How was this dataset created?
#2
by
sudeepag
- opened
Hi @nampdn-ai , thank you for this dataset. I was just curious how this dataset is different from the Flan 2022 collection, and other ones of Huggingface like https://huggingface.co/datasets/SirNeural/flan_v2 and https://huggingface.co/datasets/Open-Orca/FLAN. Is this a sample of the other ones, and if so, could you please share some additional detail around how it was created?
Thanks!
sudeepag
changed discussion title from
How was this data created?
to How was this dataset created?
Hi, I think it's not different, just filtered version of the original https://huggingface.co/datasets/Open-Orca/FLAN for my tiny model. I build a fasttext classifier to choose the task that's helpful for long term human need and strip it out.
Thanks! A few more questions:
- Are you able to share any more details about the classifier? How was this trained?
- For the tasks that you did not filter out, did you sample from them uniformly? Would you be able to share some additional details about your sampling strategy if so?
- Use GPT-3.5/4 to classify few hundred samples in different categories like: bad, normal, good. Then use that to apply to entire dataset.
- No, I just apply fasttext and choose the good one.