KRUMMHOLZ AI - PRELIMINARY CLEANED DATASETS Release: 2026-10-22 Website: https://krummholz.ai/datasets/ Five source collections currently available in four categories. Public catalog availability updated October 26, 2026; the original release date and remaining package contents are unchanged. Source datasets were retrieved, reduced to task-relevant fields, privacy-screened and checked using the methods described in their package documentation. This is a preliminary research release; it is not certified anonymous, independently domain-validated or ready for production deployment. CONTENTS Maine FIA forest measurements - PRELIMINARY 903,254 unique feature rows Maine forest inventory measurements for exploratory models of tree dimensions, biomass, carbon and forest condition. The two tree source parts are combined before deduplication. All available years are retained. License: USDA public-use data - CC0 identified in USDA’s dataset record Limit: Represents 1,095,255 tree and condition observations, with repeated feature rows stored as record_count. Download both parts. Multi-year sample; precise locations and record linkages removed. Not a current tree census or a basis for statewide estimates. NOAA MRIP 2024 fishing observations - PRELIMINARY 295,104 unique feature rows NOAA Marine Recreational Information Program 2024 recreational fishing observations across all six available waves. Catch, trip activity and measured fish size support exploratory tabular ML. These are recreational survey data, not commercial fishing or vessel-control training data. License: NOAA U.S. government public-use data Limit: Represents 392,082 retained catch, trip and size observations. Repeated feature rows use record_count. Broad state/area fields remain; fine locations and survey linkage keys are removed. Does not reproduce NOAA population estimates. BANKING77 banking customer-service intents - PRELIMINARY 13,006 queries · 77 intents Banking customer queries across all 77 intents, including card support, payments, transfers, cash withdrawals, fees, identity verification and refunds. License: CC-BY-4.0 Limit: Automated privacy screening changes some wording. Queries and intent labels only; source evaluation splits are retained but modified. No guarantee of complete anonymization or removal of semantic paraphrases. CLINC150 shopping and payment intents - PRELIMINARY 447 queries · 3 intents Crowdsourced requests for order placement, order status and declined cards. The other intents and out-of-scope examples are excluded. License: CC-BY-3.0 Limit: Automated privacy screening changes some wording. Queries and intent labels only; source evaluation splits are retained but modified. No guarantee of complete anonymization or removal of semantic paraphrases. Jaco Play - numeric trajectories - PRELIMINARY 1,085 episodes · 77,965 frames 1,085 complete tabletop-robot episodes with 77,965 ordered numeric frames and 89 task instructions, extracted from the pinned LeRobot conversion. Images and video excluded. License: CC BY 4.0 Limit: Full numeric inventory of the pinned LeRobot Jaco conversion. Images and video excluded. Coordinate units are not independently verified; no independent evaluation split. Offline research data, not validated real-robot control commands. CLEANING AND USE Customer-service texts: the currently available packages retain all 77 banking intents and the agreed CLINC service subset. Historical cleaning documentation describes the original cross-source deduplication process, including a third collection no longer distributed here. Detect/redact contact, identifier and named-entity patterns; deduplicate normalized text across sources and splits, favoring evaluation examples. Exclude conflicting labels and unsupported generated answers. Review exact counts and limitations in each cleaning audit. Environmental measurements: retain relevant source measurements and species mappings, drop identifiers, fine locations, administrative tables and all-empty features. Compact identical feature vectors into one row with record_count to preserve observed multiplicity. Counts across different tables are not independent trees, fish or people. Robotics: include the full numeric inventory of the pinned Jaco LeRobot conversion, with ordered episodes and task instructions. Exclude all imagery/video and storage bookkeeping. Check complete trajectories for duplicates; preserve repeated states within a trajectory. HOW TO OPEN Each ZIP is independently extractable. Forestry requires both ZIP parts for the complete release; concatenate tree CSV data rows under one header. Other tables remain separate. CSV and JSONL in the support packages are alternative formats of the same records. Review-sample files are subsets, not additional training examples. The website serves the cleaned ZIP files with their release documentation. Raw acquisition files and rejected personal text are not included. DOCUMENTATION CATALOG.json lists download paths, package sizes and SHA-256 hashes. Every collection includes a README, source attribution, license information and cleaning audit. Each dataset retains its source terms; there is no blanket MIT license. INDEPENDENT PRIVACY REVISION - 2026-10-23 The BANKING77 preliminary package was rescreened. Five residual fine-location spans in customer incident narratives were replaced with LOCATION placeholders. Record identifiers and all alternate/review copies were updated. The 13,006 records, 77 intents and source split counts are unchanged. Updated ZIP size and SHA-256 appear in CATALOG.json. No raw-to-clean location mapping is included. Other packages in this dated preliminary bundle are unchanged. The separate legacy homepage forestry and banking samples were also replaced with privacy-reduced editions. Forestry retains 10,000 scientific observations while removing plot/source-row linkage and raw source copies. The legacy banking sample now uses the 13,006-record cleaned corpus and omits source-row mappings. These changes are documented inside those separate ZIPs. This review combines complete schema and identifier-pattern scans, named-entity and capitalization candidate review of natural-language data, and manual assessment of flagged values. Public citation authors, generic service terminology, broad scientific geography and numeric measurements are not treated as confidential identities. Screening and data minimization reduce risk; they do not certify anonymity or preclude linkage against public source datasets.