I don´t know if anyone is interested but here it is:

Pluribus

Shout out to Kevin Wang who originally converted the hands for PT4.

Cheers.

I trust you, but just in case: when i tried downloading all the txt files as a zip, chrome itsself warned me that there is smt fishy with the files. Prob due to the fact that its all txt, but still felt weird and i canceled the dl.

anyways, when i just open one txt file itsself its normal

Very interested, but suspicious. Thanks for posting!

I don't know what a Pluribus looks like, but the download links to safe .zip and .txt files.

No virus detected. Just an EV of -1.72bb/100 over this sample. #Variance

I thought ai is taking over and pluribus crushed?

unfunny joke

They used some really sophisticated method of adjusting for variance far beyond all-in adj winnings

Wow, that´s news to me. I downloaded the ZIP file from Kevin Wang´s website and I changed all the text files one by one so I could upload them on HEM. Then I uploaded the text files to that website, so I really don´t know how it gave you that warning, sorry. Maybe download the text files only?

Seat 1: Gogo (€10000 in chips)
Seat 2: Budd (€10000 in chips)
Seat 3: Eddie (€10000 in chips)
Seat 4: Bill (€10000 in chips)
Seat 5: Pluribus (€10000 in chips)
Seat 6: MrWhite (€10000 in chips)

can someone match players to these names?

EDIT: realized theres more names here...

NoamBrown one of the creators of Pluribus answered a few questions on Reddit:

1 - A first look at the released hands shows the bot is losing at 7bb/100, offcourse this is to small sample to draw conclusions. You are using AIVAT for reducing the variance factor and then you get a positive winrate, can u explain in an easy way how this works and if other players ranges can affect this model? Also what was the real result in the 1H5B-challenge?

2 - How was the pros for this chosen? I see HU-players and a couple of tournament-players, most of the players on the list it's facing seems play mostly another format than 6-max NLH.

3 - The bot seems to never bet less than 1/2 pot on the flop after openraising and getting one caller, is this down to the "rules" you set for the bot? Highstakes players and solvers seems to prefer 1/3 or 1/4 pot on the flop in many of those situations and I'd expect the bot to arrive to similar conclusions.

1 - "First, I think it’s important for the non-poker folks to understand just how absurdly high the variance is in poker. We estimate the bot’s win rate to be 5 bb/100, which means the bot wins an average of about $5 per hand (at $50/$100 blinds with $10,000 stacks). That’s considered a high win rate, especially against this group of pros. But the standard deviation for an individual hand without variance reduction is about $1,000. Any half-decent player can make money over 10,000 hands of poker, and it’s normal for the best player in the world to lose money over 10,000 hands. (Indeed, Linus, considered by many to be the best human pro in the world at this form of poker, was down in chips in this experiment over the 10,000-hand sample.) Without variance reduction, it would have taken the pros 4 months of playing 8 hours a day, 5 days a week, to reach a meaningful sample size. Fortunately, some folks over at University of Alberta and Charles University of Prague previously developed a variance-reduction algorithm for poker called AIVAT that is provably unbiased (regardless of the other players’ ranges). We made it 100% clear to all participants before play began that we would only be evaluating the bot based on AIVAT. This ended up reducing the number of hands we needed by about 12.5x.
AIVAT is difficult to explain in a paragraph, but I can give some examples of how it works. First, if two players are all-in before all the cards are dealt, you can take the expected value over all the rollouts of the cards rather than dealing out one set of board cards. This is already a well-known and accepted form of variance reduction in the poker community, and you can see in the logs that Pluribus was very unlucky in these early all-in situations. Second, if a player is faced with an all-in bet on the river and is 50/50 between calling and folding, they could take the expected value of both actions rather than flipping a coin. Third, let’s say the bot is dealt AA and the other players are dealt weaker hands. We’d expect the bot to win money on this hand due to its lucky cards. We can reduce variance by subtracting an estimate of what we think each player should earn in this hand given all the players’ cards. This is estimated by seeing what the outcome would be if the bot played against itself in all six seats, which since it’s the same bot necessarily has zero EV. Fourth, the bot can look at its entire range, rather than the individual hand it was dealt, when evaluating its score. There’s more to AIVAT than just what I described (all details are in the paper), but that gives you a picture of how it works."

2 - "All the participants in the 5H+1AI experiment were recommended to us by other top poker pros. Some are better in tournaments or HU, but all are still considered very strong players in 6-max NLH."

3 - "Small pots on the flop are the most expensive to compute a strategy for and are also the least important, so we reduce the number of sizes the bot is allowed to choose from when betting in those situations. 1/2 pot is the smallest size we allowed it to consider betting in that kind of situation. It would probably do better if it had a 1/4 pot option, but I don’t think it makes a huge difference. It always precisely understands each opponent bet though, regardless of the size."

What is Linus nickname in the database? If you know he is down chips.

Yes agree, weird selection of players for 6max cash simulation.

Looking at the hands. It's tough to beat anything, which is playing so solid poker. Pluri plays very tight and balances the ranges postflop very well.

In a bunch of videos I have seen abt the match people claimed Mr pink would be Linus. But I don't think there is any proof. Also how do you know that he is down chips, thought he won?

sounds like bs especially over 10k hands which is nothing with smaller winrates

Thanks for doing the conversion for HEM I am chronically lazy and so I never went back to updating the code after putting it out there.

The adjustment described in the paper is not.convincing, but they don't give details.

original link is down, does anyone have an active link with the hands?

This.

meaning this is the equivalent of 125k hands?

makes it a bit more legit if it is as claimed and yet still highly inconclusive.

the 70% confidence interval with an 80stdv (not sure if that's accurate but probably not too far off) is +/- 2.5bb/100.
for 50stdv it's +/- 1.5bb/100.

and the 95% confidence interval for 80/50 is +/-4.5 and +/-3 roughly.

what that means is that 70% of the time someone plays 125k hands (with standard deviation of 80) their results will be +/-2.5bb/100 from their true winrate. 30% of the time it will be even further away than that. in isolation you might think this to be suggestive that pluribus is a loser (though far from conclusive). with so many other datasets of these types of ai crushing highly competent players it seems more likely this is a fairly common outlier dataset.

"The files in this link are deleted"
Is there a new link?

I always thought Eddie was linus in those hands

Pluribus wouldve known Tony G looked at his cards.