avemio/GRAG-CPT-HESSIAN-AI

GRAG-CPT (Continued Pre-Training) Tasks Dataset

GRAG - German-RAG - German Retrieval Augmented Generation

Dataset Summary

The CPT Tasks Dataset is a comprehensive collection designed for continued pre-training of language models, focusing on three core competencies: context-based question answering, structured reasoning, and summarization. The dataset comprises approximately 620,000 examples, with 420,000 in German and 200,000 in English.

Developed by Avemio AG, this dataset builds upon and enhances the German Wikipedia dump provided by Cohere (wikipedia-22-12-de-embeddings), expanding it with synthetic examples and structured tasks to create a robust training resource. The reasoning tasks which synthetic generation was inspired by the Paper from Tencent (“Scaling Synthetic Data Creation with 1,000,000,000 Personas”), to generate a diverse set of reasoning tasks across various domains.

Supported Tasks

Question Answering

Training examples that teach models to:

• Extract relevant information from provided context
• Generate accurate, context-based responses

Example structure: Question > Context > Context-based Answer

Structured Reasoning

Problems and solutions that develop:

• Systematic thinking approaches
• Multi-constraint problem solving
• Step-by-step solution development

Example structure: Task > Approach > Solution

Summarization

Examples that teach models to:

• Distill complex information into clear summaries
• Maintain key information while reducing length
• Structure output in bullet points or concise paragraphs

Dataset Structure

Data Subsets

Subset	Examples
Question-Answering	231,000
Reasoning-DE	200,000
Reasoning-EN	200,000
Summarization	23,000

Dataset Creation

Source Data: Question-Answering & Summarization

• Base: (cohere/wikipedia-22-12-de-embeddings)
• Enhancement: Synthetic data generation by Avemio AG
• Quality: Automatic validation and curation of examples by Open Source LLM's

Methodology: Question-Answering & Summarization

1. Extraction of base content from German Wikipedia
2. Enhancement through synthetic example generation
3. Structure addition for specific task types
4. Quality assurance and validation

Source Data: Reasoning-DE & Reasoning-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

1. Providing Persona Descriptions and rewriting in a similar style with a different focus area and name in german/english language
2. Generating Simple Logical Problems out of Persona-specific Views & Language.
3. Generating Approaches, Thinking-Steps & Solutions separately verified by Llama-3.1-70B-Instruct
4. Structure addition for specific task types
5. Quality assurance and validation

Additional Information

License

This dataset is licensed under CC-BY-SA 4.0, in accordance with the original Wikipedia content license.

Citation

@misc{avemio2024cpt,
   title={GRAG-CPT Tasks Dataset},
   author={Avemio AG, Hessian AI},
   year={2024},
   howpublished={\url{https://huggingface.co/datasets/avemio/GRAG-CPT-HESSIAN-AI/}}
}

Contributions

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

1. Follow our contribution guidelines
2. Maintain the established format for each task type
3. Provide clear documentation for new additions
4. Ensure proper licensing for all contributed content

For questions or contributions, please contact (grag@avemio.digital).