Plans?

[wassname]

I'm interested in what the plans here? Are you combining existing open source datasets or making a new one?

If you are combining them, I think you need to make some choices (large or high quality, human generated vs AI generated). Have you got an aim there?

Here's a rough list of dataset's I put together. Perhaps it will help:

all huggingface data 1

• SHP reddit based. 2023-02-23 Stanford human preferences. Comes with some premade value models twet
• alpaca_data_cleaned.json including removing as a large language model https://github.com/gururise/AlpacaDataCleaned
• HH-RLHF - Antropic RLHF 91k
• ELI5: reddit based. SHP criticises it, saying that they compare score but not time, so you can't get preferences
• b-mc2 (https://huggingface.co/datasets/b-mc2/wikihow_lists).
• oasst- from open assistant, not released yet 22k examples (probob)
• https://github.com/allenai/natural-instructions 64k
• unnatural-instructions used above and GPT3 to make 256k examples
• https://github.com/hendrycks/ethics 130k

large and low quality

• 43M https://laion.ai/blog/oig-dataset/
  • note there is a high q subset OIG-small-chip2
• https://huggingface.co/datasets/HuggingFaceH4/stack-exchange-preferences 10M used in anthropic paper
• https://github.com/teknium1/GPTeacher
• https://github.com/databrickslabs/dolly