prabakaranchandran/enterprise-dynamics-corpus-v1

Enterprise Dynamics Corpus (EDC v1)

Real-data pretraining corpus for enterprise forecasting: commerce and sales, demand and growth, finance and economic activity, IT/cloud operations, infrastructure and industrial asset telemetry. Energy, climate and transport are capped extras, not the core.

Status (2026-09-28): The default commercial_core contains 3.229 GB of verified real training Parquet, 777,504 series records across 18 sources. 46.771 GB remains toward the 50 GB commercial-core target. The full archive includes held sources: 6.273 GB across 24 verified batches. Admission, quality and category limits remain enforced. This is an incomplete corpus; availability is not yet a balanced training mixture. Machine-readable coverage and shortfalls for the commercial selection are in its coverage report. Archive coverage includes held sources. The 75/25 real/synthetic sampler is planned, not implemented. See the consumption guide and per-source contracts for channels, clocks, covariates and history lengths.

The default commercial_core includes only sources screened for commercial use and paid/free redistribution, subject to source attribution, notices and exclusions. The all config is an archive and includes explicitly excluded or review-held sources. No blanket license replaces the original terms, and this screening is not a legal warranty or a separate model-weight license.

Combined real and synthetic pretraining

Select enterprise_pretrain, split train, to load the commercial-core real records together with the enterprise pilot's 900 synthetic training scenarios. The 50 synthetic validation and 50 synthetic test scenarios are excluded. Select enterprise_synthetic to access its three separate splits. Both configurations retain canonical EDC rows.

This is an opt-in experimental training configuration; the real-only default is unchanged. Five generator families cover seasonal and intermittent demand, regime shifts, stable multivariate dynamics, and cointegration. Synthetic clocks use step indices only. The larger Mixed11 pool remains review-held and is not included. No generalization benefit has yet been measured. The planned 75/25 real/synthetic sampling ratio is not enforced by a file union; implement it in the training sampler. See the exact paths and counts and pilot documentation.

How the corpus is constructed

1. Admission. A source enters only through curation/sources.json, where it is declared once: producer, licence, attribution, upstream Hub pin, cadence and channels, chronological split boundaries and a size cap. Only producer terms that permit commercial use and redistribution are admitted to commercial_core; research-only, non-commercial, no-data-resale and unresolved terms are excluded from that selection. Historical held copies remain in the clearly labelled archive. Every source also needs a rule in the benchmark exclusion registry; configs whose fields differ from the registry are rejected.
2. Planning. curation/plan.py lists the pinned upstream files (with LFS SHA256), drops files already ingested, orders the rest by a seeded permutation and takes shards until the source cap or its category cap is reached. Batches are queued round-robin so no single source runs ahead.
3. Conversion. Each pinned file is downloaded on the curation VM, hash-checked and converted to one canonical Parquet schema: channel-first FP32 target and past-only covariates, original clock and cadence, NaN preserved, no scaling, imputation, resampling or window generation.
4. Training only. Each source contributes every observation before its family's contamination cutoff (from the benchmark exclusion registry). No validation or test sets are built here; evaluation uses separately audited benchmark tasks or downstream recipes. Seen-source temporal evaluation must be reported separately from source-unseen zero-shot evaluation.
5. Audits. Raw hash, pinned counts, source clock bounds, split disjointness, channel alignment, null/infinity checks, entity-interval overlap across batches, exact-content duplicate quarantine and a Parquet round trip with semantic digests must all pass before publication.
6. Publication. One atomic commit per batch adds partitions and a receipt; each file's server-side SHA256 is checked at that commit, then a verification record and the duplicate registry checkpoint are committed and read back. Batch-0001 was independently rebuilt on a second machine and matched the published Parquet byte for byte.

The following ledger and quality tables cover the full archive, including held sources.

Corpus ledger

Generated from verified batch receipts; training partition only unless stated. FP32 bytes = scalar slots x 4 (NaN slots included), excluding IDs and timestamps.

• Verified real training Parquet: 6.273 GB (12.5% of the 50.000 GB target); 43.727 GB remaining.
• Uncompressed training values: 22.088 GB FP32 slots (20.520 GB finite values); held-out partitions 0.000 GB FP32. Raw downloads, staging files, metadata, quality tables and the separate synthetic pool do not count toward the Parquet target.
• Training series: 1,014,439; timesteps: 2,014,197,889; scalar slots: 5,521,945,951.
• Diversity: 11 categories, 21 sources, 9 sampling intervals; largest source share 22.4%; effective number of sources 8.91 (inverse Simpson).
• Verified batches: 24; pending: none.

By category

Category	Sources	Intervals	Licences	Train series	Train timesteps	Train Parquet GB	Parquet share	Train FP32 GB	Value share
Digital and customer activity	4	1 d, 1 h, 10 min	CC-BY-4.0, CC0-1.0	144,187	915,942,738	2.330	37.1%	6.584	29.8%
IT and cloud operations	3	1 min, 5 min	CC-BY-4.0	78,193	662,693,901	2.834	45.2%	6.249	28.3%
Industrial assets and IoT telemetry	2	1 d, 1 s	CC-BY-4.0, LicenseRef-Backblaze-Drive-Stats	177,658	137,800,769	0.394	6.3%	6.240	28.3%
Geopolitics and events	2	1 d	GDELT-Terms-of-Use	15,544	100,952,029	0.357	5.7%	1.615	7.3%
Demand and sales	2	1 d, 1 h	CC-BY-4.0	53,324	110,453,112	0.095	1.5%	0.884	4.0%
Trade, supply chain and inventory	2	1 mo	LicenseRef-Eurostat-reuse-policy, LicenseRef-Statistics-Canada-Open-Licence	434,300	36,767,022	0.134	2.1%	0.205	0.9%
Prices	2	1 mo, 7 d	Etalab-Licence-Ouverte, US-Government-Public-Domain	80,967	25,534,157	0.075	1.2%	0.197	0.9%
Infrastructure and energy	1	1 h	CC-BY-4.0	1,955	17,172,720	0.034	0.5%	0.069	0.3%
Production and industry	1	1 mo	LicenseRef-Eurostat-reuse-policy	22,906	6,114,931	0.011	0.2%	0.024	0.1%
Labour and business dynamics	1	3 mo	CC0-1.0	4,813	506,734	0.005	0.1%	0.012	0.1%
Markets and market signals	1	7 d	US-Government-Public-Domain	592	259,776	0.004	0.1%	0.008	0.0%

By source

Source	Category	Licence	Interval	Channels (target+past)	Span	Train series	Train timesteps	Train Parquet GB	Parquet share	Train FP32 GB	Value share
backblaze_drive_stats	Industrial assets and IoT telemetry	LicenseRef-Backblaze-Drive-Stats	1 d	10+0	2022-01-01 to 2024-01-01	177,564	123,797,805	0.297	4.7%	4.952	22.4%
quito_hour	Digital and customer activity	CC-BY-4.0	1 h	5+0	2021-11-18 to 2023-07-28	8,531	126,292,924	1.184	18.9%	2.526	11.4%
azure_vm_traces_2017	IT and cloud operations	CC-BY-4.0	5 min	1+2	2016-11-15 to 2016-12-13	36,244	199,973,657	2.219	35.4%	2.400	10.9%
borg_cluster_data_2011	IT and cloud operations	CC-BY-4.0	5 min	2+5	2011-05-01 to 2011-05-28	23,127	83,268,724	0.528	8.4%	2.332	10.6%
anaconda_conda_downloads_hourly	Digital and customer activity	CC-BY-4.0	1 h	1+0	2017-01-01 to 2023-01-01	13,977	459,357,846	0.280	4.5%	1.837	8.3%
azure_functions_2019	IT and cloud operations	CC-BY-4.0	1 min	1+0	2019-07-15 to 2019-07-29	18,822	379,451,520	0.087	1.4%	1.518	6.9%
gdelt_v1_events_backfile	Geopolitics and events	GDELT-Terms-of-Use	1 d	4+0	1979-01-01 to 2013-04-01	7,266	88,879,275	0.283	4.5%	1.422	6.4%
petrobras_3w_real	Industrial assets and IoT telemetry	CC-BY-4.0	1 s	14+9	2011-08-30 to 2023-09-26	94	14,002,964	0.097	1.5%	1.288	5.8%
quito_min	Digital and customer activity	CC-BY-4.0	10 min	5+0	2023-07-10 to 2023-07-28	21,679	56,191,968	0.499	8.0%	1.124	5.1%
wiki_daily_100k	Digital and customer activity	CC0-1.0	1 d	1+0	2015-07-01 to 2023-01-01	100,000	274,100,000	0.367	5.8%	1.096	5.0%
freshretailnet_50k_hourly	Demand and sales	CC-BY-4.0	1 h	1+1	2024-03-28 to 2024-06-26	50,000	108,000,000	0.091	1.5%	0.864	3.9%
gdelt_v1_events_daily	Geopolitics and events	GDELT-Terms-of-Use	1 d	4+0	2019-01-01 to 2023-01-01	8,278	12,072,754	0.074	1.2%	0.193	0.9%
fr_fuel_prices_station_weekly	Prices	Etalab-Licence-Ouverte	7 d	2+0	2007-01-01 to 2024-12-30	74,219	23,769,273	0.070	1.1%	0.190	0.9%
eurostat_comext_hs4	Trade, supply chain and inventory	LicenseRef-Eurostat-reuse-policy	1 mo	2+0	2002-01-01 to 2025-01-01	53,805	14,572,809	0.083	1.3%	0.117	0.5%
statcan_cimt_hs6_2017	Trade, supply chain and inventory	LicenseRef-Statistics-Canada-Open-Licence	1 mo	1+0	2017-01-01 to 2022-01-01	380,495	22,194,213	0.051	0.8%	0.089	0.4%
bdg2_building_meters	Infrastructure and energy	CC-BY-4.0	1 h	1+0	2016-01-01 to 2017-01-01	1,955	17,172,720	0.034	0.5%	0.069	0.3%
eurostat_sts_industry_m	Production and industry	LicenseRef-Eurostat-reuse-policy	1 mo	1+0	1962-01-01 to 2024-01-01	22,906	6,114,931	0.011	0.2%	0.024	0.1%
uci_online_retail_ii	Demand and sales	CC-BY-4.0	1 d	2+0	2009-12-01 to 2011-12-09	3,324	2,453,112	0.003	0.1%	0.020	0.1%
census_qwi_state_sector_firmsize	Labour and business dynamics	CC0-1.0	3 mo	6+0	1990-01-01 to 2024-01-01	4,813	506,734	0.005	0.1%	0.012	0.1%
cftc_cot_legacy_futures_weekly	Markets and market signals	US-Government-Public-Domain	7 d	8+0	1992-10-06 to 2024-12-31	592	259,776	0.004	0.1%	0.008	0.0%
eia_energy_prices_monthly	Prices	US-Government-Public-Domain	1 mo	1+0	1973-10-01 to 2025-01-01	6,748	1,764,884	0.005	0.1%	0.007	0.0%

Attribution

• anaconda_conda_downloads_hourly (CC-BY-4.0; licence; producer: https://github.com/ContinuumIO/anaconda-package-data). Anaconda, Inc. Conda package download data, https://github.com/ContinuumIO/anaconda-package-data (CC BY 4.0). EDC changes: aggregate hourly download counts across package variants, mask documented missing collection periods, retain active package series, FP32. Historical .conda-format coverage change from June 2022 is preserved. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at 24129b3ae70b.
• azure_functions_2019 (CC-BY-4.0; licence; producer: https://github.com/Azure/AzurePublicDataset/blob/207bed67dd10090b28ad4f745b2cfd41a11aace4/AzureFunctionsDataset2019.md). Azure Functions Trace 2019 by Microsoft Azure and Microsoft Research, licensed CC-BY-4.0. Cite Shahrad et al., Serverless in the Wild: Characterizing and Optimizing the Serverless Workload at a Large Cloud Provider (USENIX ATC 2020). EDC changes: per-function and per-app per-minute invocation series on one 14-day 1-minute grid (unlisted days NaN), activity filter, FP32 canonical Parquet; counts unchanged. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at 96c6b7f095e8.
• azure_vm_traces_2017 (CC-BY-4.0; licence; producer: https://github.com/Azure/AzurePublicDataset/blob/master/AzurePublicDatasetV1.md). Original Azure VM V1 trace by Microsoft Azure and Microsoft Research, licensed CC-BY-4.0. Cite Resource Central: Understanding and Predicting Workloads for Improved Resource Management in Large Cloud Platforms (SOSP 2017). Acquired through Salesforce CloudOps/LOTSA/GIFT preprocessing. EDC changes: selection of the identified source shard, FP32 canonical Parquet representation, and chronological partitioning; numerical values and channel order retained. Acquired from Salesforce/GiftEvalPretrain at 6830b624de7e.
• backblaze_drive_stats (LicenseRef-Backblaze-Drive-Stats; licence; producer: https://www.backblaze.com/cloud-storage/resources/hard-drive-test-data). Backblaze Drive Stats (Hard Drive Data and Stats), Backblaze, Inc., https://www.backblaze.com/cloud-storage/resources/hard-drive-test-data , quarterly data Q1 2022 - Q4 2023, accessed 2026-09-27. EDC changes: select 2022-2023 only; daily SMART channels, absent readings NaN, no failure labels. Backblaze data itself must remain free and cannot be sold; users are responsible for use. Some SMART fields are vendor-packed; FP32 can round large integers. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at d8a94e96e754.
• bdg2_building_meters (CC-BY-4.0; licence; producer: https://doi.org/10.5281/zenodo.3887306). Miller, C., Kathirgamanathan, A., Picchetti, B., Arjunan, P., Park, J.Y., Nagy, Z., Raftery, P., Hobson, B.W., Shi, Z., Meggers, F. The Building Data Genome Project 2, energy meter data from the ASHRAE Great Energy Predictor III competition. Sci Data 7, 368 (2020). https://doi.org/10.1038/s41597-020-00712-x ; data: https://doi.org/10.5281/zenodo.3887306 EDC changes: raw meter readings for 2016 only, exclude Bull/Cockatoo/Hog entirely, empty cells NaN, FP32, one meter per series. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at e7ff7327a72e.
• borg_cluster_data_2011 (CC-BY-4.0; licence; producer: https://github.com/google/cluster-data/blob/master/ClusterData2011_2.md). Original trace by Google, Google Cluster Data project (Borg 2011), licensed CC-BY-4.0. Acquired through Salesforce CloudOps/LOTSA/GIFT preprocessing. EDC changes: selection of the identified source shard, FP32 canonical Parquet representation, and chronological partitioning; numerical values and channel order retained. Acquired from Salesforce/GiftEvalPretrain at 6830b624de7e.
• census_qwi_state_sector_firmsize (CC0-1.0; licence; producer: https://lehd.ces.census.gov/data/qwi/R2026Q3/). U.S. Census Bureau, "qwi__rh_fs_gs_ns_op_u," Quarterly Workforce Indicators, R2026Q3, https://lehd.ces.census.gov/data/qwi , accessed on September 26, 2026. CC0 per the Census API catalogue. EDC changes: selection of all-worker state x sector x firm-size cells and six indicators; status flags -2/-1/5 as NaN; quarters before 2024-01-01; FP32; values otherwise unchanged. Not endorsed or certified by the Census Bureau. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at d62d869af1b5.
• cftc_cot_legacy_futures_weekly (US-Government-Public-Domain; licence; producer: https://www.cftc.gov/MarketReports/CommitmentsofTraders/HistoricalCompressed/index.htm). U.S. Commodity Futures Trading Commission, Commitments of Traders (Legacy, Futures Only). Public domain; acknowledgement requested. EDC changes: one weekly series per contract market, position columns only; FP32. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at 2b6451043d83.
• eia_energy_prices_monthly (US-Government-Public-Domain; licence; producer: https://www.eia.gov/opendata/bulk/). Source: U.S. Energy Information Administration, Open Data bulk files (PET, NG, ELEC). Public domain; acknowledgement requested. EDC changes: monthly price series only, third-party-licensed series excluded; FP32. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at 4620ce9cfdd3.
• eurostat_comext_hs4 (LicenseRef-Eurostat-reuse-policy; licence; producer: https://ec.europa.eu/eurostat/api/dissemination/files/?dir=comext/COMEXT_DATA/PRODUCTS). Source: Eurostat, Comext international trade in goods. Modified by EDC (CN8->HS4 sums per EU27 reporter and flow over all partners; chapter-level confidential and chapter 98/99 codes dropped; missing months left NaN; sparse series removed; FP32). Eurostat has no responsibility for these modifications. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at 2b6451043d83.
• eurostat_sts_industry_m (LicenseRef-Eurostat-reuse-policy; licence; producer: https://ec.europa.eu/eurostat/web/short-term-business-statistics/database). Source: Eurostat, short-term business statistics (industry and construction). Modified by EDC: one index base and one adjustment per series selected, growth-rate units and country aggregates dropped, only EU/EFTA/candidate reporters kept, months from 2024 dropped, sparse series removed, FP32. Eurostat has no responsibility for these modifications or for the resulting data. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at 24129b3ae70b.
• fr_fuel_prices_station_weekly (Etalab-Licence-Ouverte; licence; producer: https://www.prix-carburants.gouv.fr/rubrique/opendata/). Ministeres economiques et financiers, Prix des carburants en France (prix-carburants.gouv.fr), Licence Ouverte / Open Licence (Etalab). EDC changes: posted-price events aggregated to Monday-start weeks (last price, posting count), out-of-range keying errors removed, no carry-forward; FP32. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at 27b7bb2956b8.
• freshretailnet_50k_hourly (CC-BY-4.0; licence; producer: https://huggingface.co/datasets/Dingdong-Inc/FreshRetailNet-50K). Dingdong Inc. FreshRetailNet-50K v1.0 (2025), https://huggingface.co/datasets/Dingdong-Inc/FreshRetailNet-50K, CC BY 4.0; Wang et al., FreshRetailNet-50K: A Stockout-Annotated Censored Demand Dataset for Latent Demand Recovery and Forecasting in Fresh Retail, arXiv:2505.16319. EDC changes: reshape producer daily rows to hourly store-product records; retain sales and stockout flag, drop daily features, convert values to FP32. No demand recovery or imputation. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at 2b6451043d83.
• gdelt_v1_events_backfile (GDELT-Terms-of-Use; licence; producer: https://www.gdeltproject.org/data.html#rawdatafiles). The GDELT Project, GDELT 1.0 Event Database, https://www.gdeltproject.org/ ; Leetaru, K. and Schrodt, P. (2013), GDELT: Global Data on Events, Location and Tone, 1979-2012, ISA Annual Convention. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at 1ad2bbbb62c5.
• gdelt_v1_events_daily (GDELT-Terms-of-Use; licence; producer: https://www.gdeltproject.org/data.html#rawdatafiles). The GDELT Project, GDELT 1.0 Event Database, https://www.gdeltproject.org/ ; Leetaru, K. and Schrodt, P. (2013), GDELT: Global Data on Events, Location and Tone, 1979-2012, ISA Annual Convention. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at 1ad2bbbb62c5.
• petrobras_3w_real (CC-BY-4.0; licence; producer: https://github.com/petrobras/3W/tree/4eaf746b2085ad76cade955545c33c67afe88bdf/dataset). Vargas, R.E.V., Munaro, C.J., Ciarelli, P.M., Medeiros, A.G., do Amaral, B.G., Barrionuevo, D.C., de Araujo, J.C.D., Ribeiro, J.L., Magalhaes, L.P. A realistic and public dataset with rare undesirable real events in oil wells. Journal of Petroleum Science and Engineering 181 (2019) 106223. https://doi.org/10.1016/j.petrol.2019.106223 ; 3W Dataset 2.0.0 data article: https://doi.org/10.1038/s41597-026-07225-z EDC changes: 32 training wells only; real measurements, overlapping timestamps merged only when equal; long gaps split, short gaps masked; labels and all-empty columns removed; FP32. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at e7ff7327a72e.
• quito_hour (CC-BY-4.0; licence; producer: https://github.com/alipay/quito). Quito corpus by Ant Group / Alipay (QuitoBench authors, arXiv 2603.26017), licensed CC BY 4.0: anonymised application-traffic telemetry across finance, e-commerce, infrastructure and IoT verticals. EDC changes: long-to-series reshaping on the exact source grid, FP32 canonical Parquet, and truncation at the QuitoBench test cutoff (2023-07-28); values, channel order and the upstream 1e-05 floor retained. Acquired from hq-bench/quito-corpus at f2d66ec7a9a8.
• quito_min (CC-BY-4.0; licence; producer: https://github.com/alipay/quito). Quito corpus by Ant Group / Alipay (QuitoBench authors, arXiv 2603.26017), licensed CC BY 4.0: anonymised application-traffic telemetry across finance, e-commerce, infrastructure and IoT verticals. EDC changes: long-to-series reshaping on the exact source grid, FP32 canonical Parquet, and truncation at the QuitoBench test cutoff (2023-07-28); values, channel order and the upstream 1e-05 floor retained. Acquired from hq-bench/quito-corpus at f2d66ec7a9a8.
• statcan_cimt_hs6_2017 (LicenseRef-Statistics-Canada-Open-Licence; licence; producer: https://www150.statcan.gc.ca/n1/pub/71-607-x/71-607-x2021004-eng.htm). Adapted from Statistics Canada, Canadian International Merchandise Trade, 2017–2021. This does not constitute an endorsement by Statistics Canada of this product. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at 7ec6ac6e9c73.
• uci_online_retail_ii (CC-BY-4.0; licence; producer: https://archive.ics.uci.edu/dataset/502/online+retail+ii). Chen, D. (2012). Online Retail II [Dataset]. UCI Machine Learning Repository. https://doi.org/10.24432/C5CG6D. Licensed CC BY 4.0. EDC changes: daily aggregation of invoice lines to net quantity and GBP revenue per SKU, customer country and store; non-product codes and zero-price stock write-offs excluded; closure days NaN; FP32 canonical Parquet; training-only. Acquired from prabakaranchandran/enterprise-dynamics-corpus-v1 at ba5626dcac3e.
• wiki_daily_100k (CC0-1.0; licence; producer: https://dumps.wikimedia.org/other/pageviews/readme.html). Wikimedia Analytics pageview counts, dedicated to CC0, distributed as the wiki_daily_100k pretraining subset by the Chronos/AutoGluon team. EDC changes: page-name identity, float64-to-float32 numerical conversion and canonical Parquet chronological partitions; no article text is admitted. Acquired from autogluon/chronos_datasets at eeecad0b82a8.

Quality by source

Medians over retained records (spikes: 90th percentile). Forecastability is 1 minus normalised spectral entropy (0 is white-noise-like); lag-1 autocorrelation near 0 also indicates noise. Rejections are hard rejects only (too short, mostly missing, constant targets, duplicates). Noise, intermittency and spikes alone are not rejection criteria. Version 1 joined observations across gaps for flat runs and lag correlation; those metrics require reprofiling before weighting or comparison with version 2. Flat-run fraction is the maximum across target channels, not evidence that every channel is constant.

Source	Records	Rejected	Quality version	Finite	Zeros	Longest flat run	Spikes p90	Lag-1 autocorr	Forecastability
anaconda_conda_downloads_hourly	13,977	0	1	1.000	0.623	0.004	0.044	0.303	0.100
azure_functions_2019	19,790	968	1	1.000	0.699	0.002	0.004	-0.035	0.401
azure_vm_traces_2017	36,244	0	1	1.000	0.000	0.000	0.121	0.707	0.199
backblaze_drive_stats	179,260	1,696	1	0.799	0.500	0.995	0.003	0.966	0.729
bdg2_building_meters	2,025	70	1	1.000	0.007	0.005	0.077	0.935	0.520
borg_cluster_data_2011	23,127	0	1	0.999	0.000	0.003	0.112	0.730	0.268
census_qwi_state_sector_firmsize	5,035	222	1	0.992	0.000	0.018	0.036	0.537	0.405
cftc_cot_legacy_futures_weekly	830	238	1	0.963	0.081	0.124	0.243	0.888	0.491
eia_energy_prices_monthly	8,906	2,158	1	0.998	0.000	0.007	0.018	0.979	0.669
eurostat_comext_hs4	53,805	0	1	1.000	0.000	0.004	0.082	0.604	0.269
eurostat_sts_industry_m	23,453	547	1	1.000	0.000	0.010	0.003	0.983	0.657
fr_fuel_prices_station_weekly	74,427	208	1	0.839	0.161	0.049	0.000	0.547	0.142
freshretailnet_50k_hourly	50,000	0	1	1.000	0.810	0.019	0.002	0.192	0.102
gdelt_v1_events_backfile	7,270	4	2	0.601	0.407	0.024	0.128	0.337	0.129
gdelt_v1_events_daily	8,282	4	2	0.819	0.190	0.010	0.067	0.290	0.090
petrobras_3w_real	155	61	1	0.678	0.129	1.000	0.319	1.000	0.762
quito_hour	8,531	0	1	1.000	0.000	0.250	0.263	0.690	0.409
quito_min	21,779	100	1	1.000	0.000	0.992	0.160	0.672	0.378
statcan_cimt_hs6_2017	381,402	907	1	0.900	0.000	0.017	0.136	0.105	0.126
uci_online_retail_ii	3,324	0	1	0.817	0.806	0.136	0.025	0.065
wiki_daily_100k	100,000	0	1	1.000	0.000	0.001	0.027	0.530	0.145

Source licenses

This collection does not replace the upstream licenses. Each source retains its own terms and attribution obligations, recorded in the batch receipts and manifest. The default commercial_core excludes Backblaze (no sale of the data itself) and the GIFT-derived Azure VM/Borg exports (conflicting packager wording, pending clarification or rebuilding from the CC BY originals). These are conservative release decisions, not findings that previous free hosting was unlawful. See license review and allowed paths.

• azure_vm_traces_2017: CC-BY-4.0. hold_packager_terms_ambiguity. Original producer data is CC BY 4.0, but the exact GIFT-Eval export combines Apache-2.0 metadata with research-purposes-only wording. Hold this copy pending clarification or independent reconstruction from original producer data.
• borg_cluster_data_2011: CC-BY-4.0. hold_packager_terms_ambiguity. Original producer data is CC BY 4.0, but the exact GIFT-Eval export combines Apache-2.0 metadata with research-purposes-only wording. Hold this copy pending clarification or independent reconstruction from original producer data.
• wiki_daily_100k: CC0-1.0. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• quito_hour: CC-BY-4.0. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• quito_min: CC-BY-4.0. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• uci_online_retail_ii: CC-BY-4.0. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• azure_functions_2019: CC-BY-4.0. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• census_qwi_state_sector_firmsize: CC0-1.0. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• eurostat_comext_hs4: LicenseRef-Eurostat-reuse-policy. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• fr_fuel_prices_station_weekly: Etalab-Licence-Ouverte. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• eia_energy_prices_monthly: US-Government-Public-Domain. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• cftc_cot_legacy_futures_weekly: US-Government-Public-Domain. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• freshretailnet_50k_hourly: CC-BY-4.0. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• anaconda_conda_downloads_hourly: CC-BY-4.0. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• eurostat_sts_industry_m: LicenseRef-Eurostat-reuse-policy. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• petrobras_3w_real: CC-BY-4.0. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• bdg2_building_meters: CC-BY-4.0. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• gdelt_v1_events_backfile: GDELT-Terms-of-Use. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• gdelt_v1_events_daily: GDELT-Terms-of-Use. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.
• backblaze_drive_stats: LicenseRef-Backblaze-Drive-Stats. excluded_no_data_resale. Producer permits sale of derivative works but prohibits selling the data itself. Excluded from commercial_core; archived free redistribution retains original conditions.
• statcan_cimt_hs6_2017: LicenseRef-Statistics-Canada-Open-Licence. screened_commercial_redistribution. Producer terms permit commercial reuse and redistribution; preserve the source-specific attribution, notices and stated exclusions.

Synthetic pool

Parametric generators from "Mix, Don't Pick" (arXiv 2606.09912; Apache-2.0 code vendored at birla-ai-labs/mix-dont-pick@6ba44a7). No real data is used. Series are index-only (no clock); every row can be regenerated from its per-series seed in provenance/synthetic/. Tracked separately from the real corpus. This is an experimental univariate pool, not a training-approved 25% allocation. Consult synthetic consumption and review flags for numerical fallbacks, clamps and generator limitations before sampling.

Batch	Generator	Series	FP32 values
synthetic-0001-mixed11	arima	45,000	46,080,000
synthetic-0001-mixed11	chaotic	45,000	46,080,000
synthetic-0001-mixed11	ets	45,000	46,080,000
synthetic-0001-mixed11	fbm	45,000	46,080,000
synthetic-0001-mixed11	garch	45,000	46,080,000
synthetic-0001-mixed11	kernelsynth	45,000	46,080,000
synthetic-0001-mixed11	sde	45,000	46,080,000
synthetic-0001-mixed11	stepfunction	45,000	46,080,000
synthetic-0001-mixed11	timesynth	45,000	46,080,000
synthetic-0001-mixed11	tsi	45,000	46,080,000
synthetic-0001-mixed11	waveform	45,000	46,080,000

Schema and loading

Each row is one entity segment: id, entity_id, source_item_id, source_family, domain, origin, partition, start, frequency_seconds, n_steps, target ([channel][time] FP32), past_feat_dynamic_real ([channel][time] FP32, past-only), upstream source_repo/source_revision/ source_path/source_row, and source index bounds.

from datasets import load_dataset
ds = load_dataset("prabakaranchandran/enterprise-dynamics-corpus-v1", "commercial_core", split="train", streaming=True)

Use the explicit paths in metadata/license_review.json for other loaders. A broad data/real/*/*/train/*.parquet glob includes restricted or review-held archive files. all is an archival config, not a commercial-release selection.

Intended use and limitations

• Pretraining data for forecasting models. Windowing, normalisation and sampling weights belong to the training pipeline and must use observed-only context; fit transforms on train only.
• Training data only. Each source stops at its contamination cutoff, so the test windows of the public benchmarks listed in the exclusion registry are never included; evaluate on those benchmarks. Hold out a development slice at training time if needed.
• Channel order is preserved; semantic channel names are not established for every source.
• Exact-duplicate checks are complete within this corpus; shifted, resampled or approximate copies across public corpora are not exhaustively detected.
• Comply with each source's attribution requirements below when redistributing.