AI Training & RAG Datasets

What you can license

Four licence types, each scoped to a different kind of AI work. Training rights are a separate licence from an API or MCP subscription: a subscription covers querying the data and working with the answer, and does not permit training on it or persisting it into a durable AI knowledge base.

• Model Training (LLM and ML) · internal training of machine-learning or large language models on CSIMarket datasets, including embeddings, classifiers, RAG systems, supervised learning and financial-reasoning development
• Derivative Model · models whose weights carry learned representations derived from the datasets. Redistribution of the raw data stays prohibited; redistribution of a derivative model is on negotiated terms
• RAG and Vector Index · for vector-search systems and for using the data as a retrieval corpus: embeddings, chunking, indexing and internal RAG workloads
• Enterprise Unlimited · unlimited internal use across departments, data-science teams and AI product groups, with optional multi-year terms

Which datasets are included

All of them. Every dataset CSIMarket sells is available under an AI licence, together or on its own.

• Industry Benchmarking · 110 industries and 13 sectors, 290 fields per period, 20+ years, true aggregates with dispersion statistics
• Company Fundamentals · statements plus the six analytics for roughly 13,000 US-listed companies, with composite scores and z-scores
• Point-in-Time Fundamentals · 50,307,195 observations across 17,020 companies, periods 1991 to 2027, with revision vintages and knowledge time preserved
• Quantitative Factor and Risk · 90 fields across 10 families, 13 years, point-in-time and survivorship-safe
• Supply-Chain and Relationship Intelligence · the commercial graph across twelve dimensions, every edge carrying its provenance, confidence and filing evidence
• Institutional Ownership · 13F ownership with concentration measures, manager portfolios, and insider transactions from Forms 3, 4 and 5
• Business Segments and Geographic Exposure · reported segment and geographic revenue, with the denominator published and disclosure completeness reported rather than padded
• Share Prices and Market Performance · daily OHLCV from 2006, index benchmarks, breadth, regime and rotation
• Total Market · the whole US-listed universe aggregated as a single line item, TTM, FY and Q
• News and Corporate Events · a normalized feed with story clustering and an event-typed stream
• Screening and Bulk · flat bulk statement rows built for pipeline ingestion
• Classification cross-walks · the 110-industry and 13-sector taxonomy mapped to 156 NAICS and 410 SIC codes

Why it trains well

• One set of definitions. Every company and every industry is measured the same way, which is what makes the data usable as a supervision signal. A model learns more reliably from “this company’s return on assets against its industry median” than from raw, inconsistently tagged filings.
• Point-in-time where it matters. The factor panel is survivorship-safe, and the point-in-time dataset keeps what the filings actually said on a past date, as first reported and as later revised, with the knowledge time behind each observation. A model trained on restated figures learns the future.
• Evidence, not assertions. Relationship edges carry their provenance, confidence and the underlying filing, so a retrieval corpus can cite rather than claim.
• Gaps are reported, not filled. Segment disclosure completeness is published instead of padded, so absent data stays absent rather than becoming a learned fiction.

How it behaves: stated, not implied

• An ordinary subscription does not grant training rights. Under the Terms of Use, use for AI or ML training, embeddings, vector databases, RAG or derivative models is not permitted unless expressly licensed. That applies to data received through the API and through MCP alike.
• MCP is not a corpus route. Asking a question over MCP and working with the answer in that session is what MCP access is for. Persisting the results into a durable knowledge base is a licensed activity, and bulk files are the route built for it.
• Raw redistribution stays prohibited. Every licence type covers internal use. A derivative model may be redistributable on negotiated terms; the underlying data is not.
• No ESG or climate data. CSIMarket holds none, so no AI licence includes any.

Delivery

Built for a pipeline rather than a browser.

• CSV · long form and wide form
• Parquet · optimized for ML pipelines
• JSON and NDJSON · for streaming ingestion
• API streaming · for automated workloads, with refresh options to match the training cycle

Best real-life usage

• Train a financial-reasoning model on normalized fundamentals and their industry context, rather than on scraped filings.
• Build a retrieval corpus an analyst assistant can cite: statements, benchmarks, relationships and events, each with its source.
• Generate embeddings for semantic search across companies, industries and business descriptions.
• Fit a derivative scoring model on the factor panel without leaking the future into the training window.