nanoTSFM-pretrain

Hill-climbing GIFT-Eval with one A100 and one hour.

Pretraining data for nanoTSFM, an open challenge where small time-series foundation models train on one GPU and forecast data they have never seen. The baseline trains on GEP-M in two minutes on an A100 and scores 0.670 relative CRPS on GIFT-Eval, zero-shot. The record page tracks the best result so far.

from datasets import load_dataset

train = load_dataset("abel-lab/nanoTSFM-pretrain", "GEP-M", split="train")
train[0]["target"]  # [variate][time] float32, NaN where a value is missing

Configs

• GEP-S (small): quick experiments and CPU debugging.
• GEP-M (medium, default): the baseline's data, and the smallest slice that gives it its full score.
• GEP-L (large): more varied data for longer runs. After 20,000 steps, GEP-L scored 0.661 and GEP-M 0.693.
• GEP-Val, GEP-Test, GEP-tasks: in-distribution diagnostics on held-out series.

Config names start with their corpus. The GEP configs come from GIFT-Eval Pretrain at revision 6830b62, which shares no series with GIFT-Eval.

Config	Series per source	Points per series	Series (train / val / test)	Points	Download
GEP-S	≤50	≤4,096	2,749 / 313 / 370	33M	94 MB
GEP-M	≤1,000	≤8,192	26,154 / 3,334 / 3,362	252M	615 MB
GEP-L	≤50,000	≤8,192	399,356 / 50,105 / 50,377	2.1B	6.2 GB
GEP-Val	≤100 validation series	≤8,192	3,279	23M	56 MB
GEP-Test	≤100 test series	≤8,192	3,306	25M	62 MB
GEP-tasks	forecast tasks		129 validation / 131 test

All GEP configs draw from the same 87 sources: every source under 250 MB, three mid-size sources, and the first shard of each of the eleven large ones (five for GEP-L). Series are sampled per source without replacement (seed 7) and keep their latest points.

Fields

Field	Type	Meaning
item_id	string	<source>/<item_id>, unique across the corpus
source	string	GIFT-Eval Pretrain dataset
freq	string	pandas frequency, such as H, 15T or W-SUN
source_row	int64	row in the source dataset
target	list of lists of float32	[variate][time]; variates share a time axis

Splits

A hash of item_id fixes each series' split in every config: 80% train, 10% validation, 10% test. A multivariate series stays whole.

import hashlib

def split_of(item_id: str) -> str:
    bucket = int(hashlib.sha256(item_id.encode()).hexdigest()[:8], 16) % 10
    return "test" if bucket == 0 else "validation" if bucket == 5 else "train"

GEP-Val, GEP-Test and GEP-tasks

GEP-Val and GEP-Test hold up to 100 held-out series per source. GEP-tasks defines one GIFT-Eval-style task per source and term (source, term, freq, seasonality, prediction_length, windows, min_length, series): forecast the last windows horizons of every series with at least min_length points. Short horizons follow GIFT-Eval's table by frequency; medium and long terms are 10× and 15× longer, for sub-daily sources only.

Rebuilding

build/build.py rebuilds every GEP config row for row from the source files in build/subset.txt:

uv run build/build.py download raw --shards 5   # 25 GB; without --shards 5, 8.6 GB and no GEP-L
uv run build/build.py all raw .

Acknowledgement

Adapted from Salesforce/GiftEvalPretrain.