Feature request

Now DataCollatorForCompletionOnlyLM assumes that there is the same number of Questions and Answers and they go strictly one after another. I want this class to support all the variants of QAAQA, QAQQA and moreover support multiple instruction_template (e.g. "user" and "tool")
There are 2 problems with implementation: 1 arises when using tools or anything more then Question or Answer templates; 2 arises when your sequence is not strictly QAQAQA (e.g. QAAQA or QAQQA)

Motivation

Fine-tuning models on agents' trajectories is becoming more and more popular, yet there is no simple tokenizators even for well-known and industry-standard formats such as working with OpenAI's format (see here for Tool Calling)
Now simple things like QAAQA or QAQQA are not supported as well as any use of tools/funciton calls/etc.
It was also mentioned a few times already - #1994 and #2545 with a few people asking for updates

Your contribution

I have already changed the masking algorithm for local use, so I will try to make a PR for that today or this later week. I am not sure tho about the speed of that so it will need to be checked.
1st problem would only be fixed if we allow it to work with lists of different Qs (question starting templates) e.g. [<|im_start|>tool", "<|im_start|>user]
2nd problem would only be fixed if we change the masking algorithm (e.g. mask all and unmask between A&Q, instead of mask everything between Q&A)