# Context is Key

Canonical source: <https://huggingface.co/datasets/ServiceNow/context-is-key>
Official code: <https://github.com/ServiceNow/context-is-key-forecasting>
Official benchmark viewer: <https://servicenow.github.io/context-is-key-forecasting/v0/>
Introducing source: [Context is Key](../../wiki/sources/context-is-key-2024.md)

## Dataset Type

Natural-language context-aided probabilistic forecasting benchmark.

## Temporal Structure

CiK contains univariate numeric time-series forecasting instances paired with natural-language context. Each instance provides a historical numeric window, textual context, future timestamps, and hidden future values for evaluation. The benchmark is designed for direct evaluation rather than dataset-specific training.

## Context Interface

The context is text and is split in the Hugging Face schema into `background`, `scenario`, and `constraints`. The benchmark tags context sources as intemporal information, future information, historical information, covariate information, and causal information.

## Actions Or Interventions

CiK is not an action-conditioned world-model dataset. Some contexts describe future events, constraints, causal relationships, or intervention-like scenarios, but these condition the forecast rather than exposing logged actions, control inputs, or a policy interface.

## Reported Scale

- 71 manually designed tasks in the paper.
- 355 examples in the Hugging Face test split, corresponding to five instances per task.
- 2,644 underlying time series across seven domains in the paper.
- Seven domains: climatology, economics, energy, mechanics, public safety, transportation, and retail.
- Sampling frequencies range from 10 minutes to monthly.
- The paper reports that 95% of tasks use real-world data and 5% use simulated dynamical systems.

## Splits And Versions

The Hugging Face dataset card says CiK is intended as a benchmark and uses splits to represent dataset versions:

- `test`: latest version of the dataset.
- `ICML2025`: version used for the ICML 2025 experimental results.

The dataset card notes that `test` corrects scaled-number issues in two bivariate SVAR causal-context tasks relative to `ICML2025`.

## Access And License Notes

The Hugging Face dataset repository lists Apache-2.0. The dataset card says the original time-series sources are public domain or CC-BY-4.0.

## Suitability Note

Use CiK to evaluate text-conditioned time-series forecasting, context understanding, constraint following, and event-aware probabilistic forecasts. Do not use it as evidence for multivariate time-series modeling, action-conditioned control, or passive numeric-only forecasting.