7 ms·
Would it be interesting to fine-tune GPT-3 on these?
by hntrader 6y ago
Would it be interesting to fine-tune GPT-3 on these?
- fumblebee 6y agoGPT-2 would suffice for something as non-sensical and rambling as dreams. No fine-tuning using the dataset, but here's the result of typing in a dream prompt into Talk-to-Transformer: [PROMPT - a dream from DreamBank] >>> I was putting together a plastic bag full of buttons and sewing paraphernalia. It was time to break out of a Hong Kong prison and I was on the fourth tier cell block. I had to wait till the window washers were busy and distract them by throwing ping pong balls so fire hoses filled with cyanide could be squirted on passers-by. I threw a rock through [...] [RESULT] >>> [...] a glass door. I was careful not to pull my stitches. I threw another rock out the window. We needed a diversion. I needed to take care of myself. The next thing I knew, I was in a helicopter about to take off. By this point my emotions were in flux, like people around a train wreck. It was hard to know what to do with them. I wanted to be strong but didn't have the will. I wanted to be a martyr. I wanted to be.
- allemagne 6y agoSometimes when I try to force the AI in AIDungeon to do something specific, the reaction is often weirdly reminiscent of the behavior of dreams. For example, I once kept trying to force an elf to teach me how to say certain phrases in their language (in an attempt to see if the AI could invent an imaginary language to some degree of consistency), and the character kept "laughing it off" and abruptly changing the subject.
- hntrader 6y agoI wonder if it's because an invented language has a gigantic number of permutations, the frequency of which swamps English-language replies. Then the most likely next word/statement that completes the input prompt is always going to be the latter due to the relative frequencies.