Clarification on training data temporal range (cutoff date) for Kronos-base

Hi team, first of all — thank you for open-sourcing Kronos. The foundation-model approach for K-line forecasting is exciting and we'd like to evaluate it seriously on US equity markets.

Before we can report any out-of-sample (OOS) performance numbers, we need to know exactly which dates in our test data Kronos has never seen during pretraining. Without this, there's a real risk of accidentally reporting memorization as predictive skill.

Could the team please clarify:

1. Training data temporal range — what is the earliest and latest date of K-line bars included in the Kronos-base pretraining corpus? Is there a hard cutoff (e.g. "bars through 2023-12-31") or was the corpus streaming up until release?

2. US equity coverage — the README mentions 45 global exchanges. Are US equities (in particular NYSE/NASDAQ SP500 constituents and SPY/QQQ ETFs) part of the training set? If yes, at what granularity (1d? 1h? 5m?).

3. Intended OOS window — from the authors' perspective, which calendar period do you consider the "clean held-out" window when users report validation metrics? (i.e. which dates are safe to backtest on without data leakage via pretraining.)

4. Fine-tune vs zero-shot recommendation — if we want to use Kronos-base on US equities from 2024 onward, would you recommend zero-shot inference, or is fine-tuning on a clean post-cutoff window necessary?

A short statement (even just "trained through YYYY-MM") in the README or model card would be hugely helpful for downstream researchers who want to report honest metrics.

Happy to contribute a PR adding this information to the README once you confirm.

Thanks!