From bd4e4ee353892fbf9a04de0450f04ffc14903740 Mon Sep 17 00:00:00 2001 From: Michael Kranzlein <8162250+mkranzlein@users.noreply.github.com> Date: Mon, 21 Aug 2023 22:27:30 -0400 Subject: [PATCH] Checked off some TODOs --- README.md | 3 --- 1 file changed, 3 deletions(-) diff --git a/README.md b/README.md index 7f43791..7505420 100644 --- a/README.md +++ b/README.md @@ -45,11 +45,8 @@ def some_function(x: Float[torch.Tensor, "10, 768"]): I recommend taking a look at the [jaxtyping docs](https://docs.kidger.site/jaxtyping/). ## TODOs -- Eval needs to account for wordpiece splits - Some long documents are too big for GPU vram right now - Batching right now should allow for single documents, but worth testing -- Chunking logic could be more clear -- Last chunk of each doc might be improperly ignored? - Eval needs final pieces put together and then needs to be tested - Decide on consistent variables for type annotations ## Cite HiPool