Update README.md

README.md

@@ -10,7 +10,8 @@ library_name: transformers
 
 # ChessBot Chess Model
 
-This is a ChessBot model for chess move prediction and position evaluation.
+This is a ChessBot model for chess move prediction and position evaluation. This model is way worse than stockfish. It is better than most humans however.
+For stronger play, reducing temperature T (lower is stronger) is suggested.
 
 ## Model Description
 
@@ -32,17 +33,20 @@ model = model.to(device)
 # Example usage
 fen = "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1"
 
-# Get the best move
+# Sample move from policy
 move = model.get_move_from_fen_no_thinking(fen, T=0.1, device=device)
 print(f"Policy-based move: {move}")
+#e2e4
 
 # Get the best move using value analysis
 value_move = model.get_best_move_value(fen, T=0, device=device)
 print(f"Value-based move: {value_move}")
+#e2e4
 
 # Get position evaluation
 position_value = model.get_position_value(fen, device=device)
 print(f"Position value [black_win, draw, white_win]: {position_value}")
+#[0.2318, 0.4618, 0.3064]
 
 # Get move probabilities
 probs = model.get_move_from_fen_no_thinking(fen, T=1, device=device, return_probs=True)
@@ -50,6 +54,12 @@ top_moves = sorted(probs.items(), key=lambda x: x[1], reverse=True)[:5]
 print("Top 5 moves:")
 for move, prob in top_moves:
     print(f"  {move}: {prob:.4f}")
+#Top 5 moves:
+#  e2e4: 0.9285
+#  d2d4: 0.0712
+#  g1f3: 0.0001
+#  e2e3: 0.0000
+#  c2c3: 0.0000
 ```
 
 ## Requirements
@@ -61,6 +71,8 @@ for move, prob in top_moves:
 
 ## Model Architecture
 
+The architecture is strongly inspired from the LCzero project. Although written in pytorch.
+
 - **Transformer layers**: 10
 - **Hidden size**: 512
 - **Feed-forward size**: 736
@@ -69,7 +81,7 @@ for move, prob in top_moves:
 
 ## Training Data
 
-This model was trained on chess game data to learn optimal move selection and position evaluation.
+This model was trained on training data from the LCzero project. It consists of around 750M chess positions. I will publish the training dataset very soon.
 
 ## Limitations