Datasets:
Almost all examples in the dataset viewer preview are explicit
As stated in the title, almost all of the examples in the dataset viewer are sexually explicit. This is also the case for the larger 6T token dataset. Some example IDs (these are just the first few examples which appear in the dataset viewer for me):
- fd27c4f3-b13c-488e-959a-5741223f5985
- 9aadd536-3c1e-4dca-96d6-2e40c02bfa01
- b247fe4e-aacc-41b5-bfdd-dd43c9f38770
- 8788f162-41a4-486d-9407-5a00434ae25e
This is a very small sample, in fact almost 100% of the examples shown in the first page of the data viewer are explicit in nature. I would have assumed that such examples would be deliberately filtered out, and I think this raises some concerns about the overall dataset quality. Am I missing something here?
On closer inspection it seems that these examples are all accurately tagged as adult content in the metadata and this topic is mentioned in the Olmo 3 paper. I apologize, I think I jumped the gun in assuming something was wrong. Nevertheless, I would suggest that at least the dataset sample be changed or organized differently. The first 39 pages (at least---I couldn't load more on my browser) of the data viewer are nothing but adult content examples, which is probably not the best advertising for an otherwise great and interesting dataset 😂
If only "adult" didn't appear so early in the alphabet ... 🙈
Thanks for the feedback @zleizzo ! We tried to specify a different dataset to be displayed in all of our datasets, but that unfortunately causes issue with HF's load_dataset. The alphabetical ordering in the data viewer is truly unfortunate, but we are definitely aware and are considering a fix for this in future releases!