OPT: Open Pre-trained Transformer Language Models

Re:OPT: Open Pre-trained Transformer Language Models by yosito on Monday May 02, 2022 @11:51PM (#31244370)

> OPT-175B has a high propensity to generate toxic language and reinforce harmful stereotypes

So they trained it on Facebook comments?

[ Parent | on HN ]

Re:OPT: Open Pre-trained Transformer Language Models by Gigachad on Tuesday May 03, 2022 @12:56AM (#31244797)

I'd think any natural language model would have the same biases we see from real humans.

[ Parent | on HN ]

Re:OPT: Open Pre-trained Transformer Language Models by tsol on Tuesday May 03, 2022 @01:36AM (#31245020)

Are there really no moderated forums that the data can be taken from? Even HN-based training data would be much more civil

[ Parent | on HN ]

Re:OPT: Open Pre-trained Transformer Language Models by Gigachad on Tuesday May 03, 2022 @01:55AM (#31245123)

A model trained on HN would spit out a 5 paragraph story about how minorities provide a negative ROI for cities. Or how the homeless need to removed from society.

[ Parent | on HN ]

Re:OPT: Open Pre-trained Transformer Language Models by can16358p on Tuesday May 03, 2022 @02:56AM (#31245479)

Don't forget that it must also generate, at some point regardless of the topic, a new terminal emulator, and an extremely positive or extremely negative opinion about how blockchain can solve a problem.

[ Parent | on HN ]

Re:OPT: Open Pre-trained Transformer Language Models by IAmEveryone on Tuesday May 03, 2022 @02:52AM (#31245458)

Sure, but it would never do something actually bad, like raising the possibility that sexual harassment might, sometimes, be an issue, or questioning the value of phrenology.

[ Parent | on HN ]

Re:OPT: Open Pre-trained Transformer Language Models by sanxiyn on Tuesday May 03, 2022 @01:41AM (#31245039)

Note that HN is included in the training data, see page 20.

[ Parent | on HN ]

Re:OPT: Open Pre-trained Transformer Language Models by IAmEveryone on Tuesday May 03, 2022 @02:53AM (#31245464)

Go figure (8)!

[ Parent | on HN ]

Re:OPT: Open Pre-trained Transformer Language Models by yosito on Tuesday May 03, 2022 @11:45AM (#31249714)

I'd think the training data is something that could be curated. Eliminating all bias might be impossible, but GIGO applies.

[ Parent | on HN ]

Re:OPT: Open Pre-trained Transformer Language Models by stephenroller on Tuesday May 03, 2022 @06:41PM (#31254508)

We trained on Reddit comments and HackerNews comments.

[ Parent | on HN ]

Re:OPT: Open Pre-trained Transformer Language Models by Rebelgecko on Tuesday May 03, 2022 @01:52AM (#31245106)

I thought Pushshift was only reddit comments?

[ Parent | on HN ]

Re:OPT: Open Pre-trained Transformer Language Models by MengerSponge on Tuesday May 03, 2022 @12:28AM (#31244624)

Does it merely reinforce harmful stereotypes? Or will it help perpetrate genocide?

[ Parent | on HN ]

Re:OPT: Open Pre-trained Transformer Language Models by rhizome on Tuesday May 03, 2022 @01:40AM (#31245033)

Tomato, tomahto.

[ Parent | on HN ]