Fixes: FallBack to LoadDataset, Dup Arguments, Savetensor Saving - #27

@garrett361 TestPaddingFree is failing because I think the runners do not have cuda

def test_padding_free(self, model_name: str) -> None:
        torch.manual_seed(42)
>       model, cfg = self.get_fa2_model_and_cfg(model_name)

Yeah expected, I should add a skip if CUDA isn't available

@garrett361 it seems they now run all tests.. previously they only run one test

Previously it was like this #25

Run pytest open_instruct/test_utils.py
============================= test session starts ==============================
platform linux -- Python 3.10.10, pytest-8.4.1, pluggy-1.6.0
rootdir: /home/runner/work/open-instruct/open-instruct
configfile: pyproject.toml
collected [7](https://github.com/garrett361/open-instruct/actions/runs/16009216897/job/45162936130#step:5:8) items

open_instruct/test_utils.py .......                                      [100%]

============================== 7 passed in 14.[8](https://github.com/garrett361/open-instruct/actions/runs/16009216897/job/45162936130#step:5:9)5s ==============================

Thought I saw that as well

It is not necessary to include tools documents in dataset_target_columns

If we look at the transformation function, first tokenization and masking happens, and then columns are removed. tools and documents columns are necessary for correctly applying the chat template (and thereby for tokenization and masking). However, after applying chat template and tokenization, the only columns needed are input_ids, labels, and attention_mask.

Of course, including tools and documents in dataset_target_columns does not impact the correctness, it will just increase the size of cached data.

Also, tools and documents are going to be ignored in default o-i sft_tulu_tokenize_and_truncate_v1. In porting changes to main, we do not have ported additional_inputs in sft_tulu_tokenize_and_truncate_v1.

@garrett361 can we merge this soon? we just ignore the failing test?