Dominion Strategy Forum

Please login or register.

Login with username, password and session length
Pages: [1]

Author Topic: GPT-3 Generated Cards  (Read 815 times)

0 Members and 1 Guest are viewing this topic.

silverspawn

  • Adventurer
  • ******
  • Online Online
  • Posts: 5001
  • Shuffle iT Username: sty.silver
  • Respect: +2674
    • View Profile
GPT-3 Generated Cards
« on: July 21, 2020, 02:35:05 pm »
+8

(GPT-3 is OpenAI's gigantic 175 billion parameter language model. It's a neural network that predicts the next word in a sequence of previous words.)

I once posted a bunch of GPT-2 generated cards in the Really Bad Card Ideas thread. But those were (a) generated with GPT-2 (which is much smaller than GPT-3) and (b) I only had access through a small prompt that barely had enough space to show it two current cards. The proper way would have been to train it on a large corpus of cards, but I couldn't do that.

Now, on the other hand, I'm using AI Dungeon. I can show it much more cards, and the results are far more impressive than they used to be. I structrured the entire thing as a conversation between an "Interviewer" and "Donald X." The Interviewer asks questions about Dominion, Donald X answers, he shows some of the existing cards, and then some new cards.

This way, I can also request particular types of cards. I'll use this thread to catalog the good ones.


Victory Cards

Yards Victory 6$
Worth 2VP for every 10 cards in your deck (round down).

Stronghold Victory 10$
Worth 3VP for every 10 cards in your deck (round down).

County Victory 6$
Worth 2VP for every 3 Silvers in your deck (round down).

Highlands Victory 5$
Worth 1VP for every 5 Treasure cards in your deck (round down).


Attack Cards

Inquisition Action Attack 5$
Each other player reveals their hand and trashes a revealed Action card costing up to 4$. You gain the trashed cards.

Begging Action Attack 2$
Gain 1VP for each Treasure card in your hand. Each other player reveals their hand and trashes a revealed Treasure card costing up to 4$. You gain the trashed cards.

Mutiny Action Attack 4$
Each other player reveals their hand and trashes a revealed Attack card costing up to 5$. You gain the trashed cards.


Villages

Swamp Village Action 3$
+1 Card
+2 Actions
+1 Buy
Trash this to gain a card costing up to 4$.

Sea Village Action 3$
+1 Card
+2 Actions
You may reveal a Treasure card from your hand. If you do, gain a Silver.

Papermaker's Village Action 2$
+1 Buy
+2 Actions
You may reveal a card with Papermaker in its name from your hand. If you do, gain a Silver.

Smelting Village Action 2$
+1 Buy
+2 Actions
Trash a card from your hand. If you do, gain a Coin token.

Silk Dyer's Village Action 3$
+1 Buy
+2 Actions
Discard any number of cards. +1 Card per card discarded.

Mobilized Village Action 5$
+2 Actions
If you have two or more Villages in play, +4 Actions.


Others

Equipment Stash Action 2$
+2 Actions
Look through your discard pile and put a card from it into your hand. If you do, put Equipment Stash on top of your deck.

Bag Lady Action 3$
Trash this or a Treasure card from your hand. Gain a card costing up to 2$ more than the trashed card.

Miscellaneous

Royal Escort Action Reaction 5$
When you would lose the game, you may reveal this from your hand to ignore that loss.

Marauder Action Attack 2$
Trash this card and another card from your hand. Gain a card costing up to 2 Coins less than the trashed card.

(Lol'd because this is probably a better card than ErrinF's card that trashed itself and another card and had potion in its cost.)

Tornado Action Reaction 6$
Trash this and all the cards in your hand and gain a Gold, 2 Provinces and a Curse.

Feast Action 4$
You may eat a food this turn. Draw until you have 5 cards in hand. You may eat another food.

Royal Blacksmith Action 5$
Trash this card. Gain a card costing up to $4.

King's Court Action 10$
Gain a Prize card.

Wood Village Action 3$
+1 Card
Gain another Village.

Lumbermill Action 4$
+1 Card
+2 Actions
You may trash this to gain a Wood token.

Salted Pork Village Action 2$
+1 Buy
You may discard any number of cards. Gain a Food token for each card you just discarded.

(What's with the food tokens  :o maybe an import from MTG.)

Bakery Village Action 4$
+1 Buy
+2 Actions
You may reveal a Bread or Pie token from your hand. If you do, gain a Baking Token.



Outtakes

Quote
After a few minutes of swooning and praising Donald, you snap out of it just in time to hear him (and the rest of the crowd) yell "SOYLENT GREEN IS PEOPLE!" Whatever that means
« Last Edit: July 28, 2020, 08:12:25 am by silverspawn »
Logged

silverspawn

  • Adventurer
  • ******
  • Online Online
  • Posts: 5001
  • Shuffle iT Username: sty.silver
  • Respect: +2674
    • View Profile
Re: GPT-3 Generated Cards
« Reply #1 on: July 22, 2020, 07:21:22 am »
0

It's clear that GPT-3 knows about dominion. It keeps suggesting cards with names of existing cards (turning up the randomness factor helped with that), and it knew about trashing and about ruins without me prompting them. I also don't think it would have gotten the mechanics right so reliably just based on my examples.

This makes me wonder whether more training data will even help... if it knows the cards anyway, it might not make a difference.
Logged

grep

  • Apprentice
  • *****
  • Offline Offline
  • Posts: 260
  • Respect: +356
    • View Profile
Re: GPT-3 Generated Cards
« Reply #2 on: July 22, 2020, 11:34:14 am »
0

I've once tried to train GPT-2 on the Dominion corpus, but achieved overfitting before perfection - probably because there aren't many cards.

http://forum.dominionstrategy.com/index.php?topic=20397.0
Logged

silverspawn

  • Adventurer
  • ******
  • Online Online
  • Posts: 5001
  • Shuffle iT Username: sty.silver
  • Respect: +2674
    • View Profile
Re: GPT-3 Generated Cards
« Reply #3 on: July 22, 2020, 01:10:40 pm »
0

I've once tried to train GPT-2 on the Dominion corpus, but achieved overfitting before perfection - probably because there aren't many cards.

http://forum.dominionstrategy.com/index.php?topic=20397.0

That's pretty neat.

Yeah, your model clearly tried much harder to imitate existing cards. Mine tends to like names of existing cards but the effect is usually unrelated to that of the original card.
Logged
Pages: [1]
 

Page created in 0.063 seconds with 21 queries.