Feature/dataset entry #2520
Feature/dataset entry #2520
Conversation
…e/dialogue-data-collator-tests
CloseChoice
requested review from
theblackcat102,
sanagno,
dvruette,
andreaskoepf and
yk
as code owners
model/model_training/tests/resources/data_collator/tokenizer.json
Outdated
Show resolved
Hide resolved
| if mode == Mode.sft: | ||
| return qa_list | ||
| elif mode == Mode.rm: | ||
| raise NotImplementedError("This is currently not implemented.") |
There was a problem hiding this comment.
Does this mean that RM training will not work after this is merged?
There was a problem hiding this comment.
Yep, that is correct. Currently it is only implemented for Dolly, which does not support rm, so no change for now. But this needs to be implemented soon.
| ] | ||
| if len(relevant_system_infos) > 0: | ||
| shuffle(relevant_system_infos) | ||
| system_tag_key_values = "\n".join([f"{k}: {v}" for k, v in relevant_system_infos]) |
There was a problem hiding this comment.
Should the values be quantized here or are they already within a certain set? (I think real values could be difficult to understand for the model)
There was a problem hiding this comment.
The keys here are one of context, lang, length, quality, humor and creativity. The values for quality, humor and creativity can be floating point numbers between 0 and 1 (this is checked by the validators and a test for this is added aswell, see here).
There was a problem hiding this comment.
yes, so if they can be anything between 0 and 1 (including e.g. 0.1826357216), then I think we should round them (e.g. to a single decimal) to make it easier for the model to understand.
closes #2708
Add pydantic basemodel class (equivalent to dataclass but with stronger guarantees) to return from dolly dataset. Add the formatting functionality in the dataset entry class.
This PR does quite a bit:
QA_SPECIAL_TOKENSand theeos_token. This class should work as a general pattern to store and format single dataset entries. This class should also remove all the formatting errors we had previously with the datasets.DialogueDataCollator. I trained a minimal tokenizer only on the tokens that are present in the tests to not bloat the code (still a lot of LOC).DialogueDataCollatorand the newly introducedDatasetEntryclass.DatasetEntryand old rows from the dataset.todos: