Forgis/hepa-corpus — an open, license-clean, streamable time-series pretraining corpus

28 sources · 95 shards · 567,637 records · 4.58 B channel-timesteps · 13.8 GB on the Hub.

Raw, un-patched, un-windowed, un-normalized continuous float32 sequences of shape (C, T), with per-record provenance, licensing, entity ids and event-label parameters. Built for HEPA-style JEPA pretraining, but nothing here is HEPA-specific.

Every byte in this repo passed a per-source redistribution audit (license_audit_v1.md). Sources that did not clear it are absent entirely — not raw, not derived, not aggregated — and are listed below with the reason.

Quickstart (streaming — no local download)

from datasets import load_dataset
import numpy as np

ds = load_dataset("Forgis/hepa-corpus", split="train", streaming=True)
r = next(iter(ds))
x = np.asarray(r["values"], dtype=np.float32)        # (C, T) raw, un-normalized
print(r["series_id"], r["source"], x.shape, r["freq"], r["entity_id"])

Filter without downloading — every selection is a query over tags, never a physical copy:

energy = ds.filter(lambda r: r["domain"] == "energy")
long   = ds.filter(lambda r: r["n_timesteps"] >= 2048)
events = ds.filter(lambda r: r["has_event"])

Or select shards up front from manifest.json and hand the URLs to your loader — that is the cheap path, because it skips the shards entirely rather than filtering their rows.

One caveat worth knowing before you build a training loop: a sequential stream reads one shard at a time, but ds.shuffle(buffer_size=N) in streaming mode has to pull from many shards at once. On this corpus that is expensive — records here are whole multi-thousand-step series, so a measured buffer_size=64 shuffle moved 3.0 GB over the wire to yield 6 records, against 93 MB for a 10-record sequential read. Shuffle at the shard level from the manifest, and reserve the in-stream buffer for small values.

---

1. Format, and why Parquet rather than WebDataset

Parquet shards under data/<source>/{train,val}-NNNNN.parquet, zstd-compressed, mean 146 MB per shard (min 0.0 MB, max 352 MB).

The plan allowed WebDataset .tar or Parquet. Parquet was chosen because:

1. Zero-dependency streaming. datasets.load_dataset(..., streaming=True) reads Parquet natively over HTTP range requests with no loader script; webdataset would be a new dependency for every consumer of the corpus.
2. Column projection. A dataloader that needs only values + entity_id never transfers the ~15 provenance/label columns. A .tar member is opaque — you pay for all of it or none.
3. The tags stay queryable server-side. The HF dataset viewer and its DuckDB endpoint index Parquet, so source / domain / quality_tags / has_event can be counted and filtered without downloading a byte. That is what makes §4's "splits are queries, not copies" real.

Shards are single-source and single-split so that every manifest row describes a homogeneous unit, and the shard target is set from a per-source compression ratio that the build measures and feeds back (zstd runs 0.12x on USHCN's integer-valued daily climate and 0.83x on ERA5 reanalysis floats — a 7x spread, so a single fixed record count would miss the band badly).

Stated plainly: most shards here are below the 100 MB floor. Two reasons, both structural. Many cleared sources hold less than 100 MB in total, so they cannot fill one shard. And the first shard of each source is sized against a deliberately pessimistic 0.7 ratio (so it can never blow past 500 MB), which undershoots for the highly compressible sources. The band is met where it matters: weatherbench_daily, which is 81% of the corpus by bytes, shards at 301 MB (min 88, max 352).

Per-record schema

field	type	meaning
series_id	string	"<source>:<native id>", globally unique
source	string	audited config name (row in the provenance table)
domain	string	energy / climate / traffic / web / health / econ / retail / other
native_C	int32	number of channels as shipped by the source — no padding, no grouping
channel_names	list[string]	the upstream column name of each channel, in order
values	list[list[float32]]	raw (C, T) — un-patched, un-windowed, un-normalized
n_timesteps	int32	T
freq	string	inferred sampling frequency (15min, h, D, W, M, Q, Y, …)
start_time_ms	int64	epoch-ms of t=0. Series are regularly sampled at freq; records where they are not carry the irregular_timestamps tag
entity_id	string	physical entity (station, grid cell, household, page) — the split key
quality_tags	list[string]	see §5
split_role	string	train | val (entity-disjoint, §4)
has_classification, class_label	bool, int32	always false / −1 — no source here ships native class labels
has_ad	bool	always false — no source here ships native anomaly labels
has_event, event_label_origin	bool, string	added_by_us under one deterministic rule, §3
event_rule_id, event_description	string	which rule, and the physical event it encodes
event_threshold, event_horizon, event_calib_end, event_channel	float32, int32×3	the parameters that exactly reconstruct the labels
n_event_positives	int32	positives under that rule, so you can filter by positive rate for free
license, license_url	string	the redistribution basis for this record

Timestamps are not stored per step. start_time_ms + freq reconstructs the index exactly for regular series and halves the on-disk size (an int64 per step costs more than the float32 value it labels). Records whose upstream index is not regular are tagged irregular_timestamps.

---

2. What is here, and what is deliberately not

The audit covered the 51-config real-world Chronos-2 subset (autogluon/chronos_datasets minus the synthetic training_corpus_* and the 14 weatherbench_hourly_* configs). The published corpus and the corpus we pretrain on are different sets, because hosting on the Hub is redistribution and downloading under a source's own research-use terms is not. Both gates are applied separately:

set	n	where it lives
audited real-world Chronos-2 subset	51	—
− non-redistributable (6 EXCLUDE + 4 LINK_ONLY)	−10	local pretraining only; never on the Hub
− conditional, pending Zenodo re-sourcing (m4_*×6, nn5, solar)	−9	local only; deferred from the Hub
= license-cleared for redistribution	32	—
− contributed zero usable records (every series shorter than 64 steps)	−4	dropped at build time
− cleared but not yet built	−0	see the resume command in §8
= published Forgis/hepa-corpus v1	28	this repo

The 32 cleared configs are selected by the predicate verdict == "INCLUDE" and redistributable is True and config not in DEFERRED_CONDITIONAL, and 28 of them actually contributed data. The ones that did not:

source	domain	why it contributed nothing
monash_car_parts	retail	every series is shorter than 64 steps
monash_m1_yearly	other	every series is shorter than 64 steps
monash_m3_yearly	other	every series is shorter than 64 steps
monash_tourism_yearly	econ	every series is shorter than 64 steps

These are low-frequency economic and demographic series — a 20-point yearly index cannot form even two P=16 patches, and this repo's existing corpus loader already uses a stricter min_len=256. They are dropped, not silently counted.

The gate keys on the redistributable boolean, not on a text match for "EXCLUDE" — a text match would silently admit the four LINK_ONLY configs, which are equally non-redistributable.

If you pretrain on this repo and compare against a HEPA number computed on the 51-config local corpus, the corpora differ and some of any gap is data, not method. In particular monash_rideshare — one of only three natively-multivariate configs in the report's set — is excluded here, so the published corpus is less natively-MV than the one the report used.

Not redistributed (never uploaded, in any form)

source	verdict	domain	origin	why it is absent
dominick	EXCLUDE	retail	origin	Kilts Center states 'These data are for academic research purposes only' with no redistribution grant; Monash's CC-BY-4.0 Zenodo re-deposit (10.5281/zenodo.4654802) cannot cure the upstream owner's restriction, so we do not re-host.
exchange_rate	EXCLUDE	econ	origin	The autogluon card claims MIT but the source repository has NO LICENSE file (GitHub license API returns 404) and its README names no upstream FX data provider; UNVERIFIED -> EXCLUDE.
m5	EXCLUDE	retail	origin	Proprietary Walmart retail data released only under Kaggle competition rules; the rules page is JavaScript-gated and could not be retrieved, so no redistribution grant can be cited. UNVERIFIED -> EXCLUDE. Kaggle's standard competition-data clause is non-commercial and participant-scoped in any case.
mexico_city_bikes	EXCLUDE	traffic	origin	The Ecobici open-data page and its terms-and-conditions page state no license, no reuse grant and no named open-data licence; the CDMX open-data portal terms page was unreachable. UNVERIFIED -> EXCLUDE.
monash_fred_md	EXCLUDE	econ	origin	FRED's legal terms forbid redistributing third-party proprietary content without the copyright holder's written permission and require permission for 'pre-approval required' copyrighted series, and the St. Louis Fed states it cannot grant permission on their behalf; FRED-MD aggregates such third-party series. Monash's CC-BY-4.0 re-deposit cannot cure this, so we do not re-host. Lead judgment requested.
monash_rideshare	EXCLUDE	traffic	origin	Weakest chain of title in the corpus: the underlying data is an individual's scrape of Uber and Lyft commercial pricing APIs posted to Kaggle with no verifiable license, so neither Kaggle nor Monash can be shown to have had the right to license it CC-BY-4.0. Lead judgment requested.
taxi_1h	LINK_ONLY	traffic	origin	The AWS Registry of Open Data lists the TLC trip records' license as the NYC.gov Terms of Use, which state 'All rights are reserved' and grant no reuse; the Apache-2.0 license on the gluon-ts fork covers that repository, not the City's underlying data. Reference the TLC source, do not re-host.
taxi_30min	LINK_ONLY	traffic	origin	The AWS Registry of Open Data lists the TLC trip records' license as the NYC.gov Terms of Use, which state 'All rights are reserved' and grant no reuse; the Apache-2.0 license on the gluon-ts fork covers that repository, not the City's underlying data. Reference the TLC source, do not re-host.
uber_tlc_daily	LINK_ONLY	traffic	origin	The FiveThirtyEight repository has NO LICENSE file (GitHub license API returns 404) and the data is a FOIL disclosure of NYC TLC records whose publisher reserves all rights. UNVERIFIED -> reference link-only, never re-host.
uber_tlc_hourly	LINK_ONLY	traffic	origin	The FiveThirtyEight repository has NO LICENSE file (GitHub license API returns 404) and the data is a FOIL disclosure of NYC TLC records whose publisher reserves all rights. UNVERIFIED -> reference link-only, never re-host.

LINK_ONLY sources are reachable at their origin link above; we do not re-host them. The four NYC TLC configs are LINK_ONLY because NYC.gov's Terms of Use read "All rights are reserved" — the Apache-2.0 on the gluon-ts fork covers that repository, not the City's data.

Deferred — cleared in principle, not shipped yet

source	license basis	status
m4_daily	CC-BY-4.0	requires_resourcing_from_zenodo
m4_hourly	CC-BY-4.0	requires_resourcing_from_zenodo
m4_monthly	CC-BY-4.0	requires_resourcing_from_zenodo
m4_quarterly	CC-BY-4.0	requires_resourcing_from_zenodo
m4_weekly	CC-BY-4.0	requires_resourcing_from_zenodo
m4_yearly	CC-BY-4.0	requires_resourcing_from_zenodo
nn5	CC-BY-4.0	requires_resourcing_from_zenodo
solar	CC-BY-4.0	requires_resourcing_from_zenodo
solar_1h	CC-BY-4.0	requires_resourcing_from_zenodo

Action required before these can ship. Re-source the identical series from the Monash Time Series Forecasting Repository Zenodo deposit named in the license link and re-attribute to Godahewa et al. (2021). They are not shipped from the autogluon copy because the origin that copy points at carries no license at all — Mcompetitions/M4-methods has no LICENSE file, NN5 publishes none, and NREL's own disclaimer was unreachable. Shipping those bytes while citing Monash's CC-BY-4.0 would mean the citation does not describe the artifact, which is precisely the defect the audit exists to prevent.

---

3. Labels

has_classification / has_AD: native only — and there are none

Every source here is a forecasting corpus. None ships class labels or anomaly annotations, so both flags are false on every record. We did not synthesize either: there is no rule over these series whose ground truth is certain, and a plausible-looking guess in a public corpus is worse than an absent column.

has_event: one deterministic rule, applied only where the event is physical

event_label_origin is added_by_us wherever it is set — no source ships native event labels. The single rule used is:

p95_threshold_crossing_v1:
y[t] = 1 iff max(values[event_channel, t:t+H]) > thr, where thr = P95(values[event_channel, :calib_end]) computed on the DISJOINT calibration region values[:, :calib_end], calib_end = floor(0.30*T), and labels are defined only for calib_end <= t <= T-H. Deterministic: no fitting, no heuristics, exactly reproducible from (event_threshold, event_horizon, event_calib_end, event_channel).

Reconstruct labels exactly:

import numpy as np
x   = np.asarray(r["values"], np.float32)[r["event_channel"]]
H, c = r["event_horizon"], r["event_calib_end"]
seg = x[c:]
w   = np.lib.stride_tricks.sliding_window_view(seg, H)[: len(seg) - H]
y   = np.nanmax(w, axis=1) > r["event_threshold"]     # labels for t = c .. T-H-1
assert y.sum() == r["n_event_positives"]

This is deterministic — no fitting, no tuning, no heuristic — and the calibration region values[:, :calib_end] is disjoint from every labelled timestep, so the threshold cannot leak the label.

It is applied only where a P95 crossing is a recognised physical event, and only to records with T ≥ 512 (below that a 30% calibration region is too small for a stable P95):

source	records labelled	event
electricity_15min	370 / 370	load spike (P95 15-min household demand exceedance)
ercot	8 / 8	grid load spike (P95 zonal demand exceedance)
monash_australian_electricity	5 / 5	grid load spike (P95 half-hourly demand exceedance)
monash_electricity_hourly	321 / 321	load spike (P95 hourly client demand exceedance)
monash_kdd_cup_2018	268 / 270	air-quality episode (P95 pollutant concentration exceedance)
monash_london_smart_meters	5,555 / 5,559	load spike (P95 half-hourly household demand exceedance)
monash_pedestrian_counts	66 / 66	crowd surge (P95 hourly pedestrian count exceedance)
monash_saugeenday	1 / 1	high-flow / flood peak (P95 daily river flow exceedance)
monash_temperature_rain	422 / 422	heat extreme (P95 daily mean-temperature exceedance)
monash_traffic	862 / 862	congestion event (P95 hourly occupancy exceedance)
monash_weather	3,010 / 3,010	per-variable, see below
ushcn_daily	1,216 / 1,218	heavy-precipitation event (P95 daily PRCP exceedance)
weatherbench_daily	34,816 / 225,280	per-variable, see below
weatherbench_weekly	34,816 / 225,280	per-variable, see below
wiki_daily_100k	100,000 / 100,000	traffic surge (P95 daily pageview exceedance)
wind_farms_hourly	328 / 328	high-output wind event (P95 hourly farm power exceedance)

weatherbench_* and monash_weather each bundle several distinct physical fields under one config name, with the field in a subset column (exported as the variable:<name> quality tag). For those, eligibility and wording are resolved per variable:

source	variable	event
monash_weather	maxtemp	heat extreme (P95 exceedance of daily maximum temperature)
monash_weather	mintemp	warm-night extreme (P95 exceedance of daily minimum temperature)
monash_weather	rain	heavy-precipitation event (P95 exceedance of daily rainfall)
monash_weather	solar	high-insolation day (P95 exceedance of daily solar exposure)
weatherbench_daily	10m_u_component_of_wind	high-wind event (P95 exceedance of 10m u-wind)
weatherbench_daily	10m_v_component_of_wind	high-wind event (P95 exceedance of 10m v-wind)
weatherbench_daily	10m_wind_speed	high-wind event (P95 exceedance of 10m wind speed)
weatherbench_daily	2m_temperature	heat extreme (P95 exceedance of 2m temperature)
weatherbench_daily	temperature	warm anomaly (P95 exceedance of air temperature)
weatherbench_daily	total_precipitation	heavy-precipitation event (P95 exceedance of total precipitation)
weatherbench_weekly	10m_u_component_of_wind	high-wind event (P95 exceedance of 10m u-wind)
weatherbench_weekly	10m_v_component_of_wind	high-wind event (P95 exceedance of 10m v-wind)
weatherbench_weekly	10m_wind_speed	high-wind event (P95 exceedance of 10m wind speed)
weatherbench_weekly	2m_temperature	heat extreme (P95 exceedance of 2m temperature)
weatherbench_weekly	temperature	warm anomaly (P95 exceedance of air temperature)
weatherbench_weekly	total_precipitation	heavy-precipitation event (P95 exceedance of total precipitation)

Everything else gets has_event=false. That deliberately includes the low-frequency economic and demographic series (monash_m1_*, monash_m3_*, monash_tourism_*, monash_cif_2016, monash_car_parts, monash_hospital, monash_nn5_weekly) — on a yearly or quarterly index a P95 crossing is trend, not an event — and the weatherbench fields not listed above (geopotential, potential_vorticity, pressure-level winds, total_cloud_cover, toa_incident_solar_radiation), where we could not assert certain ground truth.

---

4. Methodology

Normalization — in the dataloader, never on disk

Nothing here is normalized, and that is deliberate. Level and scale carry signal (exp-08 found that RevIN made a HEPA encoder blind to per-window level on energy events), so the choice of normalization is a modelling decision that must stay ablatable. Crop the raw (C, T) first, then normalize the crop in the worker. Same argument for patch size and window length: they are never baked in, so P=16 and T=2048 are choices a consumer makes, not constraints this corpus imposes.

Pseudo-multivariate grouping — also downstream

Records are stored at their native C. Most sources here are univariate (native_C=1); only ushcn_daily (5 channels) and monash_temperature_rain (78) are natively multivariate. The HEPA recipe groups c_group=8 univariate series from the same source into a (T, C=8) sample, sampling with replacement where fewer than 8 are usable, with each source contributing equally per epoch (balanced sampling). That grouping is not materialised on disk: it is a sampler choice, it changes between corpus iterations, and freezing it would make native_C a lie. Group at load time, over records filtered from one source.

De-duplication

Within a source, series are unique by upstream id. Across sources, overlap is structural rather than incidental (e.g. monash_electricity_hourly and electricity_15min are different resamplings of the same UCI donation), so it is handled by provenance, not by hashing: the source field and the manifest let a consumer drop a whole lineage. Records that are constant, or have no finite values, or are shorter than 64 steps, are dropped at build time. Records that merely contain some NaNs are kept and tagged has_nan — missingness is signal, and masking it is the dataloader's decision, not ours.

Splits are entity-disjoint, and are queries over the manifest

split_role is assigned by val iff sha1(entity_id)[:8] % 100 < 5 — a hash of the entity, never a random row or window split. All records of one physical entity (a USHCN station, a WeatherBench grid cell, a London household, a Wikipedia page) land on the same side, so a train/val pair never shares a device. entity_id is exported on every record so you can re-split by entity or by contiguous time blocks with a buffer gap without re-downloading anything.

The 5% is global, not per-source, so a source with few entities can land entirely in train and have no val-* shard at all (monash_kdd_cup_2018, for instance, has 59 entities and drew zero). That is a property of an honest entity-level split at small n, not a bug — but if you need per-source validation, re-split on entity_id yourself rather than assuming every source is represented in validation.

Because a shard is single-source and single-split, iterations and splits are manifest queries:

import json, urllib.request
m = json.load(urllib.request.urlopen(
    "https://huggingface.co/datasets/Forgis/hepa-corpus/resolve/main/manifest.json"))
v1_train = [s["shard"] for s in m["shards"]
            if s["split_role"] == "train" and "v1-chronos2" in s["corpus_versions"]]
energy_events = [s["shard"] for s in m["shards"]
                 if s["domain"] == "energy" and s["label_flags"]["has_event"]]

Adding corpus v2/v3 means appending shards with new corpus_versions tags — never re-sharding, never copying.

OOD test set: held out entirely

The six downstream event-forecasting evaluation datasets — C-MAPSS FD001, C-MAPSS FD002, C-MAPSS FD003, ETTm1, GECCO, PSM — contribute zero series to this corpus. None of them is a chronos_datasets config, and the build asserts every uploaded source name against the held-out set. Frozen-encoder transfer to those six is therefore genuinely zero-shot with respect to this corpus.

---

5. Composition

domain	sources	shards	records	rec %	channel-timesteps	ct %	size
climate	7	53	455,481	80.2%	4,066.1 M	88.8%	13.13 GB
web	1	3	100,000	17.6%	274.1 M	6.0%	0.33 GB
energy	8	15	7,240	1.3%	222.3 M	4.9%	0.35 GB
traffic	2	4	928	0.2%	18.3 M	0.4%	0.04 GB
other	6	12	3,037	0.5%	0.3 M	0.0%	0.00 GB
econ	4	8	951	0.2%	0.2 M	0.0%	0.00 GB
total	28	95	567,637	100%	4,581.3 M	100%	13.84 GB

quality_tags vocabulary: corpus_v1_chronos2, domain:<d>, freq:<f>, variable:<v>, univariate / native_mv, has_nan, irregular_timestamps, len_lt_512 / len_ge_512 / len_ge_2048, event_labeled, weak_chain_of_title, copernicus_notice_required.

Scale, stated against the target

The exp-09 plan targeted a Base-scale ~250–500 GB corpus. The realised v1 is 13.8 GB. That gap is not a shortfall of the pipeline — it is the size of the license-clean real-world Chronos-2 subset itself. The entire autogluon/chronos_datasets repo is ~895 GB, but 779 GB of that is weatherbench_hourly_* and 91 GB is synthetic training_corpus_*, both of which the report's own corpus definition excludes; the remaining real-world subset is ~25 GB before the licensing gate removes a further 19 configs. Reaching 250–500 GB requires more sources (LOTSA, v2), not more of these — the pipeline streams and scales, the data does not exist at that size under this corpus definition.

---

6. Provenance

One row per audited source. used_in_pretraining and redistributed are independent: everything is used locally under its own research-use terms; only the cleared subset is redistributed here.

source	domain	origin link	license	#series	native C	length range	freq	has_classification	has_AD	has_event	event_label_origin	used_in_pretraining	redistributed
dominick	retail	origin	proprietary-academic-use-only	—	—	—	—	no	no	—	—	yes	no — not redistributed here
electricity_15min	energy	origin	CC-BY-4.0	370	1	16,032–140,256	15min	no	no	yes	added_by_us (p95_threshold_crossing_v1)	yes	yes
ercot	energy	origin	ERCOT-Terms-of-Use (redistribution expressly permitted)	8	1	154,872–154,872	h	no	no	yes	added_by_us (p95_threshold_crossing_v1)	yes	yes
exchange_rate	econ	origin	unknown	—	—	—	—	no	no	—	—	yes	no — not redistributed here
m4_daily	other	origin	CC-BY-4.0	—	—	—	—	no	no	—	—	yes	no — deferred pending Zenodo re-sourcing
m4_hourly	other	origin	CC-BY-4.0	—	—	—	—	no	no	—	—	yes	no — deferred pending Zenodo re-sourcing
m4_monthly	other	origin	CC-BY-4.0	—	—	—	—	no	no	—	—	yes	no — deferred pending Zenodo re-sourcing
m4_quarterly	other	origin	CC-BY-4.0	—	—	—	—	no	no	—	—	yes	no — deferred pending Zenodo re-sourcing
m4_weekly	other	origin	CC-BY-4.0	—	—	—	—	no	no	—	—	yes	no — deferred pending Zenodo re-sourcing
m4_yearly	other	origin	CC-BY-4.0	—	—	—	—	no	no	—	—	yes	no — deferred pending Zenodo re-sourcing
m5	retail	origin	unknown (Kaggle competition rules, not retrievable)	—	—	—	—	no	no	—	—	yes	no — not redistributed here
mexico_city_bikes	traffic	origin	unknown	—	—	—	—	no	no	—	—	yes	no — not redistributed here
monash_australian_electricity	energy	origin	CC-BY-4.0	5	1	230,736–232,272	30min	no	no	yes	added_by_us (p95_threshold_crossing_v1)	yes	yes
monash_car_parts	retail	origin	CC-BY-4.0	—	—	—	—	no	no	—	—	yes	no — cleared, but not yet built (see §5)
monash_cif_2016	econ	origin	CC-BY-4.0	60	1	65–120	M	no	no	no	n/a	yes	yes
monash_covid_deaths	other	origin	CC-BY-4.0	233	1	212–212	D	no	no	no	n/a	yes	yes
monash_electricity_hourly	energy	origin	CC-BY-4.0	321	1	26,304–26,304	h	no	no	yes	added_by_us (p95_threshold_crossing_v1)	yes	yes
monash_electricity_weekly	energy	origin	CC-BY-4.0	321	1	156–156	W	no	no	no	n/a	yes	yes
monash_fred_md	econ	origin	conflicting: Zenodo says CC-BY-4.0, but FRED legal terms restrict third-party series	—	—	—	—	no	no	—	—	yes	no — not redistributed here
monash_hospital	other	origin	CC-BY-4.0	767	1	84–84	M	no	no	no	n/a	yes	yes
monash_kdd_cup_2018	climate	origin	CC-BY-4.0	270	1	9,504–10,920	h	no	no	partial (268/270)	added_by_us (p95_threshold_crossing_v1)	yes	yes
monash_london_smart_meters	energy	origin	CC-BY-4.0	5,559	1	288–39,648	30min	no	no	partial (5555/5559)	added_by_us (p95_threshold_crossing_v1)	yes	yes
monash_m1_monthly	other	origin	CC-BY-4.0	479	1	65–150	M	no	no	no	n/a	yes	yes
monash_m1_quarterly	other	origin	CC-BY-4.0	44	1	64–114	Q	no	no	no	n/a	yes	yes
monash_m1_yearly	other	origin	CC-BY-4.0	—	—	—	—	no	no	—	—	yes	no — cleared, but not yet built (see §5)
monash_m3_monthly	other	origin	CC-BY-4.0	1,428	1	66–144	M	no	no	no	n/a	yes	yes
monash_m3_quarterly	other	origin	CC-BY-4.0	86	1	64–72	Q	no	no	no	n/a	yes	yes
monash_m3_yearly	other	origin	CC-BY-4.0	—	—	—	—	no	no	—	—	yes	no — cleared, but not yet built (see §5)
monash_nn5_weekly	econ	origin	CC-BY-4.0	111	1	113–113	W	no	no	no	n/a	yes	yes
monash_pedestrian_counts	traffic	origin	CC-BY-4.0	66	1	576–96,424	h	no	no	yes	added_by_us (p95_threshold_crossing_v1)	yes	yes
monash_rideshare	traffic	origin	conflicting: Zenodo says CC-BY-4.0, upstream provenance is scraped commercial API data	—	—	—	—	no	no	—	—	yes	no — not redistributed here
monash_saugeenday	climate	origin	CC-BY-4.0	1	1	23,741–23,741	h	no	no	yes	added_by_us (p95_threshold_crossing_v1)	yes	yes
monash_temperature_rain	climate	origin	CC-BY-4.0	422	76	725–725	D	no	no	yes	added_by_us (p95_threshold_crossing_v1)	yes	yes
monash_tourism_monthly	econ	origin	CC-BY-4.0	366	1	91–333	M	no	no	no	n/a	yes	yes
monash_tourism_quarterly	econ	origin	CC-BY-4.0	414	1	64–130	Q	no	no	no	n/a	yes	yes
monash_tourism_yearly	econ	origin	CC-BY-4.0	—	—	—	—	no	no	—	—	yes	no — cleared, but not yet built (see §5)
monash_traffic	traffic	origin	CC-BY-4.0	862	1	17,544–17,544	h	no	no	yes	added_by_us (p95_threshold_crossing_v1)	yes	yes
monash_weather	climate	origin	CC-BY-4.0	3,010	1	1,332–65,981	D	no	no	yes	added_by_us (p95_threshold_crossing_v1)	yes	yes
nn5	econ	origin	CC-BY-4.0	—	—	—	—	no	no	—	—	yes	no — deferred pending Zenodo re-sourcing
solar	energy	origin	CC-BY-4.0	—	—	—	—	no	no	—	—	yes	no — deferred pending Zenodo re-sourcing
solar_1h	energy	origin	CC-BY-4.0	—	—	—	—	no	no	—	—	yes	no — deferred pending Zenodo re-sourcing
taxi_1h	traffic	origin	NYC.gov Terms of Use - all rights reserved (no redistribution grant)	—	—	—	—	no	no	—	—	yes	no — not redistributed here
taxi_30min	traffic	origin	NYC.gov Terms of Use - all rights reserved (no redistribution grant)	—	—	—	—	no	no	—	—	yes	no — not redistributed here
uber_tlc_daily	traffic	origin	unknown	—	—	—	—	no	no	—	—	yes	no — not redistributed here
uber_tlc_hourly	traffic	origin	unknown	—	—	—	—	no	no	—	—	yes	no — not redistributed here
ushcn_daily	climate	origin	US-Government-Public-Domain (17 U.S.C. Sec. 105)	1,218	5	5,906–59,283	D	no	no	partial (1216/1218)	added_by_us (p95_threshold_crossing_v1)	yes	yes
weatherbench_daily	climate	origin	CC-BY-4.0	225,280	1	14,609–14,610	D	no	no	partial (34816/225280)	added_by_us (p95_threshold_crossing_v1)	yes	yes
weatherbench_weekly	climate	origin	CC-BY-4.0	225,280	1	2,087–2,087	W	no	no	partial (34816/225280)	added_by_us (p95_threshold_crossing_v1)	yes	yes
wiki_daily_100k	web	origin	CC0-1.0	100,000	1	2,741–2,741	D	no	no	yes	added_by_us (p95_threshold_crossing_v1)	yes	yes
wind_farms_daily	energy	origin	CC-BY-4.0	328	1	71–366	D	no	no	no	n/a	yes	yes
wind_farms_hourly	energy	origin	CC-BY-4.0	328	1	1,715–8,784	h	no	no	yes	added_by_us (p95_threshold_crossing_v1)	yes	yes

---

7. Licensing

Hosting on the Hub is redistribution. A source ships here only if its license permits redistribution and research use, verified against the upstream rights holder — not merely against whatever a downstream re-publisher asserted. Redistribution basis for every uploaded source:

license	sources	which
CC-BY-4.0	25	electricity_15min, monash_australian_electricity, monash_cif_2016, monash_covid_deaths, monash_electricity_hourly, monash_electricity_weekly, monash_hospital, monash_kdd_cup_2018, monash_london_smart_meters, monash_m1_monthly, monash_m1_quarterly, monash_m3_monthly, monash_m3_quarterly, monash_nn5_weekly, monash_pedestrian_counts, monash_saugeenday, monash_temperature_rain, monash_tourism_monthly, monash_tourism_quarterly, monash_traffic, monash_weather, weatherbench_daily, weatherbench_weekly, wind_farms_daily, wind_farms_hourly
ERCOT-Terms-of-Use (redistribution expressly permitted)	1	ercot
US-Government-Public-Domain (17 U.S.C. Sec. 105)	1	ushcn_daily
CC0-1.0	1	wiki_daily_100k

Caveat — Monash chain of title

12 uploaded sources rest on the Monash Time Series Forecasting Repository's CC-BY-4.0 Zenodo assertion over data Monash did not originate: monash_australian_electricity, monash_cif_2016, monash_kdd_cup_2018, monash_m1_monthly, monash_m1_quarterly, monash_m3_monthly, monash_m3_quarterly, monash_tourism_monthly, monash_tourism_quarterly, monash_traffic, wind_farms_daily, wind_farms_hourly.

We ship them because Monash is a citable, DataCite-registered scholarly deposit and the upstream material is public — but the chain of title is weaker than a direct grant from the originator, and you should know that before building on them. Four Monash configs have genuinely clean chains and carry no such flag: monash_covid_deaths (JHU CSSE is itself CC-BY-4.0), monash_electricity_hourly / monash_electricity_weekly (UCI CC-BY-4.0), monash_london_smart_meters (London Datastore CC-BY). Records from flagged sources carry the weak_chain_of_title quality tag, so you can exclude them with one filter.

Required notice — Copernicus (weatherbench_daily, weatherbench_weekly)

Contains modified Copernicus Climate Change Service Information [2019]. Neither the European Commission nor ECMWF is responsible for any use that may be made of the Copernicus Information or Data it contains.

Records from these sources carry the copernicus_notice_required tag. If you redistribute them onward, carry this notice.

Attribution

Attribute per source using the provenance table. ushcn_daily is US federal public domain (17 U.S.C. §105) but still asks for the Menne et al. (2009) citation. Monash-sourced configs should cite Godahewa et al. (2021). The upstream aggregation is autogluon/chronos_datasets.

Report a problem

If you are a rights holder and believe something here should not be, open a discussion on this repo and we will remove it immediately, before adjudicating.

---

8. Reproducing this build

export HEPA_DATA_DIR=$HOME/.hepa/data
python3 experiments/exp-09/scripts/build_hf_corpus.py --configs all-cleared --shard-mb 300

The build is resumable — it checkpoints per shard and re-running skips completed sources. It will refuse any config that is not license-cleared, by name, at the command line.

Built 2026-08-05T12:08:51.769655+00:00 · manifest version 1.0.