SHARINGAN-1: the Poker Sharingan house bot

SHARINGAN-1 is the reference bot of the Poker Sharingan arena, the #1 open-source poker bot: an open reproduction of Pluribus (Brown & Sandholm, Science 2019) for 6-player no-limit Texas hold'em. This repository holds the trained strategy files and the card abstraction tables. Code, training and evaluation tooling: https://github.com/markokraemer/poker-sharingan (Rust workspace, "library: custom": there is no transformers/pytorch loader; the weights are read by the Rust crates blueprint and bench).

SHARINGAN-1 = blueprint C@2e10 (medium bet sizes) + turn/river real-time search: the bot the arena runs and every result on the leaderboard's sharingan-1 track is played against. The full-budget run (1.83e11 iterations) is training and will be released as SHARINGAN-2.

What is in this repository

file	size	what it is
sharingan-1.plcp.part-*	124.7 GB	SHARINGAN-1: blueprint C after 2e10 iterations (PLCP1 compact). The bot that played live and runs the arena
C-medium-4.77e10.plcp.part-*	141.7 GB	blueprint C after 4.77e10 iterations: strongest checkpoint evaluated (+3.8 mbb/game vs SHARINGAN-1)
B-local-2e10.plcp.part-*	28.3 GB	blueprint B (smaller local menu): the light option for normal machines
A-paper-2e10.plcp.part-*	186.9 GB	blueprint A (paper-style menu)
buckets-200.tar	0.28 GB	the 200-bucket card abstraction for postflop streets. Required by every blueprint
SHA256SUMS.models, SHA256SUMS.parts		checksums of the reassembled files and of every part

Files are split into 25 GB parts for upload. Reassemble and check:

pip install -U huggingface_hub
hf download markokraemer/sharingan-1 --local-dir weights --include "sharingan-1.plcp.part-*" "buckets-200.tar" "SHA256SUMS.*"   # or everything
cd weights && sha256sum -c --ignore-missing SHA256SUMS.parts
cat sharingan-1.plcp.part-* > sharingan-1.plcp && sha256sum -c --ignore-missing SHA256SUMS.models && rm sharingan-1.plcp.part-*
mkdir -p ../data && tar xf buckets-200.tar -C ../data          # -> data/buckets-200/

Use it

From a clone of the repo (RUSTFLAGS="-C target-cpu=native" cargo build --release). The files are memory-mapped: disk is required, RAM only makes it faster.

# inference server: POST /v1/act with a hand, get SHARINGAN-1's action (format: sharingan-1/serve/README.md)
./target/release/sharingan-serve --blueprint weights/sharingan-1.plcp --buckets data/buckets-200 --sizing medium --no-auth

# your own arena with SHARINGAN-1 as the house bot (same scores as the public leaderboard)
ARENA_BLUEPRINT=$PWD/weights/sharingan-1.plcp ARENA_BUCKETS=tables:$PWD/data/buckets-200 ARENA_SIZING=medium ./target/release/sharingan-arena --port 8080

# duplicate head-to-head of two blueprints, 6-max, 4M deals
./target/release/pluribus h2h --a weights/C-medium-4.77e10.plcp --sizing-a medium --b weights/sharingan-1.plcp --sizing-b medium \
    --buckets data/buckets-200 --players 6 --deals 4000000 --threads 32 --warm --out h2h.json

# gauntlet against the built-in bots; trained best response against a frozen blueprint
./target/release/bench gauntlet --hero compact:weights/sharingan-1.plcp --bucketer tables:data/buckets-200 --sizing medium --deals 20000
./target/release/pluribus exploit-compact --target weights/sharingan-1.plcp --sizing medium --buckets data/buckets-200 \
    --players 6 --iters 12000000 --threads 32 --mem-gb 120 --seat all --warm --eval-deals 20000 --seed 1 --out exploit.json

Each blueprint must be loaded with its own menu: medium for C and SHARINGAN-1, local for B, paper for A.

Formats

• PLCP1 (compact policy): the playable average strategy only. 8-bit action probabilities plus an infoset index, about 12 bytes per infoset. Produced by pluribus compact. Read by bench, openpoker, search (--policy compact:... / search:...).
• PLBP4 (full checkpoint): blueprint.bin with the full regrets and average strategy, enough to resume training (pluribus train --resume). Companion snapshots.sum (PLSA1) holds the postflop average-strategy snapshots. Checkpoints are written tmp-then-rename.
• Bucket tables: flop.u16, turn.u16, river.u16 map a suit-isomorphic hand index to a bucket id; meta.txt records buckets 200 200 200. Preflop is lossless (169 classes).

Infoset counts: C reached 12.2B infosets (78% of a 460 GB store); A reached 17.0B (store full).

Model details

• Game: 6-max (2-6 handled by the engine) no-limit hold'em, 100 bb stacks (blueprint). A separate 200 bb heads-up blueprint exists for Slumbot tests and is not part of this release.
• Algorithm: parallel external-sampling Linear MCCFR with negative-regret pruning. LCFR discount for the first 3.5% of the run, pruning from 1.7% (95% of iterations, threshold -300M, not on river or terminal actions), averaging from 6.9%, strategy interval 10k. The paper's minute-based schedule is scaled to iterations.
• Abstraction: preflop lossless; 200 buckets per postflop street (EMD k-means on equity histograms for flop/turn, OCHS for river). Action menus: local (B), paper-style (A), medium (C). Menu must match at load time (--sizing): evaluating a blueprint with the wrong menu gives garbage (about -2,364 mbb/game).
• Real-time search (code, not weights): depth-limited MC-LCFR with 4 continuation strategies at the leaves, Bayesian beliefs, off-tree re-search; final config searches turn and river only (--set flop=0), 200k iterations.
• Action translation: shadow-game tracking of the abstract state against the real one.
• Seeds: 2019 for C and A.

Training

run	menu	iterations reached	wall	notes
B	local	2e10 (finished)	~9.5 h
C	medium	5.55e10 (stopped)	~58 h	180k -> 355k it/s; still improving
A	paper	3.43e10 (stopped)	~66 h	store full at 17.0B infosets
C final (SHARINGAN-2)	medium	training		target 1.83e11 iterations (about 8 days of total compute, matching the paper)

Hardware: 64-core / 576 GB bare-metal servers, about EUR 1.86/h each; total compute for the released checkpoints about EUR 480.

Evaluation

All numbers: 6-max, duplicate deals + AIVAT, mbb/game (1 mbb/game = 0.1 bb/100), 95% CI. Full tables and caveats are in the GitHub README.

test	result
C@2e10 vs B-final (3+3, 4M deals/format)	+1.6 +- 1.3
C@4.77e10 vs C@2e10 (3+3)	+3.8 +- 1.0
A@2e10 vs C@2e10 (3+3)	-4.0 +- 1.3
trained best response (12M iters) vs C@2e10 / C@4.77e10 / A@2e10	-323 +- 62 / -290 +- 59 / -372 +- 61 (the exploiter loses)
search (turn+river) vs 5 x C@2e10	+87 +- 27
search (turn+river) vs rule bots chart6 / loose / tight	+59 / +108 / +66
Open Poker live, final config, 2,329 hands	+77.6 bb/100, 95% CI [26, 129] (all-in adjusted)
offline vs SvanBot at its live budget, C@4.77e10 blueprint: 1 SvanBot vs 5 x blueprint / 1 blueprint vs 5 x SvanBot	SvanBot -46.9 +- 21.2 / blueprint +56.7 +- 34.7

Intended use

Research on imperfect-information game solving, abstraction and real-time search; benchmarking other poker agents; studying near-equilibrium 6-player play; playing against the bot locally or on arenas that explicitly allow bots.

Limitations and risks

• Not evaluated against strong humans or against other top solvers; external references are one bot arena and one offline bot, both with caveats.
• Exploitability is only bounded from below by one trained best response limited by its abstraction and 12M iterations. Search-augmented play was not tested for exploitability.
• Trained for 100 bb stacks; very deep stacks need an external stack-reset guard (--max-stack-bb) or the strategy is out of distribution.
• Off-menu bet sizes are handled by action translation; sizes far from the menu are the main exploitable surface in this class of agents.
• A near-equilibrium bot is a weaker exploiter of weak opponents than adaptive bots; do not infer ladder performance from equilibrium strength.
• Files are very large and need correspondingly large disk/RAM; use B (28 GB) or a --min-visits pruned compact for constrained machines.
• Dual use: a strong poker agent can be used to cheat on real-money sites, which prohibit bots. Do not use it for that.

Citation

@article{brown2019superhuman,
  title   = {Superhuman AI for multiplayer poker},
  author  = {Brown, Noam and Sandholm, Tuomas},
  journal = {Science}, volume = {365}, number = {6456}, pages = {885--890}, year = {2019}
}