fix pretraining_ on odd datasets

[mapmeld]

Description

I'm interested in using axolotl to pretrain and then finetune on a dataset which I have on HF.

1. Fixes an issue I experienced because the training script expects pretraining to have a split named "train" and single column named "text". I propose adding split and name to the PretrainingDataset parser and a new text_column field.

2. The load_dataset(path, streaming=True) used in pretraining currently returns features: None or features: Unknown, so I propose retrieving the first record to get column names (this is later passed to remove_columns).

As an alternative config change, I could see using SFTDataset and removing the separate PretrainingDataset type

How has this been tested?

Output YAML, ran training script in CoLab notebook

Sample dataset with one column named "Comment"

pretraining_dataset:
  - path: urkopa/comments-lv
    split: train
    text_column: Comment
    type: pretrain

My dataset where each split has a different schema so I need to specify split and name

pretraining_dataset:
  - path: monsoon-nlp/greenbeing-proteins
    split: pretraining
    text_column: sequence
    name: pretraining
    type: pretrain

[NanoCode012]

Hm, is it possible to add unit tests for this?

[winglian]

this all lgtm. thanks!