🧠 Synapse-Base: The Hybrid Chess Foundation (v3.0)

---

📖 Model Details

Synapse-Base is a foundation-scale neural chess evaluator. It is the first model in the GambitFlow project to utilize a Hybrid CNN-Transformer architecture. By merging spatial perception (CNN) with long-range strategic reasoning (Transformer), Synapse-Base identifies complex patterns like pins, skewers, and multi-move tactical sequences that traditional CNNs often overlook.

• Developer: GambitFlow / Rafsan1711
• Model Type: Evaluation & Move Prediction Foundation Model
• Parent Series: Synapse Ecosystem

---

🏗️ Technical Specifications

Hybrid Architecture

Synapse-Base employs a "Residual Neck" design:

1. Residual Backbone (CNN): 20 layers of ResNet blocks for immediate board feature extraction.
2. Attention Neck (Transformer): 4 layers of Multi-Head Self-Attention (8 heads) to model relationships between distant squares (e.g., a Bishop pinning a Queen from across the board).
3. Dual Heads:
   • Value Head: Predicting the winning probability [-1.0 to 1.0].
   • Policy Head: Pre-trained weights for move suggestion (4096-logit distribution).

💾 Input Representation: The 119 Channels

To prevent "tactical blindness," the model consumes a massive 119-layer bitboard:

• 0-11: Piece placement (6 white, 6 black types).
• 12-26: Board metadata (Turn, Castling, EP, Check status).
• 27-50: Tactical Vision Maps (Attack and Defense heatmaps).
• 51-66: Coordinate Masking (Rank/File identifiers).
• 67-118: Static Positional Biases (Pre-trained PST heuristics).

---

📊 Training Data

The model was trained on the GambitFlow Elite Database:

• Source: 5,000,000+ filtered Lichess games.
• Criteria: Both players must be rated 2000 ELO or higher.
• Goal: To ensure the model learns professional strategy and avoids amateur blunders.

---

💻 How to Get Started

Quick Inference (JavaScript)

import * as ort from 'onnxruntime-web';

// Load the session
const session = await ort.InferenceSession.create('./synapse_base.onnx');

// Prepare 119-channel input (Float32Array)
const input = new Float32Array(1 * 119 * 8 * 8); 
const tensor = new ort.Tensor('float32', input, [1, 119, 8, 8]);

// Run
const { value } = await session.run({ board_state: tensor });
console.log("Evaluation:", value.data[0]);

Weights for Fine-Tuning (PyTorch)

The synapse_base.pth file contains the full state dictionary for further training.

model = SynapseBase(num_residual_blocks=20, num_filters=256)
model.load_state_dict(torch.load('synapse_base.pth'))

---

📝 Release Checklist

• Pre-trained on 5M+ Master positions.
• 119-Channel density validation (No empty layers).
• ONNX Opset 17 export validation.
• README metadata compliant with HF Hub spec.
• Verified compatibility with 18GB CPU HF Spaces.

---

⚠️ Limitations

• Endgame Precision: May struggle with precise 7-piece tablebase draws. This model is licensed under GPL v3 (GNU General Public License Version 3)

---

Part of the GambitFlow Project. Flow with the perfect move. ♟️