the dataset leaks future information

[siaochuan]

the dataset leaks future information into every sample via per-window normalization. In finetune_base_model.py (line 127) and finetune_base_model.py (line 132), x_mean/x_std are computed from the full window, which already includes the nominal forecast horizon. Both tokenizer and predictor use this same dataset path, so the leakage contaminates the whole pipeline; see finetune_tokenizer.py (line 20) and finetune_tokenizer.py (line 97). If your goal is unseen-future prediction, this is a stop-ship issue.

[EthanW1024]

致命问题，如果这样，模型不可用

[Lijuncheng75]

未来信息泄露，是真的吗？

[CjwlvCyj]

没仔细看代码，但是这是finetune的模型参数？应该和hugging face上下载到本地的原始预训练模型不一样吧，如果是data leak也只是针对为了做演示的finetune模块？有大神看了原模型可以戳戳