microsoft

/

DialoGPT-large

like 267

Follow

Microsoft 5.59k

Lol. If you don't like any of responses, just don't use the model or filter out unwanted responses.
P.S. No one owes you anything when it comes to using an open-source free model.

Yep this is a big issue, it gave me the exchange:

>> User: Are you racist?
DialoGPT: I'm not racist, but I am a racist.

oh my god. this thing is dizaster

What's up with the smooth brain comments in here?

Y'all realize this model was pre-trained on data collected from Reddit, right?

It was trying to be funny, I think. Also, there's 2 different ways you could interpret its attempt at humor, one of which is genuinely funny:

1. It's attempting sarcastic humor by facetiously implying a dislike for others along racial lines 
          but due to non-race-specific dislike of heterogeneity (less likely, less funny / not funny).
2. It's attempting sarcastic humor by facetiously implying a dislike for others (not along racial lines) due to
          dislike of heterogeneity (more likely, funnier).

Also bear in mind, this is a second-tier model in an area of application where the best-in-class models still make giant dumb errors frequently.