01 · GAME
HUNL rules and state tree
Cards, streets, blinds, terminal utilities, hand evaluation and an action abstraction from fold to all-in.
Imperfect-information game research
Build, train and study HUNL poker AI.
A source-verified research stack for CFR subgame solving, ReBeL-style recursive solving, counterfactual value learning, self-play and C++ Monte Carlo components.
HUNLTwo-player no-limit engine
CFRRegret and strategy updates
Self-playMultiprocess data generation
Python + C++Learning and simulation stack
What the public source actually contains
Every capability below maps to a visible module. Unsupported claims about OCR, automatic online play, one-click executables or guaranteed performance are intentionally excluded.
02 · SOLVING
CFR subgame solver
Regret updates, average strategies, reach probabilities, best response, expected value and exploitability utilities.
03 · RECURSION
Belief-aware recursive solving
Build subgames from public states, propagate private-hand beliefs and continue solving after sampled actions.
04 · LEARNING
Counterfactual value networks
PyTorch models estimate leaf values and support configurable value and policy learning experiments.
05 · DATA
Self-play and replay
Multiprocess generation, model synchronization, replay buffers, training batches and validation workflows.
06 · NATIVE
C++ Monte Carlo layer
A native simulator and Python-binding foundation complement the Python game and learning modules.
From public state to learned strategy
The repository separates rules, solving and learning so each stage can be tested or extended independently.
The HUNL engine creates legal public states across preflop, flop, turn and river.
CFR traversals update regret and average policy; value networks can estimate truncated leaves.
Generated counterfactual-value and policy samples flow through replay into PyTorch models.
Configuration evidence
Research parameters are visible, not hidden behind marketing
The original repository image shows iterations, regret settings, pruning, model paths and Redis configuration. It is correctly labeled here as a training configuration and terminal view.
IterationsRegret settingsModel pathsRedis
Original repository asset: training configuration and terminal output.
Historical demonstration material supplied by the repository.
Table presentation
A clear boundary between demonstration and verified capability
The table image shows a robot seat and poker-table presentation. It does not prove recognition, automation or integration with any third-party service, so this page makes no such claim.
Responsible research only
Use Deep Hold'em AI for lawful research, education, offline training and benchmarking. Do not use it for unauthorized real-time assistance, automated online play, rule evasion or unfair play.