Retrieval-Synthetic-NVDocs-v1 is a synthetic retrieval dataset generated based on NVIDIA's public content, released by NVIDIA in 2026. The dataset is generated using the NeMo Data Designer framework and is designed for training and evaluating embedding models and Retrieval-Augmented Generation (RAG) systems.

The dataset contains 15,095 records, covering 105,606 deduplicated question-answer pairs, with a data volume between 100 K and 1 M. The dataset covers various reasoning types, including factual, relational, procedural, inferential, temporal, causal, and visual reasoning, as well as multiple query types such as structural, multi-hop, and contextual queries. It also includes document artifacts such as extracted entities, key concepts, relationships, and technical terms, providing diverse training signals for retrieval and understanding tasks.

Dataset Composition

• Top-level fields: Include file_name (source file path), text (full document text), chunks (array of document chunk objects), sections_structured (formatted section strings), bundle_id (multi-document bundle identifier), document_artifacts (object containing extracted metadata and semantic information), qa_generation (object containing generated question-answer pairs), deduplicated_qa_pairs (array of deduplicated question-answer pairs), and qa_evaluations (object containing quality assessments of QA pairs).
• Chunk object fields: Each chunk contains chunk_id, doc_chunk_index, start and end (position indices), sentence_count, text, and word_count.
• Document artifact fields: Include contextual_factors, entities, insights, key_concepts, processes, relationships, technical_terms, and themes, each containing text, description, and importance.
• QA generation fields: Question-answer pairs include question, answer, query_type, reasoning_type, question_complexity, hop_count, hop_contexts, and segment_ids.
• QA evaluation fields: Include accuracy, clarity, context_support, relevance, overall, and improvements, where the first four include justification and score.