| | --- |
| | license: apache-2.0 |
| | task_categories: |
| | - reinforcement-learning |
| | language: |
| | - en |
| | - ru |
| | tags: |
| | - chess |
| | - deep-learning |
| | - stockfish |
| | - pytorch |
| | - multi-pv |
| | - chess-engine |
| | - reinforcement-learning |
| | size_categories: |
| | - 100K<n<1M |
| | --- |
| | |
| | # ♟️ Strategic Chess Dataset: Multi-PV & RL-Refined (700K+) |
| |
|
| | This is a high-performance dataset designed for training and pre-training state-of-the-art chess neural networks. It contains over **706,000 unique board positions** generated and evaluated by Stockfish 16.1. |
| |
|
| | The dataset is specifically optimized for models using **Policy & Value heads**, providing rich metadata for each state. |
| |
|
| | ## 🌟 Key Features |
| |
|
| | * **Multi-PV Intelligence:** Each position includes not just the single best move, but **3 strong alternative plans**. This allows models to learn strategic variability and fine-grained positional judgment. |
| | * **15-Channel Encoding:** Data is pre-structured for advanced architectures. It includes 12 piece layers, 1 side-to-move layer, and **2 temporal layers** (from/to squares of the last move) to eliminate tactical blindness. |
| | * **RL-Refined Accuracy:** Includes a specialized subset of **5,000+ positions** derived from Reinforcement Learning sessions. These capture "hard-to-learn" tactical blunders that were corrected by Stockfish during active self-play. |
| | * **High-Performance Processing:** The entire dataset was generated and processed using a cluster with **128+ CPU cores**, ensuring consistent and deep engine evaluation for every frame. |
| | * **Ready-to-Train Eval:** Position evaluations are pre-normalized using the $tanh(x / 300.0)$ function, mapping Stockfish centipawns to a perfect $[-1, 1]$ range for Stable MSE training. |
| |
|
| | --- |
| |
|
| | ## 📊 Data Structure |
| |
|
| | The dataset is provided in a compressed `.npz` format: |
| |
|
| | * **`states`**: `(N, 15, 8, 8)` float32 tensors representing the board state. |
| | * **`plans`**: `(N, 3, 1)` int64 array containing Multi-PV move indices ($from\_square \times 64 + to\_square$). |
| | * **`evals`**: `(N,)` float32 array of normalized position evaluations. |
| |
|
| | --- |
| |
|
| | ## 🛠️ Usage (PyTorch Example) |
| |
|
| | ```python |
| | import numpy as np |
| | import torch |
| | from torch.utils.data import Dataset |
| | |
| | class StrategicChessDataset(Dataset): |
| | def __init__(self, npz_path): |
| | data = np.load(npz_path) |
| | self.states = data['states'] |
| | self.evals = data['evals'] |
| | # Extract the primary best move from Multi-PV plans |
| | self.best_moves = data['plans'][:, 0, 0] |
| | |
| | def __len__(self): |
| | return len(self.states) |
| | |
| | def __getitem__(self, idx): |
| | state = torch.from_numpy(self.states[idx]).float() |
| | move = torch.tensor(self.best_moves[idx], dtype=torch.long) |
| | val = torch.tensor(self.evals[idx], dtype=torch.float32) |
| | return state, move, val |
| | |
| | ``` |
| | 📈 Intended Use |
| | This dataset is ideal for: |
| |
|
| | Pre-training Chess Policy-Value networks (like AlphaZero or LC0 clones). |
| |
|
| | Fine-tuning models to reduce tactical blunders. |
| |
|
| | Researching Reinforcement Learning and MCTS-based agents. |
| |
|
| | 📜 License |
| | This dataset is licensed under the Apache 2.0 License. You are free to use, modify, and distribute it for any purpose, including commercial projects. |
