Dump raw training data for the LLM-jp-3 series

[hkiyomaru]

Dump raw training data for the LLM-jp-3 series. For each training instance, the following fields should be included at least:

• token_ids: A list of token IDs for the training instance
• training_step: Training step at which the training instance was processed
• dataset: Name of the dataset from which the instance was sourced
• document_ids: IDs of the documents associated with the training instance

[hkiyomaru]

https://github.com/llm-jp/Megatron-LM/tree/nii-geniac-dump