Logical conflict between data loading and collation.

[wenqibiao]

In _load_ultrachat_conversations, the messages are constructed using only the user role:
msgs = [{"role": "user", "content": prompt}]

However, LanguageDataCollator.call implements a mandatory check that skips any sample missing an assistant turn:

if not any(m.get("role") == "assistant" for m in messages):
    continue

So, this causes all samples to be skipped during training because no assistant responses exist in the pre-processed data. Is that right?

[juhi10071998]

Hi @wenqibiao, thanks for bringing this up. Could you point to the file location where this is happening?

[wenqibiao]

Hi @wenqibiao, thanks for bringing this up. Could you point to the file location where this is happening?

ok. File location:

git branch: main
git commit: 010b220dc09890bf7646f8fad5b69fd0bc41ac5b
1) file A (_load_ultrachat_conversations): examples/dataset/make_dataset.py, line 282.
2) file B (LanguageDataCollator.__call__): modelopt/torch/utils/plugins/transformers_dataset.py, line 288.