Mafiascum.net

Dataset | Paper

Summary: I made my computer read lots of games and now it is pretty good at scumhunting. The dataset (700 games, 10000 documents) is freely available for academic use.

Nice. Less frequent higher density posting = mafia.

Sounds like how I played as mafia. "Oh looks like I need to make a post, let's make it count"

In post 1, Awoo wrote:Nice. Less frequent higher density posting = mafia.

Sounds like how I played as mafia. "Oh looks like I need to make a post, let's make it count"

Trufax, I started doing that as town in my patented "copy my scumgame as town" way…because I found it FRIGHTENINGLY EFFECTIVE as scum.

So it used to be a scumtell of mine, until I copied it into my towngame. And my towngame DID IN FACT IMPROVE after.

Scum use it, but EVERYONE should use it, basically.

In post 2, mastina wrote:

In post 1, Awoo wrote:Nice. Less frequent higher density posting = mafia.

Sounds like how I played as mafia. "Oh looks like I need to make a post, let's make it count"

Trufax, I started doing that as town in my patented "copy my scumgame as town" way…because I found it FRIGHTENINGLY EFFECTIVE as scum.

So it used to be a scumtell of mine, until I copied it into my towngame. And my towngame DID IN FACT IMPROVE after.

Scum use it, but EVERYONE should use it, basically.

This warrants some explanation.

Obviously, if literally everyone did this, the game wouldn't work.

But if MOST people did this, regardless of alignment, the game would go much better.

To give a shining example of it working well for town, you need look only to the instance which won me my banner. If you check out my posting habits, they perfectly fit the "long periods of low activity, short periods of high activity", including the town jam session.

I said what I needed to, then left until I needed to say something.

So there's benefit regardless of alignment for the style.

For a start, it fits with my real life needs. I work ~32-36 hours a week at a high stress job, and also work on a day normally most don't, without a compensating day off most work. (That is, I work 6, rather than 5, days: every weekday like most, plus Sunday.)

The hours I work change day to day, so I have no consistent circadian rhythm, wrecking havoc on my body. They change week to week, month to month.

I need time off, where I do nothing but relax. I'd die if I didn't get it. Mafia is many things; relaxing is NEVER one of them. So I need time where I'm not playing and not working, preferably every day.

Which limits when I can be around.
This is compounded by other factors: mental health (as I am bipolar, I suffer depression episodes, where I need to take care of myself), physical health (current obstacle: pneumonia, on top of my prior diagnosis of Vitamin B12 deficiency), and similar.

The in-game benefits?
This method allows time for critical thinking, but to give bursts of spontaneity when necessary. You can take time to formulate ideas, and then, express them and expand upon them off of recent posts.

It allows you to not be in the spotlight when doing so is detrimental, allowing for objective views of those who ARE, but ALSO allows you to take the spotlight when you DO need it.

Scum find it easier to exploit this strategy, but it's equally valuable to both alignments.

BBob, this is fantastic.

I have recently been following a similar Japanese project (which is more focused on creating agents than processing deception), and I will definitely be taking a close look on your paper. Where are you submitting it?

Did you have any issues with the Ethics board at your institution when making this dataset?

I'm a fan of the "few long posts" style at least partly because I don't often have time for the game, so it's one of the only ways to stay active enough to be useful.

Like mastina, I've also noticed that intentionally imitating my scum meta as town can be successful. That shouldn't make any sense, but it's probably a reflection of players being more aware of their own meta than other people are of their meta.

In post 4, Claus wrote:BBob, this is fantastic.

I have recently been following a similar Japanese project (which is more focused on creating agents than processing deception), and I will definitely be taking a close look on your paper. Where are you submitting it?

Did you have any issues with the Ethics board at your institution when making this dataset?

I was worried about this myself, but my thesis supervisor said site administrator permission was sufficient in this case.

Fantastic work!

Very fascinating

I actually did something similar recently, on sentimental analysis with amazon reviews
We used tensorflow-hubs universal text encoder, to encode text into 512d vector, and we had 3 classes for classification: negative, neutral, and positive

Yeah, this is very cool stuff.

I'm also curious about where you're submitting it.

If I get this published somewhere interesting, I'll definitely post an update here.

Can you send me the AI so I can use it in my games?

haha such a understated thread title

i'm relieved that there's still a lot of further work to do