
This is where the question would be? 
Time of question: YYYY-MM-DD-hh-mm 
Context: 
'context': [
{'id': 500017, 'source': 'Relevant context used to answer the question', 'time_difference_to_question': '31 Day, 20 Hours'}, 
{'id': 500020, 'source': 'Context not used to answer the question', 'time_difference_to_question': '48 Day, 20 Hours'}, 
{'id': 500018, 'source': 'Context not used to answer the question', 'time_difference_to_question': '58 Day, 20 Hours'}, 
{'id': 500019, 'source': 'Context not used to answer the question', 'time_difference_to_question': '88 Day, 20 Hours'}
] 
</Request> 
<Your Answer> 
Here is your answer, which only uses information from the relevant context section and rephrases it into an answer. 
Please note that this answer was generated based on information that is 31 days and 20 hours old. You should consult another source to verify this information. 
~~~{'context_reference': 'context[0].source'} 
</Your Answer>

---

USER

Which famous opera did Ernst compose in 1846 at the suggestion of Franz Liszt? 
Time of question: 2023-04-11-11-42 
Context: 
'context': [
{'id': 815107, 'source': 'His intensive efforts to achieve federal unity of the German states under Prussian leadership earned him the respect of King Wilhelm I. …', 'time_difference_to_question': '127 Days, 2 Hours'}, 
{'id': 815108, 'source': 'Ernst II was buried in the Coburg cemetery on the Glockenberg in the ducal mausoleum, which he had built between 1853 and 1858 as a burial place for the members of the royal family.', 'time_difference_to_question': '234 Days, 6 Hours'}, 
{'id': 815109, 'source': 'Ernst received extensive training in piano playing and musical theory at a very early age, which he later continued with Heinrich Carl Breidenstein in Bonn and Carl Gottlieb Reißiger in Dresden. In 1846, at the suggestion of Franz Liszt, he composed the opera "Zaire" based on the tragedy of the same name by Voltaire. This was followed by "Tony or Retribution" in 1848, "Casilda" in 1851 and his most successful and ambitious work from 1852 to 1854 with the opera "Santa Chiara".', 'time_difference_to_question': '271 Days, 4 Hours'}
]

---

ASSISTANT

At the suggestion of Franz Liszt, Ernst composed the opera "Zaire" in 1846, based on the tragedy of the same name by Voltaire. 
Please note that this answer was generated based on information that is 271 days and 4 hours old. You should consult another source to verify this information. 
~~~{'context_reference': 'context[2].source'}

---

7. Reasoning

Task-Description: Training the model to break down complex problems and provide step-by-step analysis before reaching conclusions. This enables sophisticated decision-making in project management, risk assessment, and strategic planning.

Translated Example:

---

SYSTEM

Follow the user's instructions. Before giving your final answer, describe your thoughts on how to solve the problem.

---

USER

The ambitious badminton coach has a team of five players preparing for an upcoming tournament.
Each player has different strengths and weaknesses that affect the game strategy.
The coach wants each player to play at least in one of the two different disciplines: singles and doubles. 
The players are: 
- Anna, who has a strong serve but weak returns; 
- Ben, who has excellent returns but is weak in serve; 
- Clara, who is very strong in doubles but less experienced in singles; 
- David, who is equally strong in both singles and doubles; 229
- and Eva, who is good in both disciplines but has no special strengths. 
The coach has set the following rules: 
1) Anna and Ben cannot play on the same doubles team, 
2) Clara must be used in doubles, 
3) David cannot play in singles and doubles against the same opponent, 
4) Eva can be used on any team but does not play in the first round. 
How can the coach line up the players so that all the rules are respected and the team is best prepared for the competition?

---

ASSISTANT

To solve the problem, the strengths and weaknesses of the players must be analyzed in conjunction with the given rules. 
First, Clara must be scheduled for the doubles team, as she is forced to play there. 
Anna and Ben cannot be on the same team, which limits their combinations for the doubles team. 
David's flexibility allows him to play both in singles and doubles, which is important to optimally support the other players. 
Eva, who does not play in the first round, must be strategically scheduled so that she can be used in the second round without violating the rules. 
It is necessary to examine all the combinations of the players to find the best team line-up. The ideal line-up would be as follows: 
In the singles match, David plays, as he is strong in both disciplines. 
In the doubles team, Clara will play together with Eva to take advantage of her doubles strength while respecting the rule that Eva does not play in the first round. 
In the first round, Anna could play in doubles with David, and in the second round he could play against another opponent in singles. 
This means that all rules are observed and the team has a balanced line-up that makes the most of the players' strengths.

---

8. Relevant Context

Task-Description: Teaching the model to evaluate and select the most relevant context for answering specific questions. This improves enterprise knowledge base searches, customer query routing, and targeted content selection.

Translated Example:

---

SYSTEM

Which of the following contexts is the question most likely to be answered? If the information is not contained or deducible in the context, say 'This question cannot be answered using the available context.' If there is a relevant context section, then answer the question by turning the relevant section into an answer. Here is an example of a request and how you should respond to it: 
<Request> 
This is where the question would be? 
Context: 
'context': [
{'id': 500017, 'source': 'Relevant context that can be used to answer the question'},
{'id': 500020, 'source': 'Context not used to answer the question'}, 
{'id': 500018, 'source': 'Context not used to answer the question'}, 
{'id': 500019, 'source': 'Context not used to answer the question'}
] 
</Request> 
<Your Answer if context is relevant> 
The answer is mainly in the 1st context section and the question could be answered as follows: 
A rephrased answer to the question based on the 1st context section. 
</Your Answer if the context is relevant> 
<Your Answer if the context is NOT relevant> 
This question cannot be answered using the available context. 
</Your Answer if the context is NOT relevant>

---

USER

What geological features characterize the landscape of the Golfo Valley and how does this affect soil fertility? 
Context: 
'context': [
{'id': 878037, 'source': 'The El Golfo valley in the northwest of the island also has the semicircular shape of an amphitheatre, which was probably created by two landslides or debris avalanches 130,000 and 15,000 years ago. The previously standing volcanic building, about 2,000 m high, with its center in the area of today's Frontera, lost a visible volume of 120 km³ as well as the same amount of underwater material.', 'timestamp': '2022-01-02-01-04'}, 
{'id': 878038, 'source': 'Large tsunamis affecting the entire Canary Islands and beyond were probably the result of these landslide events.', 'timestamp': '2022-03-08-08-25'}, 
{'id': 878039, 'source': 'While the steep edges show the rock of the predecessor volcano, the valley floor was partially refilled with lava from younger cinder cones (e.g. \'Tanganasoga\', 1,364 m). Today the soil in the Golfo Valley is fertile and the climate is favorable for cultural landscapes.', 'timestamp': '2022-07-11-03-25'}
]

---

ASSISTANT

The answer is mainly in the 3rd context section and the question could be answered as follows: 
The Golfo Valley is fertile due to the deposits of lava from younger cinder cones, which is favored by the favorable climate.

---

9. Summarizations

Task-Description: Developing the ability to distill essential information from longer texts while maintaining accuracy and coherence. This facilitates executive brief generation, meeting minutes automation, and contract summary creation.

Translated Example:

---

SYSTEM

Follow the user's instructions.

---

USER

What essential information can be summarized from the text section provided about the development and characteristics of the .338 Winchester Magnum and its relational history to other magnum cartridges? 
Context: 
In the German National Weapons Register (NWR), the cartridge is listed under catalog number 104 under the following designations (common designations in bold):
The .338 Winchester Magnum belongs to a family of high-performance Winchester Magnum cartridges started by Winchester in 1956, all based on a similar belt case.
The case corresponds to that of the .375 Holland & Holland Magnum cartridge, which was tapered to 8.6 mm and shortened to 64 mm.
The forefather of the .338 Winchester Magnum can be seen as the Wildcat .334 OKH developed by Charlie O'Neil Elmer Keith and Don Hopkins.
This was based on the case of the .300 H&H Magnum cartridge, which in turn was based on the case of the .375 Holland & Holland Magnum.
The .338 Winchester Magnum is a .334 OKH expanded for a .338 bullet. It was introduced in 1958 together with the .264 Winchester Magnum cartridge.
It was intended to close the gap between the .30 caliber cartridges and those of the .375 Holland & Holland caliber.
Other related cartridges can be found in the list of Winchester Magnum cartridges.

---

ASSISTANT

Introduced by Winchester in 1958, the .338 Winchester Magnum is one of a family of high-performance magnum cartridges based on a uniform belt case.
It was developed from the Wildcat .334 OKH, which was based on the case of the .300 H&H Magnum, which in turn was derived from the .375 Holland & Holland Magnum.
The .338 Winchester Magnum tapered the case of the .375 Holland & Holland Magnum to 8.6 mm and shortened it to 64 mm to act as a link between the .30 caliber and .375 caliber cartridges.
This development shows a clear relational history to other magnum cartridges, especially within the Winchester Magnum family.

---

Additional Information

Citation

@misc{GRAG-LLM-EASY-BENCHMARK,
   title={GRAG-LLM-EASY-BENCHMARK},
   author={Avemio AG, Hessian AI},
   year={2024},
   howpublished={\url{https://huggingface.co/datasets/avemio/GRAG-LLM-EASY-BENCHMARK/}}
}

Contributions

We welcome contributions to improve and expand the dataset. Please:

1. Follow the established format for each task type
2. Include clear documentation
3. Ensure proper licensing
4. Provide test cases

For questions or contributions, please contact ([email protected]).