Were the pretrained HF checkpoints retrained after the per-window normalization leakage fix (#227)?

[rlan35]

Hi, thanks for open-sourcing Kronos.

I went through the data pipeline to check the future-leakage concerns from #227 and #265, and wanted to confirm my reading of the current code:

• Inference path looks clean. KronosPredictor.predict() computes the z-score mean/std on the lookback window only (model/kronos.py:544), so no future information leaks at predict time.
• The finetune dataset bug from the dataset leaks future information #227 also appears fixed. finetune/dataset.py:107-117 now computes x_mean/x_std strictly from the lookback portion (past_x = x[:past_len]) rather than over the full lookback+horizon window.

What's still unclear to me — and I think this is what #277 was asking before it went quiet — is whether the pretrained checkpoints on HuggingFace (NeoQuasar/Kronos-small / -base / the tokenizer) were themselves retrained after this normalization fix, or whether they predate it. If the released weights were pretrained with the old full-window normalization, the leakage would effectively be baked into the checkpoints even though the current dataset code is correct.

Could you clarify:

1. Were the released HF weights retrained after the normalization fix, or are they from before it?
2. If they predate the fix, are there plans to re-release retrained checkpoints?

Thanks!