Alternative data datasets
Browse by provider, category, marketplace, access, delivery and license
An English-language NLP corpus of cleaned earnings call transcripts gathered from public investor-relations pages, sized between 10,000 and 100,000 documents for financial analysis and LLM work.
S&P 500 earnings call episodes from 2005 to 2025 are released as static tabular rows, each identified by a company-quarter episode identifier and containing full transcripts along with SEC press materials, suited to supervised or reinforcement-style experiments.
Indian stock market price history spanning 2000 to 2026 for over 2,500 NSE-listed stocks and indices, packaged in large Parquet shards around 1.5 GB each for efficient streaming.
Onchain Agent Builder Bench 367 multiple-choice questions on building onchain (crypto) AI agents: which stack layers and standards a build needs, what ERC-8004, x402, ERC-4337, ERC-7715, MCP, A2A and SATI are for, how to read what a package does with a key, 148 glossary terms, and what an assistant should decline. Release v1.0, dated 2026-10-02. Live version and the tools that answer these questio
Citation This dataset builds on edgar-corpus.
sec-filings is a Hugging Face dataset published by Sicheng-Chroma that distributes pre-chunked U.S. SEC filings paired with dense vector embeddings. It is distributed in optimized Parquet format and is intended for retrieval and similarity research over regulatory disclosures.
View source information and access options.
View source information and access options.
Original TIFF drawings and grant full-text XML for 165,917 U.S. design patents omitted from the AI4Patents/IMPACT collection, filling gaps including 161,093 grants from 2023 to 2026 and 4,824 earlier patents missing from IMPACT.
Earnings Call Transcript Lite is a small public-records dataset of corporate earnings call transcripts paired with short reference summaries. It is distributed as a CSV and is published on the Hugging Face Hub under an unstated license.
A community-compiled collection of stock-market-related tweets in English, stored as CSV, hundreds of thousands of posts (CC BY 4.0). Retail chatter feeds sentiment-signal research and market-tone studies.
tech-job-postings-labeled is a small public dataset of labeled technology job postings published on Hugging Face by SuhaibAtef. It comprises roughly 8,000 records tagged with source labels and is distributed in JSON.
Vector embeddings of Apple patent text totalling about 4 million entries, stored as parquet shards with a 768-dimensional float64 feature per record.
jobpostingsamples is a small sample dataset of US job postings published by talanAI. It is distributed in CSV format and is tagged for use with the Hugging Face datasets, pandas, polars, and mlcroissant libraries.
The us-patents dataset is a collection of ~ 8M US patent grants and applications from 1976-2025, cleaned, filtered, and formatted for pre-training of language models. Document Format corpus_id: Unique integer key with no semantic value. filing_date: The filing date of the grant or application. In case of duplicates, earliest filing date from the duplicate cluster. patent_type: The type of patent.
Text of SEC filings pulled from EDGAR spanning millions of records in English (Apache 2.0): 10-K, 10-Q, 8-K, annual reports and exhibits, tagged for text-generation and classification. A practical base corpus for earnings- and filings-driven signals, from risk factors and MD&A to management commentary.
View source information and access options.
PatentPulse PatentPulse is a provenance-preserving corpus of USPTO grants and published patent applications extracted from official weekly XML bulk releases. This immutable Parquet snapshot normalizes the project's historical append-only JSONL into one schema. Historical partial snapshot. This release was captured on 2026-08-28 while the upstream backfill was still in progress. It is a stable, cit
Indian Domestic Airline Flights 2018-2025 (TsFile) Apache TsFile version of Gokul99400/IndianDomesticAirlineDataset. Overview Indian domestic airline schedule records covering 2018-2025 across ~100 airports: airline, flight number, route, operating days of week, scheduled departure/arrival times and the schedule validity window (validFrom..validTo). The repo ships two CSVs: Air-Clean.csv (33,734 d
SEC 10-K Filings Embeddings Chunked text from 49 SEC 10-K annual reports for seven large US companies, with a 1536-dimensional dense embedding per chunk. Built for the Pinecone 101 workshop as a ready-to-import vector dataset. Records 15,244 chunks Filings 49 (10-K) Companies AAPL, AMZN, GOOGL, META, MSFT, NVDA, TSLA Fiscal years 2018–2025 (7 per company) Vector 1536-dim float32, intended for cosi
Qwen3-4B Patent Embeddings (专利特征向量数据集) 📌 简介 (Introduction) 本数据集包含了海量专利文本的特征向量（Embeddings）。所有特征向量均使用强大的 Qwen3-embedding-4B 模型提取。 数据领域: 专利 (Patents) 特征提取模型: Qwen3-embedding-4B 数据格式: .mmap (Memory-mapped file, numpy-compatible) + .tsv 数据精度: float16 📂 数据结构 (Data Structure) 由于数据量庞大，在生成时使用了 4 张 GPU 进行分布式计算，因此数据被划分为 4 个分片（Shards）：shard_0 到 shard_3。 每个 Shard 目录下包含以下核心文件： embeddings.float16.mmap: 核心的特征向量矩阵
View source information and access options.
A collection of Parquet files providing one record per Chinese invention patent family granted between 2003 and 2019, capturing a cumulative quality assessment pipeline that includes metrics such as semantic knowledge recombination, strict historical semantic novelty, and IPC-based robustness measures across global patent publications.
munich2drive_edgar Surround-camera driving clips from a real Munich HD map built by SimForge, each paired with the ego vehicle's full trajectory. Every clip is a nine-panel video -- six surround cameras at the ego's real sensor extrinsics, a chase view, a bird's-eye view, and the frame-locked CommonRoad ground truth -- so what the model sees and what the data says are visible side by side. The ego
Dataset details:- This dataset is the 2nd iteration following bugs in 1st dataset. The initial data suffered with followoing cases:- (i) The failed reference_answers generation(due error totalling 23) primarly because of 2 reasons/exceptions:- (a) There was normal limit(300 in 1st request) and worst case limit(450 in 3rd request) number of tokens for consolidated 4 refernce_answers per chunk and i
Dataset details:- This dataset is basically mapping of final anchor-positive pair data with their refernce answer. The given input data considered because:- (i) it had the had purest anchor-positive pairs with semantically bound anchors with context/positive. (ii) gave us the best result on final embedding fine tuning model. The anchor-context(positive)-reference_answer data has been generated via
A knowledge graph assembled from 575,778 ClinicalTrials.gov registrations and their associated arms, outcomes, sites, sponsors, conditions, interventions, MeSH codes, and PubMed citations, containing 7,628,735 nodes and 15,531,427 edges, delivered as 623 MB of Parquet files (compared to about 7 GB of raw JSON) and built using Samyama Graph.
View source information and access options.
View source information and access options.
A snapshot of processed ClinicalTrials.gov records filtered to completed-recruitment Phase 3 and Phase 4 studies, plus trials of medical devices or behavioral interventions, totaling 156,887 trials as of March 2025, used for an agentic medical retrieval and evidence-grounding system.
A 10% sample of over 573,000 ClinicalTrials.gov studies augmented with AI-classified therapeutic areas, sponsor categorization, outcome groupings, and duration metrics, offered for portfolio and market landscape analysis.
Cleaned, quality-flagged insider transactions and buying clusters, Jan 2024
10-K annual filings from EDGAR in Parquet, tens of thousands of documents. The standard annual-disclosure corpus for fundamentals, risk-factor and management-discussion analysis.
The linkedin-job-postings dataset, published by xanderios on Hugging Face, contains a single CSV table of U.S. LinkedIn job listings with about 33,000 rows. It is tagged with a MIT license on the platform, though the underlying LinkedIn terms of service place restrictions on redistribution and use.
Minute-level open-high-low-close-volume bars for Indian equities, hundreds of millions of rows in Parquet under MIT. Gives high-frequency microstructure across the Indian cash market for intraday and market-on-close strategies.
The restaurant_reviews dataset, published on the Hugging Face Hub by user yav1327, is a tabular and text collection of restaurant listings paired with user reviews. It is distributed in Parquet format with a single train split of 13,144 rows and an unstated license.
patent-spec-xml is a dataset of patent specification documents in XML format, published by Yehoon and licensed under unspecified terms. It is catalogued on the alternative-data directory as a public-records resource.
The 2024 Earnings Call Transcript dataset, published by yeong-hwan on Hugging Face, aggregates quarterly earnings call dialogue for individual stock tickers. It is distributed as a JSON table with 1,904 rows in a single train split.
CoER RL Project page · Paper · Code Executable task configurations for Stage 2 bilateral Co-PPO, with separate training and internal-validation splits. These are not generated rollouts or the official evaluation panel. Contents Split File Size train train.parquet 12,705 configurations validation validation.parquet 3,186 configurations Training and validation configurations are disjoint. Internal v
Over 635,000 registered clinical studies from ClinicalTrials.gov together with a supplementary pharma intelligence table of roughly 1.06 million rows that map those trials to FDA drug approvals, distributed as part of an open data initiative for academic and personal use.
The tech-job-postings-salaries-sample dataset is a 500-row public sample of a larger scraping-based feed of technology job postings drawn from public applicant tracking system endpoints such as Greenhouse, Lever, Ashby, Workable and SmartRecruiter. It is published by DataForge (Zalize) for evaluation and preview use.
The Tech Job Postings & Salary Dataset from Zalize Data contains over 449,000 monthly job postings scraped from public applicant tracking systems at US and EU technology companies. It is distributed as parquet files under a CC BY-NC 4.0 license and is part of the DataForge Open Data program.
Derived from USPTO and PatentsView bulk releases, this dataset catalogs 9.1 million U.S. patents from 1976 onward along with complete citation links, assignees, inventors, and CPC codes, with the full-graph edition exceeding 255 million rows.
Normalized stock-trade disclosures filed by U.S. House and Senate members under the STOCK Act, regenerated on a schedule by the repository's own pipeline rather than curated manually.
ZipLime US Insider Trading Disclosures (PIT) [!CAUTION] Use knowledge_date, not transaction_date, when backtesting. transaction_date says when a trade occurred; knowledge_date says when the filing became observable through EDGAR. Using the former as the signal date introduces look-ahead bias. visible = trades.filter(pl.col("knowledge_date") <= simulation_time) This dataset normalizes corporate-ins
US Short Interest This dataset publishes no short-interest data, and that is the finding, not a bug. Every US source of equity short positioning was audited on 13 September 2026 before any of it was collected. Two forbid the publication pattern and the third does not exist yet. What is published here is the audit itself — which clause, on which page, read on which day — plus the adapters that will
A curated collection of SEC EDGAR filings from recent and upcoming IPO companies, containing 5,179 text segments sourced from 261 filings, formatted for large language model training and financial analysis.