| --- |
| license: mit |
| tags: |
| - reinforcement-learning |
| - poker |
| - stable-baselines3 |
| - ppo |
| - texas-holdem |
| library_name: stable-baselines3 |
| --- |
| |
| # PokerForge PPO Poker Bots |
|
|
| This repository contains runtime artifacts for PokerForge, a full-stack AI poker |
| platform built around heads-up No-Limit Texas Hold'em abstractions. |
|
|
| ## Files |
|
|
| - `models/medium/ppo_medium_final.zip` - PPO medium bot trained for 1M timesteps vs easy. |
| - `models/hard/ppo_hard_final.zip` - PPO hard bot trained for 5M timesteps vs frozen medium. |
| - `models/*/best_model.zip` - best checkpoints from training/evaluation callbacks. |
| - `reports/evaluation_report.*` - latest reproducible bot-vs-bot evaluation report. |
| - `reports/representative_hands.json` - replay-ready sample hand logs for the frontend dashboard. |
|
|
| ## Runtime Contract |
|
|
| - Framework: Stable-Baselines3 PPO |
| - Observation space: `Box(18,)` |
| - Action space: `Discrete(3)` where `0=fold`, `1=check/call`, `2=raise` |
| - Expected local paths inside PokerForge: |
| - `backend/data/models/medium/ppo_medium_final.zip` |
| - `backend/data/models/hard/ppo_hard_final.zip` |
|
|
| ## Evaluation Summary |
|
|
| The latest evaluation report is included under `reports/`. The current honest |
| finding is that medium and hard both beat easy, while hard only shows a marginal, |
| statistically weak edge over medium. This is attributed mainly to the limited |
| 3-action abstraction creating a ceiling on behavioral differentiation. |
|
|
| ## Reproduce In PokerForge |
|
|
| ```bash |
| cd backend |
| python tools/download_models.py --repo-id Rushisagar221/pokerforge-bots --if-missing |
| python server.py |
| ``` |
|
|
| ## Manifest |
|
|
| ```json |
| { |
| "repo_id": "Rushisagar221/pokerforge-bots", |
| "generated_at": "2026-04-23T14:08:03.040902", |
| "artifacts": [ |
| { |
| "path": "models/medium/ppo_medium_final.zip", |
| "bytes": 162131, |
| "sha256": "b8ed8a7217de2bc790af71a0dbdc6a5a9fd695fcf541351bb965549d3c20c126" |
| }, |
| { |
| "path": "models/medium/best_model.zip", |
| "bytes": 162116, |
| "sha256": "31d26001f967b7d221af016ec1a4c5b1a33f32b71630cb2eea3bf9c8a2e59956" |
| }, |
| { |
| "path": "models/hard/ppo_hard_final.zip", |
| "bytes": 165087, |
| "sha256": "ac3b23fd8188713cd25bcbd1585cfc213d1a05b6254c56ba59ee7119de5896e1" |
| }, |
| { |
| "path": "models/hard/best_model.zip", |
| "bytes": 165087, |
| "sha256": "ac3b23fd8188713cd25bcbd1585cfc213d1a05b6254c56ba59ee7119de5896e1" |
| }, |
| { |
| "path": "reports/evaluation_report.json", |
| "bytes": 51143, |
| "sha256": "1ce452e2e57e67965f13337cb12a736cf658467e3db70ea20b52eaeddb67532a" |
| }, |
| { |
| "path": "reports/evaluation_report.md", |
| "bytes": 2486, |
| "sha256": "2ed28899b86090b97bdacbae1273c4366202344e42ab8b93013cdc260db378bb" |
| }, |
| { |
| "path": "reports/representative_hands.json", |
| "bytes": 64864, |
| "sha256": "33054b06fbfe819dbfb98faafa32534ff149bbd254f9250fb405786fbd2ecaf3" |
| } |
| ] |
| } |
| ``` |
|
|
