Empty-history inputs return ±1e5 quantiles (fix in weights + repro)
#9
by lawrencehui - opened
Hi, I found that Chronos-2 returns extreme quantiles when the context is entirely NaN, e.g. a 10–90% range of −142,000 to +131,000. It shows up in GIFT-Eval's Bitbrains tasks, where servers that were off for the whole context window come back during the forecast. Repro and a suggested one-line code fix are in amazon-science/chronos-forecasting#544.
As an experiment I also fixed it in the weights: 1,536 weights in block 11's MLP (one previously dead neuron), with outputs bit-identical whenever the context has two or more real values. It's here if useful: lawrencehui/chronos-2-empty-history-fix. Bitbrains CRPS improves up to 33%, nothing else changes.
The code fix is the right one upstream. Thanks for open-sourcing the model.