Add ShareChat Dataset

[raye22]

A new published human-LLM conversation dataset, collected from five different platforms: ChatGPT, Claude, Gemini, Grok, and Perplexity.

[mlabonne]

Thanks for the addition! I don't see how to access the conversations, however. I only see a "plain_text" column (but it's only partial) and the links in "url" don't work for me (e.g., https://chatgpt.com/share/67770737-9244-8006-9592-8e08ab2df3ec returns a 404 error).

Do you know how to get the conversations?

[raye22]

Thanks so much for the quick response!

To access the full conversations, you need to group the rows by the url column. Within each URL group, you can use the message index and role columns to sequence the interaction between the LLM and the user. And regarding the data quality points you raised:

1. I checked the empty rows for the plain_text column, and they currently account for less than 1% of the data for most platforms (exception for 2% for grok).
2. And for the broken urls, the scraping was performed a couple of months ago. Some links (like the one you tested) may now return a 404 error if the original user deleted or made the chat private in the meantime. However, I manually checked a random sample of the ChatGPT URLs, and about 70% are still active.

Given the unique multi-platform and longer interaction data included here, I believe this would be a valuable resource for the community and a great addition to your repository!