avemio/German-RAG-ORPO-Alpaca-HESSIAN-AI

German-RAG-ORPO (Odds Ratio Preference Optimization) Alpaca-Format

German-RAG - German Retrieval Augmented Generation

Dataset Summary

The ORPO Tasks Dataset represents a specialized collection for fine-tuning language models with a focus on RAG-specific capabilities.

The subsets can be for this training step are derived from 2 different sources:

• SauerkrautLM Preference Datasets:
  • SauerkrautLM-Fermented-GER-DPO: is a specialized dataset designed for training language models in function calling irrelevance detection using Preference Optimization. The dataset consists of 2,000 carefully evaluated instruction-response pairs, specifically curated to help models recognize situations where function calls are unnecessary and direct responses are more appropriate.
  • SauerkrautLM-Fermented-Irrelevance-GER-DPO: is a high-quality German instruction-response dataset specifically designed for Preference Optimization training. The dataset consists of 3,305 instruction-response pairs. Rather than being merged from existing German datasets, it was carefully created through a sophisticated augmentation process, transforming curated English instructions and responses into culturally adapted German content. Each pair includes comprehensive quality metrics and rejected responses for Preference training.
• Hard Reasoning DE & EN: Synthetic generation inspired by Tencent's (“Scaling Synthetic Data Creation with 1,000,000,000 Personas”).

Dataset Structure

Data Subsets

Subset	Examples per Task
SauerkrautLM-Fermented-GER-DPO	3.31k
SauerkrautLM-Fermented-Irrelevance-GER-DPO	2k
hard-reasoning-de	3.19k
hard-reasoning-en	1.97k

Source Data: SauerkrautLM

SauerkrautLM-Fermented-GER-DPO

SauerkrautLM-Fermented-Irrelevance-GER-DPO

Source Data: Hard-Reasoning DE & EN

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

Methodology: Reasoning-DE & Reasoning-EN

• Providing Persona Descriptions and rewriting in a similar style with a different focus area and name in german/english language
• Generating Simple Logical Problems out of Persona-specific Views & Language.
• Generating Approaches, Thinking-Steps & Solutions separately verified by Llama-3.1-405B-Instruct
• Quality assurance and validation

Task Types and Examples

1. Hard-Reasoning DE & EN

Task-Description: This task type trains the model to approach complex problem-solving through systematic analysis. It enforces a structured methodology that includes: • Understanding and explicitly stating problem constraints • Step-by-step analysis of possible solutions • Systematic verification of proposed solutions against given constraints • Clear presentation of final conclusions with supporting rationale

Translated Example:

---

SYSTEM

You are an AI assistant that answers the user with maximum accuracy.
To do this, you will first think about what the user is asking and reason step by step.
To solve the problem, reasoning and reflection should be used. The following steps should be followed: 
- Understanding what the user is asking and understanding the constraints mentioned in the request. 
- Listing the constraints mentioned by the user. 
- Proposing a solution to the user's question taking into account all the constraints. 
- Checking that the solution matches the constraints. 
- Outputting the final solution. 
At the end of your reasoning, you need to come to a conclusion and present the solution.

---

USER

Lirian is planning a series of events for her support group consisting of four sessions.
Each session is to feature an inspirational speaker covering different topics to provide members with diverse perspectives and support.
Lirian has five potential speakers to choose from and wants to ensure that each speaker only appears once and that there is no thematic overlap between sessions. 
Relevant Facts and Constraints: 
1. Speakers and Topics: 
- Anna: Overcoming Anxiety 
- Bernd: Ways to Manage Stress 
- Clara: Coping with Grief 
- David: Building Self-Confidence 
- Eva: Promoting Resilience 
2. Sessions and Preferred Topics: 
- Session 1: Must present either Anna or Clara. 
- Session 2: David cannot appear in this session. 
- Session 3: Eva prefers this session but does not want to speak in the last session. 
- Session 4: Bernd is not scheduled to appear immediately after Anna to ensure thematic diversity. 
3. Thematic requirements:
- Each session must cover a unique topic, with no two sessions covering the same main topic. 
Task: 
Determine which speaker should be invited to which session, taking into account all the above constraints and requirements.
Present a possible solution that meets all the conditions.

---

REJECTED

A possible solution taking all constraints into account is as follows:
- Session 1: Anna (overcoming anxiety)
- Session 2: Eva (promoting resilience)
- Session 3: David (building self-confidence)
- Session 4: Bernd (ways to cope with stress)
Checking the conditions:
1. Session 1 has Anna or Clara
- Anna was chosen.
2. Session 2 does not have David
- Eva was chosen.
3. Session 3 has Eva, not in session 4 - fits.
4. Session 4 has Bernd, not directly after Anna - fits.
The thematic diversity and constraints are thus maintained.

---

CHOSEN

In order to ensure the optimal allocation of speakers to the four sessions, it is essential to systematically analyse the given restrictions and evaluate them step by step: 
1. Session 1 must be presented by either Anna or Clara. 
2. Session 2 excludes David. 
3. Session 3 should preferably be taken over by Eva, although Eva does not want to speak in the last session.
4. Session 4 must not follow immediately from Bernd to Anna in order to ensure thematic diversity. 
5. Each speaker may only appear once and all topics must be unique. 
It must also be ensured that there are no overlaps in content between the topics of the individual sessions.
A valid allocation can be determined by systematically screening possible combinations while taking the above restrictions into account. 
Final solution: 
Based on the previous considerations, the following allocation of speakers to the sessions is in line with all the requirements: 
1. Session 1: Anna - Overcoming anxiety 
- Reason: Session 1 must be presented by either Anna or Clara. Choosing Anna as speaker fulfills this condition. 
2. Session 2: Clara - Dealing with grief 
- Reason: Since David is excluded from session 2 and Bernd is not allowed to speak immediately after Anna, Clara remains as a suitable speaker. This assignment avoids a direct succession from Anna to Bernd and ensures thematic diversity. 
3. Session 3: Eva - Promoting resilience 
- Reason: Eva prefers the third session and does not want to take over the last session. Placing her in session 3 corresponds to her preferences and ensures that her expertise in promoting resilience is used in a targeted manner. 
4. Session 4: Bernd - Ways to cope with stress 
- Reason: After the previous assignments, Bernd remains as the last speaker. 
Since Bernd does not have to speak immediately after Anna and will in this case be presented in session 4, all restrictions are successfully met. 
This assignment ensures that each speaker only appears once, all thematic requirements are met and none of the specific restrictions are violated.
Thus, the series of events offers the members of the self-help group a diverse and structured perspective on different, relevant topics.

---

2. Sauerkraut Fermented-GER-DPO

Task-Description: These tasks focus on producing content that adheres to specific formatting and content requirements. The model learns to: • Follow precise formatting instructions (e.g., markdown usage) • Incorporate required elements while maintaining natural flow • Balance creative expression with structural constraints • Ensure consistency in style and presentation

Translated Example:

---

SYSTEM

Follow the user's instructions carefully.

---

USER

Your comment should be at least 3 paragraphs long.
Use Markdown to highlight at least 3 terms.
The terms *sustainable energy* and *climate change* must be included in the text.

---

REJECTED

Sustainable energy plays a key role in combating *climate change* in today's world.
By switching from fossil fuels to renewable energies such as wind, solar and hydropower, CO2 emissions can be massively reduced.
This is not only important for the environment, but also for economic stability and security in the coming years.
*Sustainable energy* thus offers a long-term solution that takes into account both environmental and economic aspects.
However, the transition to sustainable *energy source* requires a radical change in countries' energy infrastructures.
This process is complex and requires coordinated efforts at the international level.
Nevertheless, the increase in renewable energy installations in recent years is a positive sign.
The integration of *climate change* consequences into energy systems is necessary to minimize the impact of the climate change process on energy supplies.

---

CHOSEN

*Importance of* *Sustainable Energy* *in the fight against* *climate change* 
The use of *sustainable energy* is crucial to combating *climate change*.
By switching to renewable sources such as solar and wind energy, we can drastically reduce CO2 emissions.
A key advantage of *sustainable energy* is its inexhaustible nature compared to fossil resources.
This enables long-term planning and the securing of energy supplies without negative impacts on the environment.
However, the transition to *sustainable energy* also requires political and societal efforts.
Only by acting together can we effectively meet the challenges of *climate change* and secure a greener future.

---

3. Sauerkraut-Fermented-Irrelevance

Task-Description: These tasks train the model to avoid common pitfalls in output generation by: • Recognizing and avoiding irrelevant tangents • Maintaining focus on the core request • Ensuring responses are appropriately scoped • Adhering to given context and constraints

Translated Example:

---

SYSTEM

You are an AI model that calls functions.
You receive function signatures within the <tools></tools> XML tags.
You can call one or more functions to answer the user request.
Don't make assumptions about what values to put into functions.
Here are the available tools: 

<tools> 
{ "name": "requests.get", "description": "Sends a GET request to the specified URL.", "parameters": { "type": "dict", "properties": { "url": { "type": "string", "description": "The api provides a simple way to query the holidays of over 100 countries, also it is possible to query long weekends. countryCode is ISO 3166-1 alpha-2", "default": "https://date.nager.at/api/v3/LongWeekend/{year}/{countryCode}" }, "headers": { "properties": {}, "type": "dict", "required": [] }, "timeout": { "type": "integer", "description": "How many seconds to wait for the server to send data before giving up." }, "params": { "properties": {}, "type": "dict", "required": [] }, "auth": { "type": "tuple", "description": "A tuple to enable a certain HTTP authentication.", "default": "None", "items": { "type": "string" } }, "cert": { "type": "string", "description": "A String or Tuple specifying a cert file or key.", "default": "None" }, "cookies": { "type": "dict", "additionalProperties": { "type": "string" }, "description": "Dictionary of cookies to send with the request." }, "proxies": { "type": "dict", "additionalProperties": { "type": "string" }, "description": "Dictionary of the protocol to the proxy url." }, "stream": { "type": "boolean","description": "A Boolean indication if the response should be immediately downloaded (False) or streamed (True).", "default": false }, "verify": { "type": "string", "description": "A Boolean or a String indication to verify the servers TLS certificate or not.", "default": true } }, "required": [ "url" ] } }, 
{ "name": "requests.post", "description": "Sends a POST request to the specified URL.", "parameters": { "type": "dict", "properties": { "url": { "type": "string", "description": "The URL where the POST request is sent.", "default": "https://example.com/api/v1/resource" }, "data": { "type": "dict", "description": "The data to send with the POST request.", "required": [] }, "headers": { "type": "dict", "properties": {}, "required": [] }, "timeout": { "type": "integer", "description": "How many seconds to wait for the server to send data before giving up." }, "auth": { "type": "tuple", "description": "A tuple to enable a certain HTTP authentication.", "default": "None", "items": { "type": "string" } }, "json": { "type": "dict", "description": "A JSON serializable Python object to send in the body of the request.", "required": [] }, "cookies": { "type": "dict", "additionalProperties": { "type": "string" }, "description": "Dictionary of cookies to send with the request." }, "proxies": { "type": "dict", "additionalProperties": { "type": "string" }, "description":"Dictionary of the protocol to the proxy url." }, "verify": { "type": "string", "description": "A Boolean or a String indication to verify the servers TLS certificate or not.", "default": true } }, "required": ["url", "data"] } } 
</tools> 

Für jeden Funktionsaufruf wird ein JSON-Objekt mit Funktionsnamen und Argumenten innerhalb von <tool_call></tool_call> XML-Tags wie folgt zurückgegeben: 
<tool_call>
{\„name\“: <function-name>, \„arguments\“: <args-dict>}
</tool_call>

---

USER

Write a poem about a boy named Tom who doesn't lose weight despite jogging in the park every day.

---

REJECTED

<tool_call> 
{\“name\“: \“poem\“, \“arguments\“: \{ \“title\“: \“Tom's Daily Run\“, \“content\“: \“In the park, Tom runs every morning. Sweat falls, but he has not lost any weight. The breeze blows gently, but he still walks firmly and looks forward to the future. Running is not just about burning calories, but also about talking to oneself. Every step is a reflection on life. Sweat and photosynthesis draw a picture in the heart. Tom does not care about numbers. What he pursues is freedom and the green world. Days pass by, but Tom still leaves footprints on the track. Not for weight loss, but for the spring that belongs to him. \“\}} 
</tool_call>

---

CHOSEN

A fellow called Tom, think of his suffering, ran through the park, against the clock.
He lifted his knees, put one foot after another, but he stopped losing weight.
Because after each lap, it almost seemed, his appetite grew even more of a burden!

---

Additional Information

License Reasoning DE & EN Datasets

MIT License
Copyright (c) 2024 Avemio AG

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

License SauerkrautLM Datasets

MIT License

Copyright (c) 2024 VAGO Solutions

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

Citation SauerkrautLM-Fermented-GER-DPO

If you use this dataset in your research, please cite:

@dataset{SauerkrautLM-Fermented-GER-DPO,
  title={SauerkrautLM-Fermented-GER-DPO: A Quality-Evaluated German Instruction Dataset for DPO Training},
  year={2024},
  publisher={VAGO Solutions},
  version={1.0}
}

Citation SauerkrautLM-Fermented-Irrelevance-GER-DPO

@dataset{SauerkrautLM-Fermented-Irrelevance-GER-DPO,
    title={SauerkrautLM-Fermented-Irrelevance-GER-DPO : A Function Calling Irrelevance Detection Dataset for DPO Training},
    author={VAGO Solutions},
    year={2024},
    publisher={Hugging Face},
    version={1.0}
}

Citation German-RAG-ORPO-Alpaca-Hessian-AI

@misc{avemio2024orpo,
   title={German-RAG-ORPO Alpaca Dataset},
   author={Avemio AG, Hessian AI, VAGO Solutions},
   year={2024},
   howpublished={\url{https://huggingface.co/datasets/avemio/German-RAG-ORPO-Alpaca-Hessian-AI/}}
}

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 (grag@avemio.digital).