fin-ai-lab/Market-1T-1Hz-2019H2-2020-sparse

Market-1T — 1 Hz, July 2019 → December 2020 (sparse)

Code: fin-ai-lab/tfwm downloads this data, trains the TFWM encoders on it, and reproduces the paper's tables.

230,025 ticker-days of 1 Hz US equity market data over 376 trading days (2019-07-01 → 2020-12-31). The window deliberately straddles the February–March 2020 COVID crash and the recovery through 2020, so it contains a genuine regime break rather than a single stationary market.

Three layouts, three repositories

Repository	One record	Size	Use it for
fin-ai-lab/Market-1T-1Hz-2019H2-2020-sparse	one ticker-day, only the seconds that carry data, 07:00–20:00 ET (pre-market and after-hours included)	~27 GB	the raw source: your own fill policy, and the only copy of pre-market / after-hours data
fin-ai-lab/Market-1T-1Hz-2019H2-2020-dense	one ticker-day on the filled 1 Hz grid, regular session only (09:30–16:00 ET)	~23 GB	self-supervised training and probing (the SSL / LeJEPA encoders trained on this)
fin-ai-lab/Market-1T-1Hz-2019H2-2020-daystore	one trading day: every ticker × channel × second, plus precomputed targets, regular session only (09:30–16:00 ET)	~25 GB	cross-sectional / supervised training (the supervised encoders trained on this)

All three carry the same nine channels: bid_price, vwap_all, high, low, ask_price, bid_size, ask_size, volume, n. The encoders' 11 extra view-information channels are computed at load time and are not stored.

This repository: sparse

MosaicML Streaming (MDS) shards under 1Hz_mosaic_mnth_sparse/YYYY/MM/, zstd-compressed, shuffled within each month. Each record covers 07:00–20:00 ET: pre-market, the regular session and after-hours. Only the seconds that carry data are stored. The two dense repositories keep only 09:30–16:00, so the pre-market and after-hours data exist only here. That is also where sparse storage pays off most. Outside the regular session, observed seconds are few and far apart (in a sample of August 2020 records, the 6.5 extended hours held only ~5% of a record's rows), and a filled grid there would be almost all padding. Inside the regular session the data is nearly dense, and the dense layouts compress better (see the size column above). Some NaNs are retained. Nothing is pre-filled, so the fill policy is yours.

Field	Type	Meaning
ticker	str	Symbol
date	str	YYYY-MM-DD
ts_interval	int32 (n_steps,)	Unix time (s) of each row
features	float32 (n_steps, 9)	One row per observed second

from huggingface_hub import snapshot_download
from streaming import StreamingDataset

root = snapshot_download("fin-ai-lab/Market-1T-1Hz-2019H2-2020-sparse", repo_type="dataset",
                         allow_patterns=["1Hz_mosaic_mnth_sparse/2020/08/*"])
ds = StreamingDataset(local=f"{root}/1Hz_mosaic_mnth_sparse/2020/08", shuffle=True, batch_size=64)
rec = ds[0]            # rec["features"].shape == (n_steps, 9)

The dense repository was built from these shards by the project's densifier (sf-densify), which fills the regular session and drops extended hours.

With the project code

fin-ai-lab/tfwm downloads what a run needs by default: uv run train.py fetches the months in its training and evaluation spans into market1t/, verifies every file, decompresses the day store, and builds the cross-sectional target tables. To fetch a range by hand:

uv run python -m market_jepa.market1t 2019-07 2020-01

Log in first (hf auth login). The Hub allows 1,000 API requests per 5 minutes per account, and one month is several hundred files per layout, so a large download pauses for rate limits; the code waits and resumes. To stream from the Hub instead of downloading, use machine=hub (or pass any hf://datasets/fin-ai-lab/Market-1T-1Hz-2019H2-2020-dense/1Hz_mosaic_mnth root).

Related

• TFWM Pre-Trained Encoders: the encoders trained on this data