sharingan-2 / README.md
markokraemer's picture
Card: exploiter result is a sanity check, not an exploitability measure
c97c473 verified
|
Raw History Blame Contribute Delete
5.17 kB
---
license: apache-2.0
library_name: custom
language:
- en
tags:
- poker
- game-theory
- cfr
- mccfr
- reinforcement-learning
- imperfect-information-games
- no-limit-holdem
- rust
pipeline_tag: reinforcement-learning
model-index:
- name: SHARINGAN-2
results: []
---
# SHARINGAN-2: the full-length Poker Sharingan blueprint
SHARINGAN-2 is blueprint C of the [Poker Sharingan](https://sharingan.poker) Pluribus reproduction (Brown & Sandholm, *Science* 2019) for 6-player
no-limit Texas hold'em, trained to the full **1.83e11 iterations**. It is the successor of
[SHARINGAN-1](https://huggingface.co/markokraemer/sharingan-1) (the same run at 2e10 iterations). Code, training and evaluation tooling:
https://github.com/markokraemer/poker-sharingan (Rust workspace, "library: custom": the weights are read by the Rust crates `blueprint` and `bench`).
The public arena's `sharingan-1` track keeps SHARINGAN-1 as its opponent for the duration of the current competition, so scores stay comparable.
## Results
All numbers: 6-max, duplicate deals + AIVAT, mbb/game (1 mbb/game = 0.1 bb/100), 95% CI, 4M deals per format.
| test | 3 + 3 seats | 1 SHARINGAN-2 vs 5 | 5 vs 1 (the other side's score) |
|---|---|---|---|
| **SHARINGAN-2 vs SHARINGAN-1 (C@2e10)** | **+5.09 +- 1.16** | +8.60 +- 1.89 | -7.64 +- 1.88 |
| SHARINGAN-2 vs C@4.77e10 (strongest earlier checkpoint) | **+1.97 +- 0.97** | +3.76 +- 1.57 | -1.95 +- 1.58 |
| test | result |
|---|---|
| trained best response (12M iterations, every seat) vs SHARINGAN-2 | the exploiter loses -247 +- 58 (SHARINGAN-1: -323 +- 62). A true best response cannot lose, so this run is far too short to measure exploitability: a sanity check that it finds no easy leak, not evidence of robustness |
| gauntlet, 6-max, vs nit / TAG / LAG / maniac / calling station / random | +125 +- 35 / +303 +- 56 / +933 +- 85 / +1,622 +- 99 / +2,572 +- 118 / +4,004 +- 197 |
It is a 6-max blueprint: heads-up (2-max) play is out of distribution (it loses to the simple built-in bots there).
## What is in this repository
| file | size | what it is |
|---|---|---|
| `sharingan-2.plcp.part-*` | 169.9 GB | **SHARINGAN-2**: blueprint C after 1.83e11 iterations (PLCP1 compact, 14.5B infosets, `medium` bet sizes) |
| `buckets-200.tar` | 0.28 GB | the 200-bucket card abstraction for postflop streets (same as SHARINGAN-1). Required |
| `SHA256SUMS.models`, `SHA256SUMS.parts` | | checksums of the reassembled file and of every part |
Files are split into 25 GB parts for upload. Reassemble and check:
```bash
pip install -U huggingface_hub
hf download markokraemer/sharingan-2 --local-dir weights
cd weights && sha256sum -c SHA256SUMS.parts
cat sharingan-2.plcp.part-* > sharingan-2.plcp && sha256sum -c SHA256SUMS.models && rm sharingan-2.plcp.part-*
mkdir -p ../data && tar xf buckets-200.tar -C ../data # -> data/buckets-200/
```
## Use it
From a clone of [the repo](https://github.com/markokraemer/poker-sharingan) (`RUSTFLAGS="-C target-cpu=native" cargo build --release`). The file is
memory-mapped: disk is required, RAM only makes it faster. Load it with the `medium` menu.
```bash
# inference server: POST /v1/act with a hand, get SHARINGAN-2's action (format: sharingan-1/serve/README.md)
./target/release/sharingan-serve --blueprint weights/sharingan-2.plcp --buckets data/buckets-200 --sizing medium --no-auth
# your own arena with SHARINGAN-2 as the house bot
ARENA_BLUEPRINT=$PWD/weights/sharingan-2.plcp ARENA_BUCKETS=tables:$PWD/data/buckets-200 ARENA_SIZING=medium ./target/release/sharingan-arena --port 8080
# the head-to-head above
./target/release/pluribus h2h --a weights/sharingan-2.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
```
## Model details
* **Game:** 6-max no-limit hold'em, 100 bb stacks.
* **Algorithm:** parallel external-sampling **Linear MCCFR** with negative-regret pruning; the paper's schedule scaled to iterations. Run C, seed 2019,
trained to 1.83e11 iterations (the paper's budget), on 64-core / 576 GB bare-metal servers, resumed across restarts from checkpoints.
* **Abstraction:** preflop lossless; 200 buckets per postflop street. Action menu `medium` (must match at load time).
* **Real-time search** is code, not weights (see SHARINGAN-1's card and the repo); the numbers above are blueprint against blueprint.
## Limitations and risks
* Not evaluated against strong humans; exploitability is only bounded from below by one trained best response.
* Trained for 100 bb, 6-max. Off-menu bet sizes are handled by action translation, the main exploitable surface of this class of agents.
* A near-equilibrium bot is a weaker exploiter of weak opponents than adaptive bots.
* **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
```bibtex
@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}
}
```