--- language: - en tags: - chess - reinforcement-learning - resnet - transformer - gambitflow - synapse-edge license: cc-by-nc-4.0 library_name: onnx metrics: - accuracy - mse pipeline_tag: zero-shot-classification --- # ♟️ GambitFlow Synapse-Edge v1 (Flagship)
![GambitFlow Banner](https://capsule-render.vercel.app/api?type=waving&color=0:2c3e50,100:000000&height=200§ion=header&text=Synapse-Edge%20v1&fontSize=50&animation=fadeIn&fontAlignY=35&desc=Next-Gen%20Hybrid%20Chess%20Intelligence&descAlignY=60) [![License: CC BY-NC 4.0](https://img.shields.io/badge/License-CC%20BY--NC%204.0-lightgrey.svg)](https://creativecommons.org/licenses/by-nc/4.0/) ![Architecture](https://img.shields.io/badge/Architecture-Hybrid%20ResNet--Transformer-orange) ![Parameters](https://img.shields.io/badge/Params-~16.5M-blue) ![Format](https://img.shields.io/badge/Format-ONNX%20(Opset%2017)-green) ![Status](https://img.shields.io/badge/Status-Flagship%20Release-red) [**Dataset Hub**](https://huggingface.co/datasets/GambitFlow/Synapse-Edge-Data) • [**Community Support**](https://huggingface.co/Rafs-an09002)
## 🌟 Model Overview **Synapse-Edge v1** is the definitive flagship chess AI from **GambitFlow**. Representing the pinnacle of our Gen-3 research, it shatters the limitations of pure convolutional models by introducing a massive **Hybrid ResNet-Transformer architecture**. While previous models like *Nexus-Core* excelled at recognizing spatial patterns, Synapse-Edge v1 masters **long-range tactical dependencies** and **strategic sequencing**, making it our most "human-like" yet superhumanly sharp engine to date. --- ## 🏗️ Technical Architecture The model utilizes a sophisticated multi-stage processing pipeline: ### 1. The Input: 119-Channel Rich Feature Map Instead of a simple 12-channel board state, Synapse-Edge v1 processes **119 discrete information layers** per position: - **Piece Occupancy (12):** Fundamental bitboards for all pieces. - **Attack Influence Maps (12):** Explicit spatial "heatmaps" of which squares are under fire. - **Auxiliary Metadata (95):** Castling rights, side to move, check status, en passant targets, and board history. ### 2. The Backbone: SE-ResNet-20 - **20 Residual Blocks** ensure deep feature extraction. - **Squeeze-and-Excitation (SE) Attention** modules in every block allow the network to dynamically recalibrate piece importance based on the position. ### 3. The Neck: Transformer Sequence Fusion - **4 Transformer Layers** process the board as a 64-square sequence. - This allows the model to understand **tactical causality** (e.g., *"If I move here, the pinned knight will be attacked 3 moves later"*). ### 4. Multi-Head Prediction System The model doesn't just evaluate; it understands the game through four specialized heads: - **Policy Head:** Predicts the most likely master-level move from **4,672 possible UCI combinations**. - **Value Head:** Provides a rock-solid evaluation in the range **[-1, +1]**. - **Tactical Head:** A binary classifier that flags **"Sharpness"** (detects Forks, Pins, and Skewers instantly). - **Phase Head:** Dynamically identifies game phases (**Opening, Middlegame, Endgame**) to adjust playing style. --- ## 📊 Training Details ### Distributed 4-Worker Sharding Synapse-Edge v1 was trained using a **High-Efficiency Distributed Pipeline**: - **Dataset:** Over **5.5 million elite positional samples** (Elo 2000+) + **3 million tactical puzzles**. - **Execution:** The database was split into 4 shards and trained simultaneously across 4 independent Google Colab instances. - **Synthesis:** The final model is a **Synchronized Ensemble** where weights from all four shards were merged and averaged to create a "Master Brain" with collective knowledge. | Specification | Value | | :--- | :--- | | **Total Parameters** | 16,494,757 | | **Total Samples** | 8.5 Million | | **Training Device** | 4x Tesla T4 GPUs (Distributed) | | **Optimizer** | AdamW (1e-4) | | **Precision** | Mixed (FP16/FP32) | --- ## 🚀 Usage & Implementation The model is exported in **ONNX (Opset 17)** for maximum cross-platform compatibility. ### Quick Start with Python ```python import onnxruntime as ort import numpy as np # Initialize the flagship engine session = ort.InferenceSession("synapse_edge_v1.onnx") # Prepare your input (119 channels) # dummy_input = np.random.randn(1, 119, 8, 8).astype(np.float32) # Run Multi-Head Inference policy, value, tactical, phase = session.run(None, {"input": dummy_input}) print(f"Value Score: {value[0][0]}") print(f"Tactical Sharpness: {tactical[0][0]}") ``` --- ## 🛣️ Roadmap: The Path to Superhuman Strength Synapse-Edge v1 is not a finished product, but the beginning of a **Continuous Development Pipeline**: 1. **v1 (Current):** Master-level baseline trained on master games and puzzles. 2. **v1.1 - v1.5:** Iterative fine-tuning on refined elite datasets. 3. **v2 (Self-Play):** The model will play against itself for weeks, generating "Alien Strategies" to surpass human theory. 4. **v3 (Final Flagship):** Full Reinforcement Learning (RL) integration aiming for **3500+ Elo**. --- ## 🛡️ Limitations & Bias - **Inference Latency:** Due to the Transformer layers, inference on CPU is slower than Nexus-Core (~100-300ms per position). For optimal performance, use GPU-based ONNX Runtime. - **Endgame Accuracy:** Without tablebases, very complex endgames (e.g., KBNK) may require more search depth. ## 📜 License This model is released under the **Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)** license. Commercial use without prior permission is prohibited. --- **Model Authors:** [Rafsan / GambitFlow](https://huggingface.co/Rafs-an09002) **Project Mission:** Democratizing Superhuman Chess AI through Neural Innovation. 🚀♟️