--- license: cc-by-nc-sa-4.0 task_categories: - question-answering - text-retrieval language: - en tags: - arXiv - RAG - scholarly-data - semantic-search - chromadb - embeddings --- # Dataset Card for Scholarly-Epistemic-Engine: arXiv cs.AI Corpus and Embeddings This dataset contains the processed text, metadata, and semantic vector embeddings of approximately 90,000 scholarly articles from the arXiv Computer Science - Artificial Intelligence (cs.AI) category, spanning from 1993 to December 2024. It is designed to support Retrieval-Augmented Generation (RAG) systems and semantic knowledge discovery. ## Dataset Details ### Dataset Description This dataset was generated as part of the research paper *"A Four-Stage Retrieval-Augmented Generation System for Semantic Knowledge Elicitation"*. It aims to address information overload in the AI domain by providing a ready-to-use, diachronic semantic knowledge base. The raw PDF documents were statelessly extracted, cleaned, and segmented into 5-sentence chunks. These chunks (totaling 7,496,671) were then embedded using the `sentence-transformers/all-mpnet-base-v2` model and indexed in a ChromaDB database. The dataset includes the raw extracted text, metadata, and the full pre-computed vector database (>50 GB). - **Curated by:** Aman Bhardwaj, Jeet Bhardwaj, and Mir Shahnawaz Ahmad - **Language(s) (NLP):** English (`en`) - **License:** CC-BY-NC-SA-4.0 (Creative Commons Attribution-NonCommercial-ShareAlike 4.0) **License & Copyright Disclaimer:** *The code used to generate this dataset is licensed under the MIT License on GitHub. However, this dataset itself is distributed under the CC BY-NC-SA 4.0 license strictly for academic and non-commercial research purposes. This dataset is a derivative work compiled from pre-print publications hosted on [arXiv.org](https://arxiv.org). The copyright for the underlying raw text and original manuscripts remains with the respective authors or their publishers. Users of this dataset must adhere to arXiv's terms of use and API usage guidelines.* ### Dataset Sources - **Repository (Source Code):** [Scholarly-Epistemic-Engine](https://github.com/GitHub-AmanBhardwaj/Scholarly-Epistemic-Engine) - **Paper:** *A Four-Stage Retrieval-Augmented Generation System for Semantic Knowledge Elicitation* ## Uses ### Direct Use This dataset is intended for researchers and developers building or evaluating academic AI systems. Direct use cases include: - Benchmarking Retrieval-Augmented Generation (RAG) architectures. - Developing domain-specific semantic search engines for scientific literature. - Analyzing temporal trends and diachronic synthesis within the AI research domain. - Evaluating cross-encoder reranking and LLM-based fusion techniques. ### Out-of-Scope Use - **Commercial Use:** Training commercial Large Language Models (LLMs) or building monetized platforms is strictly prohibited by the NC license and respects the original authors' copyright. - Generating definitive factual claims without verifying the original source manuscripts (hallucination risk). ## Dataset Structure The repository contains several key files and directories: - `metadata.csv`: Contains the master metadata catalog for all 89,375 scraped papers (Paper ID, Title, Publication Date, URL) collected during Phase 1. - `v1.csv`, `v2.csv`, `v3.csv`, `v4.csv`: These represent the iterative stages of the full-text extraction process using the metadata. `v4.csv` contains the final, complete raw extraction dataset before data cleaning was applied. - `final_data.csv`: The consolidated, cleaned, and statelessly extracted full text of 87,984 processed PDFs. This was generated by processing `v4.csv` to remove file corruption and missing data artifacts (Phase 2 output). - `chroma_db_ai_papers_mpnet.rar`: The compressed ChromaDB vector database containing 7,496,671 chunks and their corresponding 768-dimensional `all-mpnet-base-v2` embeddings. To use this database, download and extract the archive. The internal directory structure is as follows: ```text chroma_db_ai_papers_mpnet/ ├── chroma.sqlite3 # SQLite database containing metadata and document text └── 353b9c15-6644-4264-96ff-a99e49f860f1/ # hnswlib vector index directory ├── data_level0.bin ├── header.bin ├── index_metadata.pickle ├── length.bin └── link_lists.bin > **⚠️Note on ChromaDB Compatibility:** If you experience hardware or architecture loading errors with this pre-computed database (e.g., on Mac ARM/Apple Silicon), you can easily rebuild the index locally. The raw text is provided in `final_data.csv`, and the exact `vectorization.py` script required to rebuild it is available in the `data_pipeline/` folder of our [GitHub Repository](https://github.com/GitHub-AmanBhardwaj/Scholarly-Epistemic-Engine). ## Dataset Creation ### Curation Rationale The rapid growth of AI literature has created an acute information overload, making it difficult for researchers to track trends or synthesize ideas across decades. Existing datasets often assume text is pre-processed or lack the scale necessary for comprehensive semantic elicitation. This dataset provides a complete, chunked, and embedded corpus to facilitate immediate research into advanced academic RAG workflows. ### Source Data #### Data Collection and Processing 1. **Metadata Acquisition (Phase 1):** Iterative scraping of the arXiv API (cs.AI category) from January 1993 to December 2024, utilizing delay mechanisms and error handling to respect rate limits. 2. **Text Extraction (Phase 2):** Resumable, stateless PDF downloading and text extraction in a Kaggle computing environment. Excluded ~1.5% of papers due to file corruption/missing data. 3. **Chunking & Vectorization (Phase 3):** Text was cleaned, tokenized into sentences using NLTK `punkt`, and grouped into chunks of 5 sentences. Each chunk was embedded using `all-mpnet-base-v2` (768 dimensions) via a local GPU (NVIDIA RTX 5090) and stored in ChromaDB. #### Who are the source data producers? The source data producers are academic researchers, students, and industry professionals who submitted their pre-print manuscripts to the open-access arXiv repository under the Computer Science - Artificial Intelligence (cs.AI) category. ### Personal and Sensitive Information This dataset contains minimal personal information, limited to the names and institutional affiliations of the authors as publicly provided in their arXiv submissions. No sensitive personal information is deliberately included. ## Bias, Risks, and Limitations - **Domain Limitation:** The dataset is strictly confined to the `cs.AI` category. Cross-disciplinary queries (e.g., AI in bioinformatics) may suffer if the relevant papers were categorized elsewhere on arXiv. - **Extraction Artifacts:** Despite text cleaning, the raw extraction from PDFs may contain OCR artifacts, broken equations, or malformed tables. - **Temporal Cutoff:** The dataset contains papers up to December 2024 and does not reflect research published after this date. - **Embedding Bias:** The vector representations inherit any systemic biases present in the `all-mpnet-base-v2` model's training data. ### Recommendations Users should employ advanced retrieval mechanisms (such as the Cross-Encoder re-ranking and Adaptive LLM Fusion detailed in the accompanying paper) to mitigate the noise of broad retrieval and ensure semantic faithfulness when generating answers from this dataset. ## Dataset Card Authors Aman Bhardwaj, Jeet Bhardwaj, Mir Shahnawaz Ahmad ## Dataset Card Contact For questions regarding the dataset or methodology, please contact: Aman Bhardwaj (whyamanbhardwaj@gmail.com).