Is it possible to take data from the dataset sequentially during training?

[pitkad]

Hello!
For my task there is a large dataset generated automatically. This dataset includes a number of incorrect pairs, where "output" is not related to "input" in any way.
I came up with the idea of filtering this dataset by analyzing the "loss" function for each element using the following command:

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py
--stage sft
--model_name_or_path krshtfr01
--do_train
--dataset test-reseach
--template default
--finetuning_type lora
--lora_target q_proj,v_proj
--output_dir test-reseach4
--overwrite_cache
--per_device_train_batch_size 1
--gradient_accumulation_steps 1
--lr_scheduler_type cosine
--logging_steps 1
--save_steps 100000
--learning_rate 1e-59
--num_train_epochs 1.0
--plot_loss
--fp16

It turned out that incorrect elements are clearly visible by a high “loss”, however, the elements from the dataset are not selected sequentially and it is impossible to make a correspondence between the dataset and the result.

In this regard, I have a question:
Is it possible to ensure that during training the elements are taken strictly sequentially from the beginning to the end of the file? Or at least so that “trainer_log.jsonl” includes information about the element number in the dataset.