add src folder and webhook

app.py

@@ -9,20 +9,28 @@ This Gradio app provides:
 Leaderboard data is stored in a private HuggingFace dataset for persistence.
 """
 
+import hashlib
+import hmac
 import io
 import os
+import sys
 from datetime import datetime
 from pathlib import Path
 from typing import Optional
 
 import gradio as gr
 import pandas as pd
+from fastapi import FastAPI, Request, BackgroundTasks
+
+# Create FastAPI app for webhook
+fastapi_app = FastAPI()
 
 # Configuration
 ORGANIZATION = os.environ.get("HF_ORGANIZATION", "LLM-course")
 LEADERBOARD_DATASET = os.environ.get("LEADERBOARD_DATASET", f"{ORGANIZATION}/chess-challenge-leaderboard")
 LEADERBOARD_FILENAME = "leaderboard.csv"
 HF_TOKEN = os.environ.get("HF_TOKEN")  # Required for private dataset access
+WEBHOOK_SECRET = os.environ.get("WEBHOOK_SECRET", "")  # For webhook verification
 
 STOCKFISH_LEVELS = {
     "Beginner (Level 0)": 0,
@@ -293,9 +301,9 @@ def evaluate_legal_moves(
     """Evaluate a model's legal move generation."""
     try:
         import sys
-        sys.path.insert(0, str(Path(__file__).parent.parent / "src"))
+        sys.path.insert(0, str(Path(__file__).parent))
         
-        from chess_challenge.evaluate import ChessEvaluator, load_model_from_hub
+        from src.evaluate import ChessEvaluator, load_model_from_hub
         
         progress(0, desc="Loading model...")
         model, tokenizer = load_model_from_hub(model_id)
@@ -355,9 +363,9 @@ def evaluate_winrate(
     """Evaluate a model's win rate against Stockfish."""
     try:
         import sys
-        sys.path.insert(0, str(Path(__file__).parent.parent / "src"))
+        sys.path.insert(0, str(Path(__file__).parent))
         
-        from chess_challenge.evaluate import ChessEvaluator, load_model_from_hub
+        from src.evaluate import ChessEvaluator, load_model_from_hub
         
         progress(0, desc="Loading model...")
         model, tokenizer = load_model_from_hub(model_id)
@@ -419,9 +427,9 @@ def evaluate_model(
     try:
         # Import evaluation code
         import sys
-        sys.path.insert(0, str(Path(__file__).parent.parent / "src"))
+        sys.path.insert(0, str(Path(__file__).parent))
         
-        from chess_challenge.evaluate import ChessEvaluator, load_model_from_hub
+        from src.evaluate import ChessEvaluator, load_model_from_hub
         
         progress(0, desc="Loading model...")
         model, tokenizer = load_model_from_hub(model_id)
@@ -660,5 +668,104 @@ with gr.Blocks(
             )
 
 
+# =============================================================================
+# WEBHOOK HANDLERS FOR AUTOMATIC EVALUATION
+# =============================================================================
+
+def verify_webhook_signature(payload: bytes, signature: str) -> bool:
+    """Verify the webhook signature from Hugging Face."""
+    if not WEBHOOK_SECRET:
+        print("⚠️ WEBHOOK_SECRET not set - skipping signature verification")
+        return True
+    expected = hmac.new(WEBHOOK_SECRET.encode(), payload, hashlib.sha256).hexdigest()
+    return hmac.compare_digest(f"sha256={expected}", signature)
+
+
+def run_auto_evaluation(model_id: str):
+    """Run model evaluation in background after webhook trigger."""
+    try:
+        print(f"🚀 Auto-evaluating new model: {model_id}")
+        
+        # Import evaluation functions
+        sys.path.insert(0, str(Path(__file__).parent))
+        from src.evaluate import ChessEvaluator, load_model_from_hub
+        
+        # Load model
+        model, tokenizer = load_model_from_hub(model_id)
+        
+        # Run legal moves evaluation (quick first pass)
+        evaluator = ChessEvaluator(
+            model=model,
+            tokenizer=tokenizer,
+            stockfish_level=1,
+        )
+        results = evaluator.evaluate_legal_moves(n_positions=100, verbose=True)
+        
+        # Update leaderboard
+        leaderboard = load_leaderboard()
+        entry = next((e for e in leaderboard if e["model_id"] == model_id), None)
+        if entry is None:
+            entry = {"model_id": model_id}
+            leaderboard.append(entry)
+        
+        entry.update({
+            "legal_rate": results.get("legal_rate_with_retry", 0),
+            "legal_rate_first_try": results.get("legal_rate_first_try", 0),
+            "last_updated": datetime.now().strftime("%Y-%m-%d %H:%M"),
+        })
+        
+        save_leaderboard(leaderboard)
+        print(f"✅ Auto-evaluation complete for {model_id}: legal_rate={results.get('legal_rate_with_retry', 0):.1%}")
+        
+    except Exception as e:
+        print(f"❌ Auto-evaluation failed for {model_id}: {e}")
+        import traceback
+        traceback.print_exc()
+
+
+@fastapi_app.post("/webhook")
+async def handle_webhook(request: Request, background_tasks: BackgroundTasks):
+    """Handle incoming webhooks from Hugging Face."""
+    payload = await request.body()
+    signature = request.headers.get("X-Webhook-Signature", "")
+    
+    # Verify signature
+    if not verify_webhook_signature(payload, signature):
+        print("��� Webhook signature verification failed")
+        return {"error": "Invalid signature"}, 403
+    
+    data = await request.json()
+    event = data.get("event", {})
+    event_type = event.get("action")
+    repo = data.get("repo", {})
+    repo_type = repo.get("type")
+    repo_name = repo.get("name")
+    
+    print(f"📥 Webhook received: {event_type} for {repo_type}/{repo_name}")
+    
+    # Only process model creation/updates in our organization
+    if repo_type == "model" and repo_name and repo_name.startswith(f"{ORGANIZATION}/"):
+        if event_type in ["create", "update"]:
+            # Check if it's a chess model
+            if "chess" in repo_name.lower():
+                print(f"🎯 Queuing evaluation for chess model: {repo_name}")
+                background_tasks.add_task(run_auto_evaluation, repo_name)
+                return {"status": "evaluation_queued", "model": repo_name}
+            else:
+                print(f"⏭️ Skipping non-chess model: {repo_name}")
+    
+    return {"status": "ignored"}
+
+
+@fastapi_app.get("/health")
+async def health_check():
+    """Health check endpoint."""
+    return {"status": "healthy", "organization": ORGANIZATION}
+
+
+# Mount Gradio app to FastAPI
+fastapi_app = gr.mount_gradio_app(fastapi_app, demo, path="/")
+
 if __name__ == "__main__":
-    demo.launch()
+    import uvicorn
+    uvicorn.run(fastapi_app, host="0.0.0.0", port=7860)

requirements.txt

@@ -5,3 +5,5 @@ python-chess>=1.999
 huggingface-hub>=0.20.0
 datasets>=2.14.0
 pandas>=2.0.0
+fastapi
+uvicorn

src/__init__.py

@@ -0,0 +1,13 @@
+"""Chess Challenge source module."""
+
+from .model import ChessConfig, ChessForCausalLM
+from .tokenizer import ChessTokenizer
+from .evaluate import ChessEvaluator, load_model_from_hub
+
+__all__ = [
+    "ChessConfig",
+    "ChessForCausalLM", 
+    "ChessTokenizer",
+    "ChessEvaluator",
+    "load_model_from_hub",
+]

src/evaluate.py

@@ -0,0 +1,619 @@
+"""
+Evaluation script for the Chess Challenge.
+
+This script evaluates a trained chess model by playing games against
+Stockfish and computing ELO ratings.
+"""
+
+from __future__ import annotations
+
+import argparse
+import random
+from dataclasses import dataclass
+from typing import List, Optional, Tuple
+
+import torch
+
+
+@dataclass
+class GameResult:
+    """Result of a single game."""
+    moves: List[str]
+    result: str  # "1-0", "0-1", or "1/2-1/2"
+    model_color: str  # "white" or "black"
+    termination: str  # "checkmate", "stalemate", "illegal_move", "max_moves", etc.
+    illegal_move_count: int
+
+
+class ChessEvaluator:
+    """
+    Evaluator for chess models.
+    
+    This class handles playing games between a trained model and Stockfish,
+    tracking results, and computing ELO ratings.
+    """
+    
+    def __init__(
+        self,
+        model,
+        tokenizer,
+        stockfish_path: Optional[str] = None,
+        stockfish_level: int = 1,
+        max_retries: int = 3,
+        device: str = "cuda" if torch.cuda.is_available() else "cpu",
+    ):
+        """
+        Initialize the evaluator.
+        
+        Args:
+            model: The trained chess model.
+            tokenizer: The chess tokenizer.
+            stockfish_path: Path to Stockfish executable.
+            stockfish_level: Stockfish skill level (0-20).
+            max_retries: Maximum retries for illegal moves.
+            device: Device to run the model on.
+        """
+        self.model = model.to(device)
+        self.tokenizer = tokenizer
+        self.max_retries = max_retries
+        self.device = device
+        
+        # Initialize Stockfish
+        try:
+            import chess
+            import chess.engine
+            
+            self.chess = chess
+            
+            if stockfish_path is None:
+                # Try common paths
+                import shutil
+                stockfish_path = shutil.which("stockfish")
+            
+            if stockfish_path:
+                self.engine = chess.engine.SimpleEngine.popen_uci(stockfish_path)
+                self.engine.configure({"Skill Level": stockfish_level})
+            else:
+                print("WARNING: Stockfish not found. Install it for full evaluation.")
+                self.engine = None
+                
+        except ImportError:
+            raise ImportError(
+                "python-chess is required for evaluation. "
+                "Install it with: pip install python-chess"
+            )
+    
+    def __del__(self):
+        """Clean up Stockfish engine."""
+        if hasattr(self, 'engine') and self.engine:
+            self.engine.quit()
+    
+    def _convert_board_to_moves(self, board) -> str:
+        """Convert board move history to model input format."""
+        moves = []
+        temp_board = self.chess.Board()
+        
+        for move in board.move_stack:
+            # Get piece and color
+            color = "W" if temp_board.turn == self.chess.WHITE else "B"
+            piece = temp_board.piece_at(move.from_square)
+            piece_letter = piece.symbol().upper() if piece else "P"
+            
+            # Get squares
+            from_sq = self.chess.square_name(move.from_square)
+            to_sq = self.chess.square_name(move.to_square)
+            
+            move_str = f"{color}{piece_letter}{from_sq}{to_sq}"
+            
+            # Add promotion
+            if move.promotion:
+                move_str += f"={self.chess.piece_symbol(move.promotion).upper()}"
+            
+            # Add capture suffix
+            if temp_board.is_capture(move):
+                move_str += "(x)"
+            
+            # Add check/checkmate suffix
+            temp_board.push(move)
+            if temp_board.is_checkmate():
+                move_str = move_str.replace("(x)", "(x+*)") if "(x)" in move_str else move_str + "(+*)"
+            elif temp_board.is_check():
+                move_str = move_str.replace("(x)", "(x+)") if "(x)" in move_str else move_str + "(+)"
+            
+            # Handle castling
+            if piece_letter == "K" and abs(ord(from_sq[0]) - ord(to_sq[0])) > 1:
+                if to_sq[0] == 'g':  # Kingside
+                    move_str = move_str.split("(")[0] + "(o)"
+                else:  # Queenside
+                    move_str = move_str.split("(")[0] + "(O)"
+            
+            moves.append(move_str)
+        
+        return " ".join(moves)
+    
+    def _get_model_move(
+        self,
+        board,
+        temperature: float = 0.7,
+        top_k: int = 10,
+    ) -> Tuple[Optional[str], int]:
+        """
+        Get the model's next move prediction.
+        
+        Returns:
+            Tuple of (UCI move string, number of retries used).
+        """
+        self.model.eval()
+        
+        # Convert board to input format
+        moves_str = self._convert_board_to_moves(board)
+        
+        # Add BOS token if no moves yet
+        if not moves_str:
+            input_text = self.tokenizer.bos_token
+        else:
+            input_text = self.tokenizer.bos_token + " " + moves_str
+        
+        # Tokenize
+        inputs = self.tokenizer(
+            input_text,
+            return_tensors="pt",
+            truncation=True,
+            max_length=self.model.config.n_ctx - 1,
+        ).to(self.device)
+        
+        # Try to generate a legal move
+        for retry in range(self.max_retries):
+            with torch.no_grad():
+                outputs = self.model(**inputs)
+                logits = outputs.logits[:, -1, :] / temperature
+                
+                # Apply top-k filtering
+                if top_k > 0:
+                    indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+                    logits[indices_to_remove] = float("-inf")
+                
+                # Sample
+                probs = torch.softmax(logits, dim=-1)
+                next_token = torch.multinomial(probs, num_samples=1)
+            
+            # Decode the move
+            move_token = self.tokenizer.decode(next_token[0])
+            
+            # Convert to UCI
+            if len(move_token) >= 6:
+                uci_move = move_token[2:4] + move_token[4:6]
+                
+                # Handle promotion
+                if "=" in move_token:
+                    promo_idx = move_token.index("=")
+                    uci_move += move_token[promo_idx + 1].lower()
+                
+                try:
+                    move = self.chess.Move.from_uci(uci_move)
+                    if move in board.legal_moves:
+                        return uci_move, retry
+                except (ValueError, self.chess.InvalidMoveError):
+                    pass
+            
+            # Mask out the tried token for next retry
+            logits[0, next_token[0]] = float("-inf")
+        
+        return None, self.max_retries
+    
+    def _get_stockfish_move(self, board, time_limit: float = 0.1) -> str:
+        """Get Stockfish's move."""
+        if self.engine is None:
+            raise RuntimeError("Stockfish engine not initialized")
+        
+        result = self.engine.play(board, self.chess.engine.Limit(time=time_limit))
+        return result.move.uci()
+    
+    def play_game(
+        self,
+        model_color: str = "white",
+        max_moves: int = 200,
+        temperature: float = 0.7,
+    ) -> GameResult:
+        """
+        Play a single game between the model and Stockfish.
+        
+        Args:
+            model_color: "white" or "black".
+            max_moves: Maximum number of moves before draw.
+            temperature: Sampling temperature for model.
+        
+        Returns:
+            GameResult with the game details.
+        """
+        board = self.chess.Board()
+        moves = []
+        illegal_move_count = 0
+        
+        model_is_white = model_color == "white"
+        
+        while not board.is_game_over() and len(moves) < max_moves:
+            is_model_turn = (board.turn == self.chess.WHITE) == model_is_white
+            
+            if is_model_turn:
+                # Model's turn
+                uci_move, retries = self._get_model_move(board, temperature)
+                illegal_move_count += retries
+                
+                if uci_move is None:
+                    # Model couldn't find a legal move
+                    return GameResult(
+                        moves=moves,
+                        result="0-1" if model_is_white else "1-0",
+                        model_color=model_color,
+                        termination="illegal_move",
+                        illegal_move_count=illegal_move_count + 1,
+                    )
+                
+                move = self.chess.Move.from_uci(uci_move)
+            else:
+                # Stockfish's turn
+                if self.engine:
+                    uci_move = self._get_stockfish_move(board)
+                    move = self.chess.Move.from_uci(uci_move)
+                else:
+                    # Random move if no engine
+                    move = random.choice(list(board.legal_moves))
+            
+            board.push(move)
+            moves.append(move.uci())
+        
+        # Determine result
+        if board.is_checkmate():
+            if board.turn == self.chess.WHITE:
+                result = "0-1"  # Black wins
+            else:
+                result = "1-0"  # White wins
+            termination = "checkmate"
+        elif board.is_stalemate():
+            result = "1/2-1/2"
+            termination = "stalemate"
+        elif board.is_insufficient_material():
+            result = "1/2-1/2"
+            termination = "insufficient_material"
+        elif board.can_claim_draw():
+            result = "1/2-1/2"
+            termination = "draw_claim"
+        elif len(moves) >= max_moves:
+            result = "1/2-1/2"
+            termination = "max_moves"
+        else:
+            result = "1/2-1/2"
+            termination = "unknown"
+        
+        return GameResult(
+            moves=moves,
+            result=result,
+            model_color=model_color,
+            termination=termination,
+            illegal_move_count=illegal_move_count,
+        )
+    
+    def evaluate_legal_moves(
+        self,
+        n_positions: int = 1000,
+        temperature: float = 0.7,
+        verbose: bool = True,
+    ) -> dict:
+        """
+        Evaluate the model's ability to generate legal moves.
+        
+        This evaluation only checks if the model generates legal moves,
+        without playing full games. Useful as a first-pass evaluation.
+        
+        Args:
+            n_positions: Number of positions to test.
+            temperature: Sampling temperature.
+            verbose: Whether to print progress.
+        
+        Returns:
+            Dictionary with legal move statistics.
+        """
+        results = {
+            "total_positions": 0,
+            "legal_first_try": 0,
+            "legal_with_retry": 0,
+            "illegal_all_retries": 0,
+            "positions": [],
+        }
+        
+        # Generate random positions by playing random moves
+        for i in range(n_positions):
+            board = self.chess.Board()
+            
+            # Play random number of moves (5-40) to get varied positions
+            n_random_moves = random.randint(5, 40)
+            for _ in range(n_random_moves):
+                if board.is_game_over():
+                    break
+                move = random.choice(list(board.legal_moves))
+                board.push(move)
+            
+            if board.is_game_over():
+                continue  # Skip terminal positions
+            
+            results["total_positions"] += 1
+            
+            # Test model's move generation
+            uci_move, retries = self._get_model_move(board, temperature)
+            
+            position_result = {
+                "fen": board.fen(),
+                "move_number": len(board.move_stack),
+                "legal": uci_move is not None,
+                "retries": retries,
+            }
+            results["positions"].append(position_result)
+            
+            if uci_move is not None:
+                if retries == 0:
+                    results["legal_first_try"] += 1
+                else:
+                    results["legal_with_retry"] += 1
+            else:
+                results["illegal_all_retries"] += 1
+            
+            if verbose and (i + 1) % 100 == 0:
+                legal_rate = (results["legal_first_try"] + results["legal_with_retry"]) / results["total_positions"]
+                print(f"  Positions: {i + 1}/{n_positions} | Legal rate: {legal_rate:.1%}")
+        
+        # Calculate statistics
+        total = results["total_positions"]
+        if total > 0:
+            results["legal_rate_first_try"] = results["legal_first_try"] / total
+            results["legal_rate_with_retry"] = (results["legal_first_try"] + results["legal_with_retry"]) / total
+            results["illegal_rate"] = results["illegal_all_retries"] / total
+        else:
+            results["legal_rate_first_try"] = 0
+            results["legal_rate_with_retry"] = 0
+            results["illegal_rate"] = 1
+        
+        return results
+    
+    def evaluate(
+        self,
+        n_games: int = 100,
+        temperature: float = 0.7,
+        verbose: bool = True,
+    ) -> dict:
+        """
+        Run a full win-rate evaluation of the model against Stockfish.
+        
+        Args:
+            n_games: Number of games to play.
+            temperature: Sampling temperature.
+            verbose: Whether to print progress.
+        
+        Returns:
+            Dictionary with evaluation metrics.
+        """
+        results = {
+            "wins": 0,
+            "losses": 0,
+            "draws": 0,
+            "illegal_moves": 0,
+            "total_moves": 0,
+            "games": [],
+        }
+        
+        for i in range(n_games):
+            # Alternate colors
+            model_color = "white" if i % 2 == 0 else "black"
+            
+            game = self.play_game(
+                model_color=model_color,
+                temperature=temperature,
+            )
+            
+            results["games"].append(game)
+            results["total_moves"] += len(game.moves)
+            results["illegal_moves"] += game.illegal_move_count
+            
+            # Count result
+            if game.result == "1/2-1/2":
+                results["draws"] += 1
+            elif (game.result == "1-0" and model_color == "white") or \
+                 (game.result == "0-1" and model_color == "black"):
+                results["wins"] += 1
+            else:
+                results["losses"] += 1
+            
+            if verbose and (i + 1) % 10 == 0:
+                print(f"  Games: {i + 1}/{n_games} | "
+                      f"W: {results['wins']} L: {results['losses']} D: {results['draws']}")
+        
+        # Calculate statistics
+        total = results["wins"] + results["losses"] + results["draws"]
+        results["win_rate"] = results["wins"] / total if total > 0 else 0
+        results["draw_rate"] = results["draws"] / total if total > 0 else 0
+        results["loss_rate"] = results["losses"] / total if total > 0 else 0
+
+        total_attempts = results["total_moves"] + results["illegal_moves"]
+
+        # Average length counts both legal moves and illegal attempts so early illegal terminations
+        # don't show as near-zero length games.
+        results["avg_game_length"] = total_attempts / total if total > 0 else 0
+
+        # Illegal move rate: illegal attempts over total attempts
+        results["illegal_move_rate"] = results["illegal_moves"] / total_attempts if total_attempts > 0 else 0
+        
+        # Estimate ELO (simplified)
+        # Stockfish Level 1 is approximately 1350 ELO
+        stockfish_elo = 1350
+        if results["win_rate"] > 0 or results["loss_rate"] > 0:
+            score = results["wins"] + 0.5 * results["draws"]
+            expected = total * 0.5  # Expected score against equal opponent
+            
+            # Simple ELO estimation
+            if score > 0:
+                win_ratio = score / total
+                if win_ratio > 0 and win_ratio < 1:
+                    elo_diff = -400 * (1 - 2 * win_ratio) / (1 if win_ratio > 0.5 else -1)
+                    results["estimated_elo"] = stockfish_elo + elo_diff
+                else:
+                    results["estimated_elo"] = stockfish_elo + (400 if win_ratio >= 1 else -400)
+            else:
+                results["estimated_elo"] = stockfish_elo - 400
+        else:
+            results["estimated_elo"] = None
+        
+        return results
+
+
+def load_model_from_hub(model_id: str, device: str = "auto"):
+    """
+    Load a model from the Hugging Face Hub.
+    
+    Args:
+        model_id: Model ID on Hugging Face Hub.
+        device: Device to load the model on.
+    
+    Returns:
+        Tuple of (model, tokenizer).
+    """
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+    
+    # Import to register custom classes (use relative import or handle both cases)
+    try:
+        from src.model import ChessConfig, ChessForCausalLM
+    except ImportError:
+        from .model import ChessConfig, ChessForCausalLM
+    
+    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
+    model = AutoModelForCausalLM.from_pretrained(
+        model_id,
+        trust_remote_code=True,
+        device_map=device,
+    )
+    
+    return model, tokenizer
+
+
+def main():
+    """Main evaluation function."""
+    parser = argparse.ArgumentParser(description="Evaluate a chess model")
+    
+    parser.add_argument(
+        "--model_path", type=str, required=True,
+        help="Path to the model or Hugging Face model ID"
+    )
+    parser.add_argument(
+        "--mode", type=str, default="both", choices=["legal", "winrate", "both"],
+        help="Evaluation mode: 'legal' for legal move rate, 'winrate' for games, 'both' for both"
+    )
+    parser.add_argument(
+        "--stockfish_path", type=str, default=None,
+        help="Path to Stockfish executable"
+    )
+    parser.add_argument(
+        "--stockfish_level", type=int, default=1,
+        help="Stockfish skill level (0-20)"
+    )
+    parser.add_argument(
+        "--n_positions", type=int, default=500,
+        help="Number of positions for legal move evaluation"
+    )
+    parser.add_argument(
+        "--n_games", type=int, default=100,
+        help="Number of games to play for win rate evaluation"
+    )
+    parser.add_argument(
+        "--temperature", type=float, default=0.7,
+        help="Sampling temperature"
+    )
+    
+    args = parser.parse_args()
+    
+    print("=" * 60)
+    print("CHESS CHALLENGE - EVALUATION")
+    print("=" * 60)
+    
+    # Load model
+    print(f"\nLoading model from: {args.model_path}")
+    
+    if "/" in args.model_path and not args.model_path.startswith("."):
+        # Assume Hugging Face model ID
+        model, tokenizer = load_model_from_hub(args.model_path)
+    else:
+        # Local path
+        from transformers import AutoModelForCausalLM
+        try:
+            from src.tokenizer import ChessTokenizer
+            from src.model import ChessConfig, ChessForCausalLM
+        except ImportError:
+            from .tokenizer import ChessTokenizer
+            from .model import ChessConfig, ChessForCausalLM
+        
+        tokenizer = ChessTokenizer.from_pretrained(args.model_path)
+        model = AutoModelForCausalLM.from_pretrained(args.model_path)
+    
+    # Create evaluator
+    print(f"\nSetting up evaluator...")
+    evaluator = ChessEvaluator(
+        model=model,
+        tokenizer=tokenizer,
+        stockfish_path=args.stockfish_path,
+        stockfish_level=args.stockfish_level,
+    )
+    
+    # Run legal move evaluation
+    if args.mode in ["legal", "both"]:
+        print(f"\n" + "=" * 60)
+        print("PHASE 1: LEGAL MOVE EVALUATION")
+        print("=" * 60)
+        print(f"Testing {args.n_positions} random positions...")
+        
+        legal_results = evaluator.evaluate_legal_moves(
+            n_positions=args.n_positions,
+            temperature=args.temperature,
+            verbose=True,
+        )
+        
+        print("\n" + "-" * 40)
+        print("LEGAL MOVE RESULTS")
+        print("-" * 40)
+        print(f"  Positions tested:     {legal_results['total_positions']}")
+        print(f"  Legal (1st try):      {legal_results['legal_first_try']} ({legal_results['legal_rate_first_try']:.1%})")
+        print(f"  Legal (with retry):   {legal_results['legal_first_try'] + legal_results['legal_with_retry']} ({legal_results['legal_rate_with_retry']:.1%})")
+        print(f"  Always illegal:       {legal_results['illegal_all_retries']} ({legal_results['illegal_rate']:.1%})")
+    
+    # Run win rate evaluation
+    if args.mode in ["winrate", "both"]:
+        print(f"\n" + "=" * 60)
+        print("PHASE 2: WIN RATE EVALUATION")
+        print("=" * 60)
+        print(f"Playing {args.n_games} games against Stockfish (Level {args.stockfish_level})...")
+        
+        winrate_results = evaluator.evaluate(
+            n_games=args.n_games,
+            temperature=args.temperature,
+            verbose=True,
+        )
+        
+        print("\n" + "-" * 40)
+        print("WIN RATE RESULTS")
+        print("-" * 40)
+        print(f"  Wins:   {winrate_results['wins']}")
+        print(f"  Losses: {winrate_results['losses']}")
+        print(f"  Draws:  {winrate_results['draws']}")
+        print(f"\n  Win Rate:  {winrate_results['win_rate']:.1%}")
+        print(f"  Draw Rate: {winrate_results['draw_rate']:.1%}")
+        print(f"  Loss Rate: {winrate_results['loss_rate']:.1%}")
+        print(f"\n  Avg Game Length: {winrate_results['avg_game_length']:.1f} moves")
+        print(f"  Illegal Move Rate: {winrate_results['illegal_move_rate']:.2%}")
+        
+        if winrate_results["estimated_elo"]:
+            print(f"\n  Estimated ELO: {winrate_results['estimated_elo']:.0f}")
+    
+    print("\n" + "=" * 60)
+    print("EVALUATION COMPLETE")
+    print("=" * 60)
+
+
+if __name__ == "__main__":
+    main()

src/model.py

@@ -0,0 +1,437 @@
+"""
+Chess Transformer Model for the Chess Challenge.
+
+This module provides a simple GPT-style transformer architecture
+designed to fit within the 1M parameter constraint.
+
+Key components:
+- ChessConfig: Configuration class for model hyperparameters
+- ChessForCausalLM: The main model class for next-move prediction
+"""
+
+from __future__ import annotations
+
+import math
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import PretrainedConfig, PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+
+class ChessConfig(PretrainedConfig):
+    """
+    Configuration class for the Chess Transformer model.
+    
+    This configuration is designed for a ~1M parameter model.
+    Students can adjust these values to explore different architectures.
+    
+    Parameter budget breakdown (with default values):
+    - Embeddings (vocab): 1200 x 128 = 153,600
+    - Position Embeddings: 256 x 128 = 32,768
+    - Transformer Layers: 6 x ~120,000 = ~720,000
+    - LM Head (with weight tying): 0 (shared with embeddings)
+    - Total: ~906,000 parameters
+    
+    Attributes:
+        vocab_size: Size of the vocabulary (number of unique moves).
+        n_embd: Embedding dimension (d_model).
+        n_layer: Number of transformer layers.
+        n_head: Number of attention heads.
+        n_ctx: Maximum sequence length (context window).
+        n_inner: Feed-forward inner dimension (default: 3 * n_embd).
+        dropout: Dropout probability.
+        layer_norm_epsilon: Epsilon for layer normalization.
+        tie_weights: Whether to tie embedding and output weights.
+    """
+    
+    model_type = "chess_transformer"
+    
+    def __init__(
+        self,
+        vocab_size: int = 1200,
+        n_embd: int = 128,
+        n_layer: int = 6,
+        n_head: int = 4,
+        n_ctx: int = 256,
+        n_inner: Optional[int] = None,
+        dropout: float = 0.1,
+        layer_norm_epsilon: float = 1e-5,
+        tie_weights: bool = True,
+        pad_token_id: int = 0,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        **kwargs,
+    ):
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            **kwargs,
+        )
+        
+        self.vocab_size = vocab_size
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.n_ctx = n_ctx
+        self.n_inner = n_inner if n_inner is not None else 3 * n_embd  # Reduced from 4x to 3x
+        self.dropout = dropout
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.tie_weights = tie_weights
+        # Inform HF base class about tying behavior
+        self.tie_word_embeddings = bool(tie_weights)
+
+
+class MultiHeadAttention(nn.Module):
+    """
+    Multi-head self-attention module.
+    
+    This is a standard scaled dot-product attention implementation
+    with causal masking for autoregressive generation.
+    """
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        
+        assert config.n_embd % config.n_head == 0, \
+            f"n_embd ({config.n_embd}) must be divisible by n_head ({config.n_head})"
+        
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.head_dim = config.n_embd // config.n_head
+        
+        # Combined QKV projection for efficiency
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd)
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd)
+        
+        self.dropout = nn.Dropout(config.dropout)
+        
+        # Causal mask (will be created on first forward pass)
+        self.register_buffer(
+            "bias",
+            torch.tril(torch.ones(config.n_ctx, config.n_ctx)).view(
+                1, 1, config.n_ctx, config.n_ctx
+            ),
+            persistent=False,
+        )
+    
+    def forward(
+        self,
+        x: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, seq_len, _ = x.size()
+        
+        # Compute Q, K, V
+        qkv = self.c_attn(x)
+        q, k, v = qkv.split(self.n_embd, dim=2)
+        
+        # Reshape for multi-head attention
+        q = q.view(batch_size, seq_len, self.n_head, self.head_dim).transpose(1, 2)
+        k = k.view(batch_size, seq_len, self.n_head, self.head_dim).transpose(1, 2)
+        v = v.view(batch_size, seq_len, self.n_head, self.head_dim).transpose(1, 2)
+        
+        # Scaled dot-product attention
+        attn_weights = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
+        
+        # Apply causal mask
+        causal_mask = self.bias[:, :, :seq_len, :seq_len]
+        attn_weights = attn_weights.masked_fill(causal_mask == 0, float("-inf"))
+        
+        # Apply attention mask (for padding)
+        if attention_mask is not None:
+            # attention_mask shape: (batch_size, seq_len) -> (batch_size, 1, 1, seq_len)
+            attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+            attn_weights = attn_weights.masked_fill(attention_mask == 0, float("-inf"))
+        
+        attn_weights = F.softmax(attn_weights, dim=-1)
+        attn_weights = self.dropout(attn_weights)
+        
+        # Apply attention to values
+        attn_output = torch.matmul(attn_weights, v)
+        
+        # Reshape back
+        attn_output = attn_output.transpose(1, 2).contiguous().view(
+            batch_size, seq_len, self.n_embd
+        )
+        
+        # Output projection
+        attn_output = self.c_proj(attn_output)
+        
+        return attn_output
+
+
+class FeedForward(nn.Module):
+    """
+    Feed-forward network (MLP) module.
+    
+    Standard two-layer MLP with GELU activation.
+    """
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        
+        self.c_fc = nn.Linear(config.n_embd, config.n_inner)
+        self.c_proj = nn.Linear(config.n_inner, config.n_embd)
+        self.dropout = nn.Dropout(config.dropout)
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.c_fc(x)
+        x = F.gelu(x)
+        x = self.c_proj(x)
+        x = self.dropout(x)
+        return x
+
+
+class TransformerBlock(nn.Module):
+    """
+    A single transformer block with attention and feed-forward layers.
+    
+    Uses pre-normalization (LayerNorm before attention/FFN) for better
+    training stability.
+    """
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        
+        self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.attn = MultiHeadAttention(config)
+        self.ln_2 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.mlp = FeedForward(config)
+    
+    def forward(
+        self,
+        x: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        # Pre-norm attention
+        x = x + self.attn(self.ln_1(x), attention_mask=attention_mask)
+        # Pre-norm FFN
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class ChessForCausalLM(PreTrainedModel):
+    """
+    Chess Transformer for Causal Language Modeling (next-move prediction).
+    
+    This model is designed to predict the next chess move given a sequence
+    of previous moves. It uses a GPT-style architecture with:
+    - Token embeddings for chess moves
+    - Learned positional embeddings
+    - Stacked transformer blocks
+    - Linear head for next-token prediction
+    
+    The model supports weight tying between the embedding layer and the
+    output projection to save parameters.
+    
+    Example:
+        >>> config = ChessConfig(vocab_size=1200, n_embd=128, n_layer=6)
+        >>> model = ChessForCausalLM(config)
+        >>> inputs = {"input_ids": torch.tensor([[1, 42, 87]])}
+        >>> outputs = model(**inputs)
+        >>> next_move_logits = outputs.logits[:, -1, :]
+    """
+    
+    config_class = ChessConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    # Suppress missing-key warning for tied lm_head when loading
+    keys_to_ignore_on_load_missing = ["lm_head.weight"]
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__(config)
+        
+        # Token and position embeddings
+        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
+        self.wpe = nn.Embedding(config.n_ctx, config.n_embd)
+        
+        self.drop = nn.Dropout(config.dropout)
+        
+        # Transformer blocks
+        self.h = nn.ModuleList([
+            TransformerBlock(config) for _ in range(config.n_layer)
+        ])
+        
+        # Final layer norm
+        self.ln_f = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        
+        # Output head
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+        
+        # Declare tied weights for proper serialization
+        if config.tie_weights:
+            self._tied_weights_keys = ["lm_head.weight"]
+        
+        # Initialize weights
+        self.post_init()
+        
+        # Tie weights if configured
+        if config.tie_weights:
+            self.tie_weights()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.wte
+
+    def set_input_embeddings(self, new_embeddings: nn.Module):
+        self.wte = new_embeddings
+        if getattr(self.config, "tie_weights", False):
+            self.tie_weights()
+
+    def get_output_embeddings(self) -> nn.Module:
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings: nn.Module):
+        self.lm_head = new_embeddings
+
+    def tie_weights(self):
+        # Use HF helper to tie or clone depending on config
+        if getattr(self.config, "tie_weights", False) or getattr(self.config, "tie_word_embeddings", False):
+            self._tie_or_clone_weights(self.lm_head, self.wte)
+    
+    def _init_weights(self, module: nn.Module):
+        """Initialize weights following GPT-2 style."""
+        if isinstance(module, nn.Linear):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
+            if module.bias is not None:
+                torch.nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
+        elif isinstance(module, nn.LayerNorm):
+            torch.nn.init.ones_(module.weight)
+            torch.nn.init.zeros_(module.bias)
+    
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        """
+        Forward pass of the model.
+        
+        Args:
+            input_ids: Token IDs of shape (batch_size, seq_len).
+            attention_mask: Attention mask of shape (batch_size, seq_len).
+            position_ids: Position IDs of shape (batch_size, seq_len).
+            labels: Labels for language modeling loss.
+            return_dict: Whether to return a ModelOutput object.
+        
+        Returns:
+            CausalLMOutputWithPast containing loss (if labels provided) and logits.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        
+        batch_size, seq_len = input_ids.size()
+        device = input_ids.device
+        
+        # Create position IDs if not provided
+        if position_ids is None:
+            position_ids = torch.arange(seq_len, device=device).unsqueeze(0).expand(batch_size, -1)
+        
+        # Get embeddings
+        token_embeds = self.wte(input_ids)
+        position_embeds = self.wpe(position_ids)
+        hidden_states = self.drop(token_embeds + position_embeds)
+        
+        # Pass through transformer blocks
+        for block in self.h:
+            hidden_states = block(hidden_states, attention_mask=attention_mask)
+        
+        # Final layer norm
+        hidden_states = self.ln_f(hidden_states)
+        
+        # Get logits
+        logits = self.lm_head(hidden_states)
+        
+        # Compute loss if labels are provided
+        loss = None
+        if labels is not None:
+            # Shift logits and labels for next-token prediction
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            
+            # Flatten for cross-entropy
+            loss_fct = nn.CrossEntropyLoss(ignore_index=self.config.pad_token_id)
+            loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)),
+                shift_labels.view(-1),
+            )
+        
+        if not return_dict:
+            output = (logits,)
+            return ((loss,) + output) if loss is not None else output
+        
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=None,
+            hidden_states=None,
+            attentions=None,
+        )
+    
+    @torch.no_grad()
+    def generate_move(
+        self,
+        input_ids: torch.LongTensor,
+        temperature: float = 1.0,
+        top_k: Optional[int] = None,
+        top_p: Optional[float] = None,
+    ) -> int:
+        """
+        Generate the next move given a sequence of moves.
+        
+        Args:
+            input_ids: Token IDs of shape (1, seq_len).
+            temperature: Sampling temperature (1.0 = no change).
+            top_k: If set, only sample from top k tokens.
+            top_p: If set, use nucleus sampling with this threshold.
+        
+        Returns:
+            The token ID of the predicted next move.
+        """
+        self.eval()
+        
+        # Get logits for the last position
+        outputs = self(input_ids)
+        logits = outputs.logits[:, -1, :] / temperature
+        
+        # Apply top-k filtering
+        if top_k is not None:
+            indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+            logits[indices_to_remove] = float("-inf")
+        
+        # Apply top-p (nucleus) filtering
+        if top_p is not None:
+            sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+            cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+            
+            # Remove tokens with cumulative probability above the threshold
+            sorted_indices_to_remove = cumulative_probs > top_p
+            sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+            sorted_indices_to_remove[..., 0] = 0
+            
+            indices_to_remove = sorted_indices_to_remove.scatter(
+                dim=-1, index=sorted_indices, src=sorted_indices_to_remove
+            )
+            logits[indices_to_remove] = float("-inf")
+        
+        # Sample from the distribution
+        probs = F.softmax(logits, dim=-1)
+        next_token = torch.multinomial(probs, num_samples=1)
+        
+        return next_token.item()
+
+
+# Register the model with Auto classes for easy loading
+from transformers import AutoConfig, AutoModelForCausalLM
+
+AutoConfig.register("chess_transformer", ChessConfig)
+AutoModelForCausalLM.register(ChessConfig, ChessForCausalLM)

src/tokenizer.py

@@ -0,0 +1,278 @@
+"""
+Custom Chess Tokenizer for the Chess Challenge.
+
+This tokenizer treats each move as a single token using the extended UCI notation
+from the Lichess dataset (e.g., WPe2e4, BNg8f6).
+
+The dataset format uses:
+- W/B prefix for White/Black
+- Piece letter: P=Pawn, N=Knight, B=Bishop, R=Rook, Q=Queen, K=King
+- Source and destination squares (e.g., e2e4)
+- Special suffixes: (x)=capture, (+)=check, (+*)=checkmate, (o)/(O)=castling
+"""
+
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+from typing import Dict, List, Optional
+
+from transformers import PreTrainedTokenizer
+
+
+class ChessTokenizer(PreTrainedTokenizer):
+    """
+    A custom tokenizer for chess moves using extended UCI notation.
+    
+    This tokenizer maps each possible chess move to a unique token ID.
+    The vocabulary is built from the training dataset to ensure all moves
+    encountered during training have a corresponding token.
+    
+    Example:
+        >>> tokenizer = ChessTokenizer()
+        >>> tokenizer.encode("WPe2e4 BPe7e5")
+        [1, 42, 87, 2]  # [BOS, e2e4, e7e5, EOS]
+    """
+    
+    model_input_names = ["input_ids", "attention_mask"]
+    vocab_files_names = {"vocab_file": "vocab.json"}
+    
+    # Special tokens
+    PAD_TOKEN = "[PAD]"
+    BOS_TOKEN = "[BOS]"
+    EOS_TOKEN = "[EOS]"
+    UNK_TOKEN = "[UNK]"
+    
+    def __init__(
+        self,
+        vocab_file: Optional[str] = None,
+        vocab: Optional[Dict[str, int]] = None,
+        **kwargs,
+    ):
+        """
+        Initialize the chess tokenizer.
+        
+        Args:
+            vocab_file: Path to a JSON file containing the vocabulary mapping.
+            vocab: Dictionary mapping tokens to IDs (alternative to vocab_file).
+            **kwargs: Additional arguments passed to PreTrainedTokenizer.
+        """
+        # Initialize special tokens
+        self._pad_token = self.PAD_TOKEN
+        self._bos_token = self.BOS_TOKEN
+        self._eos_token = self.EOS_TOKEN
+        self._unk_token = self.UNK_TOKEN
+
+        # Remove any duplicate special-token entries passed through kwargs
+        # to avoid "multiple values for keyword" errors when loading from disk.
+        kwargs.pop("pad_token", None)
+        kwargs.pop("bos_token", None)
+        kwargs.pop("eos_token", None)
+        kwargs.pop("unk_token", None)
+        
+        # Load or create vocabulary
+        if vocab is not None:
+            self._vocab = vocab
+        elif vocab_file is not None and os.path.exists(vocab_file):
+            with open(vocab_file, "r", encoding="utf-8") as f:
+                self._vocab = json.load(f)
+        else:
+            # Create a minimal vocabulary with just special tokens
+            # The full vocabulary should be built from the dataset
+            self._vocab = self._create_default_vocab()
+        
+        # Create reverse mapping
+        self._ids_to_tokens = {v: k for k, v in self._vocab.items()}
+        
+        # Call parent init AFTER setting up vocab
+        super().__init__(
+            pad_token=self._pad_token,
+            bos_token=self._bos_token,
+            eos_token=self._eos_token,
+            unk_token=self._unk_token,
+            **kwargs,
+        )
+    
+    def _create_default_vocab(self) -> Dict[str, int]:
+        """
+        Create a minimal default vocabulary with just special tokens.
+        
+        For the full vocabulary, use `build_vocab_from_dataset()`.
+        This minimal vocab is just a placeholder - you should build from data.
+        """
+        special_tokens = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]
+        vocab = {token: idx for idx, token in enumerate(special_tokens)}
+        return vocab
+    
+    @classmethod
+    def build_vocab_from_iterator(
+        cls,
+        iterator,
+        min_frequency: int = 1,
+    ) -> "ChessTokenizer":
+        """
+        Build a tokenizer vocabulary from an iterator of game strings.
+        
+        Args:
+            iterator: An iterator yielding game strings (space-separated moves).
+            min_frequency: Minimum frequency for a token to be included.
+        
+        Returns:
+            A ChessTokenizer with the built vocabulary.
+        """
+        from collections import Counter
+        
+        token_counts = Counter()
+        
+        for game in iterator:
+            moves = game.strip().split()
+            token_counts.update(moves)
+        
+        # Filter by frequency
+        tokens = [
+            token for token, count in token_counts.items()
+            if count >= min_frequency
+        ]
+        
+        # Sort for reproducibility
+        tokens = sorted(tokens)
+        
+        # Build vocabulary
+        special_tokens = [cls.PAD_TOKEN, cls.BOS_TOKEN, cls.EOS_TOKEN, cls.UNK_TOKEN]
+        vocab = {token: idx for idx, token in enumerate(special_tokens + tokens)}
+        
+        return cls(vocab=vocab)
+    
+    @classmethod
+    def build_vocab_from_dataset(
+        cls,
+        dataset_name: str = "dlouapre/lichess_2025-01_1M",
+        split: str = "train",
+        column: str = "text",
+        min_frequency: int = 500,
+        max_samples: Optional[int] = 100000,
+    ) -> "ChessTokenizer":
+        """
+        Build a tokenizer vocabulary from a Hugging Face dataset.
+        
+        Args:
+            dataset_name: Name of the dataset on Hugging Face Hub.
+            split: Dataset split to use.
+            column: Column containing the game strings.
+            min_frequency: Minimum frequency for a token to be included (default: 500).
+            max_samples: Maximum number of samples to process (default: 100k).
+        
+        Returns:
+            A ChessTokenizer with the built vocabulary.
+        """
+        from datasets import load_dataset
+        
+        dataset = load_dataset(dataset_name, split=split)
+        
+        if max_samples is not None:
+            dataset = dataset.select(range(min(max_samples, len(dataset))))
+        
+        def game_iterator():
+            for example in dataset:
+                yield example[column]
+        
+        return cls.build_vocab_from_iterator(game_iterator(), min_frequency=min_frequency)
+    
+    @property
+    def vocab_size(self) -> int:
+        """Return the size of the vocabulary."""
+        return len(self._vocab)
+    
+    def get_vocab(self) -> Dict[str, int]:
+        """Return the vocabulary as a dictionary."""
+        return dict(self._vocab)
+    
+    def _tokenize(self, text: str) -> List[str]:
+        """
+        Tokenize a string of moves into a list of tokens.
+        
+        Args:
+            text: A string of space-separated moves.
+        
+        Returns:
+            List of move tokens.
+        """
+        return text.strip().split()
+    
+    def _convert_token_to_id(self, token: str) -> int:
+        """Convert a token to its ID."""
+        return self._vocab.get(token, self._vocab.get(self.UNK_TOKEN, 0))
+    
+    def _convert_id_to_token(self, index: int) -> str:
+        """Convert an ID to its token."""
+        return self._ids_to_tokens.get(index, self.UNK_TOKEN)
+    
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        """Convert a list of tokens back to a string."""
+        # Filter out special tokens for cleaner output
+        special = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}
+        return " ".join(t for t in tokens if t not in special)
+    
+    def save_vocabulary(
+        self,
+        save_directory: str,
+        filename_prefix: Optional[str] = None,
+    ) -> tuple:
+        """
+        Save the vocabulary to a JSON file.
+        
+        Args:
+            save_directory: Directory to save the vocabulary.
+            filename_prefix: Optional prefix for the filename.
+        
+        Returns:
+            Tuple containing the path to the saved vocabulary file.
+        """
+        if not os.path.isdir(save_directory):
+            os.makedirs(save_directory, exist_ok=True)
+        
+        vocab_file = os.path.join(
+            save_directory,
+            (filename_prefix + "-" if filename_prefix else "") + "vocab.json",
+        )
+        
+        with open(vocab_file, "w", encoding="utf-8") as f:
+            json.dump(self._vocab, f, ensure_ascii=False, indent=2)
+        
+        return (vocab_file,)
+
+
+def count_vocab_from_dataset(
+    dataset_name: str = "dlouapre/lichess_2025-01_1M",
+    split: str = "train",
+    column: str = "text",
+    max_samples: Optional[int] = 10000,
+) -> Dict[str, int]:
+    """
+    Count token frequencies in a dataset (useful for vocabulary analysis).
+    
+    Args:
+        dataset_name: Name of the dataset on Hugging Face Hub.
+        split: Dataset split to use.
+        column: Column containing the game strings.
+        max_samples: Maximum number of samples to process.
+    
+    Returns:
+        Dictionary mapping tokens to their frequencies.
+    """
+    from collections import Counter
+    from datasets import load_dataset
+    
+    dataset = load_dataset(dataset_name, split=split)
+    
+    if max_samples is not None:
+        dataset = dataset.select(range(min(max_samples, len(dataset))))
+    
+    token_counts = Counter()
+    
+    for example in dataset:
+        moves = example[column].strip().split()
+        token_counts.update(moves)
+    
+    return dict(token_counts)