README.md

metadata

license: other
license_name: per-source
license_link: https://github.com/maxgfr/skycmd/blob/main/DATA_LICENSES.md
pretty_name: skycmd time series (s6-cast training)
task_categories:
  - time-series-forecasting
tags:
  - skycmd
configs:
  - config_name: games
    data_files: data/games.parquet
  - config_name: london_smart_meters
    data_files: data/london_smart_meters.parquet
  - config_name: pedestrian_counts
    data_files: data/pedestrian_counts.parquet
  - config_name: rideshare
    data_files: data/rideshare.parquet
  - config_name: ushcn_daily
    data_files: data/ushcn_daily.parquet
  - config_name: wiki_daily
    data_files: data/wiki_daily.parquet
  - config_name: wind_farms_daily
    data_files: data/wind_farms_daily.parquet
  - config_name: wind_farms_hourly
    data_files: data/wind_farms_hourly.parquet

skycmd time series (s6-cast training)

The training series of the s6-cast forecaster: public series absent from GIFT-Eval and Chronos' zero-shot benchmark, and the state variables of the skycmd games.

Part of skycmd, tiny decision models trained from scratch in Rust that run in the browser. Written by sky-data fetch-series and sky-train cast-synth --games at commit 84c6d854. Licences of every set: DATA_LICENSES.md.

Each row is one series: id (<source>:<series>) and values (a list of floats, in time order). The s6-cast trainer holds out one series in ten, by id, for validation; windows and synthetic series (sky-train cast-synth) are drawn from these at training time.

Sources and licences

Each config is one source and keeps its source's licence (share-alike sources stay share-alike). Attribution goes to the upstream authors; the conversion is ours.

config	upstream	licence	how it is used	rows
games	ours: state variables of the skycmd games (sky-train cast-synth --games)	apache-2.0	60 oracle episodes per game	1,549
london_smart_meters	autogluon/chronos_datasets monash_london_smart_meters (Monash archive, Godahewa et al. 2021)	cc-by-4.0	first 2,000 series, last 8,192 points	2,000
pedestrian_counts	autogluon/chronos_datasets monash_pedestrian_counts (Monash archive, Godahewa et al. 2021)	cc-by-4.0	all series	66
rideshare	autogluon/chronos_datasets monash_rideshare (Monash archive, Godahewa et al. 2021)	cc-by-4.0	every numeric column of the first 200 rows	200
ushcn_daily	autogluon/chronos_datasets ushcn_daily (NOAA NCEI)	public domain (U.S. Government work)	every numeric column of the first 1,300 rows	1,300
wiki_daily	autogluon/chronos_datasets wiki_daily_100k (Wikimedia pageviews)	cc0-1.0	first 5,000 series	5,000
wind_farms_daily	autogluon/chronos_datasets wind_farms_daily (Monash archive, Godahewa et al. 2021)	cc-by-4.0	all series, last 8,192 points	337
wind_farms_hourly	autogluon/chronos_datasets wind_farms_hourly (Monash archive, Godahewa et al. 2021)	cc-by-4.0	all series, last 8,192 points	337

Files

file	rows
data/games.parquet	1,549
data/london_smart_meters.parquet	2,000
data/pedestrian_counts.parquet	66
data/rideshare.parquet	200
data/ushcn_daily.parquet	1,300
data/wiki_daily.parquet	5,000
data/wind_farms_daily.parquet	337
data/wind_farms_hourly.parquet	337

from datasets import load_dataset
ds = load_dataset("maxgfr/skycmd-series", "games")

skycmd wrote no row with a language model: the targets come from the upstream labels (some upstream sets carry their own teacher soft labels, see their cards), the games' oracles or our generators.