RAG Scientific QA Dataset: 711 Question-Answer Pairs from ArXiv Papers

711 high-quality Question-Answering pairs synthetically generated from 200 ArXiv papers in Computer Science fields like AI and CL. The dataset was created by author xunnhi using gpt-4o-mini with strict anti-hallucination rules and was last updated on 2026-06-18. It is designed for fine-tuning Large Language Models on Retrieval-Augmented Generation tasks.

Use Cases

• Fine-tuning LLMs for scientific QA based on the 711 synthetically generated question-answer pairs.
• Benchmarking RAG system performance on computer science literature based on the ArXiv paper source.
• Training models to answer questions with reduced hallucination based on the described strict generation rules.

Strengths

• Contains 711 high-quality QA pairs, providing a substantial corpus for model tuning.
• Source data comprises 200 ArXiv papers, offering a foundation in computer science topics.
• Generation method employed gpt-4o-mini with strict rules aimed at preventing hallucination.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.
• Last updated 2026-06-18 17:52:27; freshness should be verified as the date is in the future.

Provenance

Source

200 ArXiv papers from Computer Science subfields (AI, CL, LG, IR).

Collection Method

Synthetically generated using gpt-4o-mini with strict anti-hallucination rules.

Freshness

Last updated 2026-06-18 17:52:27.