Multi-subject RLVR
Multithematic QA corpus from the Chinese exam ExamQA, including nearly 580,000 question-answer pairs translated into English. The data covers 48 university disciplines (law, medicine, medicine, computer science, economics, etc.) classified into 4 main areas: sciences, humanities, social and applied sciences.
Description
Multi-subject RLVR is a massively multi-thematic academic question and answer dataset. Based on the Chinese ExaMQA corpus, it offers QA pairs translated into English, with automatic categorization of topics using GPT-4O-mini.
What is this dataset for?
- Train or evaluate multi-thematic QA models
- Perform fine-tuning on academic comprehension tasks
- Serve as a training base for RLHF or RLAIF approaches in an educational context
Can it be enriched or improved?
Yes. We could:
- Add explanations for each answer (step-by-step)
- Label difficulty levels
- Translate questions into other languages or reintroduce distractors for multiple choice QA
🔎 In summary
🧠 Recommended for
- Educational AI developers
- Multithematic QA researchers
- RLHF projects
🔧 Compatible tools
- OpenAssistant
- LangChain
- Transformers
- LLama
- Mistral
- Claude
💡 Tip
Filter the desired fields (STEM, law, economics, etc.) according to your use cases to specialize your model.
Frequently Asked Questions
Does this dataset only contain multiple choice questions?
No, the distractors have been removed to convert each example into a free QA pair (question + answer).
Is it possible to filter by field or subject?
Yes, each QA is automatically classified into one of 48 subjects or grouped into 4 general areas (STEM, social sciences, etc.).
Can this dataset be used to train an educational chatbot model?
Yes, it's a great choice for fine-tuning an academic assistant or an automated tutoring model.
