gpt-adventure/ai_dungeon.py, branch main This is a clone of AI dungeon https://code.roseh.moe/gpt-adventure/atom/ai_dungeon.py?h=main 2022-02-16T05:24:37Z Rename to GPT adventure. 2022-02-16T05:24:37Z Rose Hogenson rhogenson@posteo.net 2022-02-16T05:24:37Z urn:sha1:169efa63904e3850a5bd698e4f62d870488305b9 Add a doctor scenario. 2022-02-15T18:13:21Z Rose Hogenson rhogenson@posteo.net 2022-02-15T18:13:21Z urn:sha1:66dfe3e616ba0d2c5d90afac4acc8e10ee736169 Use a better sentence detection strategy. 2022-02-15T18:12:40Z Rose Hogenson rhogenson@posteo.net 2022-02-15T18:12:40Z urn:sha1:60cd7954af55e91c97e90a1d8ae3e051c93dded5 It's actually pretty easy to see whether a snippet is a complete sentence, just append an extra word and see if it tokenizes into two sentences or just one. This should be a lot more reliable. Wrap the prologue. 2022-02-15T17:02:56Z Rose Hogenson rhogenson@posteo.net 2022-02-15T17:02:56Z urn:sha1:a4d6412fdaad7fae011d7103d4e40b01091b112d Sometimes the AI generates really long lines. Use gpt2 by default. 2022-02-15T17:02:32Z Rose Hogenson rhogenson@posteo.net 2022-02-15T17:02:32Z urn:sha1:37d8729645cfb2eefc92192b1d67b2476b81b3c4 I'm actually getting way better results from gpt2. Now gpt2 is my new best friend! Increase the temperature to 0.9. 2022-02-15T17:02:16Z Rose Hogenson rhogenson@posteo.net 2022-02-15T17:02:16Z urn:sha1:4817605a0c52a52dc1b9fcae6774887990eb8a3f Use real terminal width instead of assuming 72. 2022-02-15T15:13:19Z Rose Hogenson rhogenson@posteo.net 2022-02-15T15:13:19Z urn:sha1:08fef4b9ef771494bd2291edcfebf42b3d8ac5fd Auto-download nltk packages if they're missing. 2022-02-15T15:04:12Z Rose Hogenson rhogenson@posteo.net 2022-02-15T15:04:12Z urn:sha1:a906adb9ea2073ee9c23913585ff821558717ac8 Use an nltk tokenizer to trim sentence output. 2022-02-15T14:55:05Z Rose Hogenson rhogenson@posteo.net 2022-02-15T14:55:05Z urn:sha1:185851ec959d99c7c92309781bf5212c5433dc05 It kind of feels bad to depend on nltk and hugging face, but we're doing the best we can. Too bad that nltk isn't easier to install, too. Use a naive regexp to strip extra output. 2022-02-15T14:22:00Z Rose Hogenson rhogenson@posteo.net 2022-02-15T14:22:00Z urn:sha1:87cbdc4ec89fa8d54ab3123449c6397fe235f29c The sentence tokenizer wasn't correct because not all sentences end in periods. The current regexp doesn't handle punctuation inside words such as apostrophes or abbreviations, but we're trying it. Maybe a tokenizer would be better?