50
$\begingroup$

Has ChatGPT used highly rated and upvoted questions/answers from Stack Overflow in its training data?

For me it makes complete sense to take answers that have upwards of 100 upvotes and include them in your training data, but people around me seem to think this hypothesis doesn't make sense. Is there a way to confirm this?

$\endgroup$
6
  • 3
    $\begingroup$ It's worth noting that even with the option you suggest, the answers produced by ChatGPT wouldn't be more correct. Maybe they would be slightly better written, but this is not the issue. $\endgroup$
    – Erwan
    Commented Jan 10, 2023 at 16:30
  • 6
    $\begingroup$ Think about the attrition this inbreeding will cause to the quality of SO answers.. People ask ChatGPT for solutions which come in part from SO, then they post the answers to SO, ChatGPT learns from these answers but doesn't get them perfect, so the next time someone asks ChatGPT for a similar answer, the result is a little more incorrect, and is then again posted to SO.. $\endgroup$
    – tplive
    Commented Jan 11, 2023 at 9:56
  • $\begingroup$ @tplive. That's exactly the reason SO has a feedback system. In my experience, the answers supposedly knowledgeable people post are often no walk in the park. $\endgroup$ Commented Jan 11, 2023 at 18:28
  • $\begingroup$ It certainly didn't learn from stackexchange to apologize for its mistakes. $\endgroup$
    – user11235
    Commented Jan 24, 2023 at 9:40
  • 1
    $\begingroup$ It would possibly make more sense to use Accepted answers rather than the most upvoted ones. Numbers of upvotes - especially on Stack Overflow (as opposed to other SE sites) - are more reflective of the number of people who wanted to know the answer than the quality of the answer. There are some very good answers with few upvotes simply because it's not a common problem. $\endgroup$
    – komodosp
    Commented Mar 21, 2023 at 12:22

3 Answers 3

59
$\begingroup$

ChatGPT is in the Large Language Models (LLM) category. The most (in)famous GPT model is probably GPT-3, because since then, researchers realized that LLMs mostly follow a predictable scaling law, thus the more data and the bigger model, the better. It is accurate to say that ChatGPT was trained with Stack Overflow data, but it should be all Stack Overflow instead of just most upvoted answers/comments.

The Wikipedia page of GPT-3 and their paper mentions that GPT-3 was trained on multiple datasets, and one of which is the Common Crawl, which basically crawls everything on the Internet. Some data pre-processing was done before training, but the authors did not mention removing the comments, so we can say that it is all Stack Overflow data.

If we look at the Common Crawl data in Sep 2022, there is indeed the domain com.stackoverflow in their list. Thus, while ChatGPT was trained on Stack Overflow data, it is trained on all Stack Overflow data instead of just most upvoted answers.

However, if you think ChatGPT's code output is of high quality, think again, because Stack Overflow temporarily bans ChatGPT because the average rate of getting correct answers from ChatGPT is too low, the posting of answers created by ChatGPT is substantially harmful to the site and to users who are asking and looking for correct answers (quoted from the link).

Here is the justification of them:

The primary problem is that while the answers which ChatGPT produces have a high rate of being incorrect, they typically look like they might be good and the answers are very easy to produce. There are also many people trying out ChatGPT to create answers, without the expertise or willingness to verify that the answer is correct prior to posting. Because such answers are so easy to produce, a large number of people are posting a lot of answers. The volume of these answers (thousands) and the fact that the answers often require a detailed read by someone with at least some subject matter expertise in order to determine that the answer is actually bad has effectively swamped our volunteer-based quality curation infrastructure.

EDIT: a comment below also confirmed that GPT-2, the predecessor of GPT-3 and ChatGPT, was trained with Stack Overflow data.

$\endgroup$
6
  • 3
    $\begingroup$ How do you know it was trained on all data? The crawler might have accessed all data. But how do we know this data hasn't been preprocessed? E.g. remove comments etc. $\endgroup$ Commented Jan 9, 2023 at 14:32
  • 4
    $\begingroup$ @ThomasWeller See Appendix A of the paper. I don't think they removed comments. $\endgroup$
    – wizzwizz4
    Commented Jan 9, 2023 at 17:02
  • 1
    $\begingroup$ Note that because the average rate of getting correct answers from ChatGPT is too low is not sourced in the link you've provided. $\endgroup$ Commented Jan 10, 2023 at 0:39
  • 1
    $\begingroup$ I would argue that (at least in the general population), ChatGPT is now the most famous GPT model, by quite a large margin. $\endgroup$
    – Arthur
    Commented Jan 10, 2023 at 9:55
  • 4
    $\begingroup$ It is also known that GPT2 was trained on Stack Exchange, and Stack Overflow. $\endgroup$
    – David
    Commented Jan 11, 2023 at 1:39
2
$\begingroup$

Technically speaking, we don't know whether ChatGPT used Stack Overflow data for the training. That is because ChatGPT is a proprietary model, and OpenAI hasn't published training details to the public.

We know that, from the GPT-3 paper [1], the GPT-3 training dataset includes Common Crawl. However, there is no indication of the relation between GPT-3.5 series (including ChatGPT) and GPT-3 models according to OpenAI's account.

See also this question Are GPT-3.5 series models based on GPT-3? for the discussion on GPT-3 vs. GPT-3.5. I add: the token set is different between davinci (GPT-3) and text-davinci-003 (GPT-3.5) (source) thus the later cannot be a simple fine-tuning of the former.

I think, though, it is natural to infer OpenAI uses StackOverflow data because it is generally considered high quality [2] and easily obtainable. Also, if OpenAI ever uses web crawling data, it will see many StackOverflow and its clone sites.

$\endgroup$
0
$\begingroup$

it was trained using supervised from open ai team for initializing the conversational model and for more customize style, it can highly possible using chosen stackoverflow sample dataset since the answer of the question from the website was moderated. ChatGPT was take the advantage of GPT-3 model that has been trained from various corpus of data and can perform few-shot learning.

$\endgroup$
1
  • 2
    $\begingroup$ As it’s currently written, your answer is unclear. Please edit to add additional details that will help others understand how this addresses the question asked. You can find more information on how to write good answers in the help center. $\endgroup$
    – Community Bot
    Commented Apr 12, 2023 at 15:32

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .