Chess Challenge submission by shubhasanket

README.md

@@ -0,0 +1,26 @@
+---
+library_name: transformers
+tags:
+- chess
+- llm-course
+- chess-challenge
+license: mit
+---
+
+# shubha-chess-1m-structured-v2
+
+Chess model submitted to the LLM Course Chess Challenge.
+
+## Submission Info
+
+- **Submitted by**: [shubhasanket](https://huggingface.co/shubhasanket)
+- **Parameters**: 986,576
+- **Organization**: LLM-course
+
+## Model Details
+
+- **Architecture**: Chess Transformer (GPT-style)
+- **Vocab size**: 143
+- **Embedding dim**: 144
+- **Layers**: 4
+- **Heads**: 8

config.json

@@ -0,0 +1,20 @@
+{
+  "architectures": [
+    "ChessForCausalLM"
+  ],
+  "bos_token_id": 1,
+  "dropout": 0.1,
+  "dtype": "float32",
+  "eos_token_id": 2,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "chess_transformer",
+  "n_ctx": 256,
+  "n_embd": 144,
+  "n_head": 8,
+  "n_inner": 512,
+  "n_layer": 4,
+  "pad_token_id": 0,
+  "tie_weights": true,
+  "transformers_version": "4.57.6",
+  "vocab_size": 143
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2115203092db9b17ff723b1c4f00bd4a534189fda6d931849f733ae1ccf350fa
+size 3950720

special_tokens_map.json

@@ -0,0 +1,6 @@
+{
+  "bos_token": "[BOS]",
+  "eos_token": "[EOS]",
+  "pad_token": "[PAD]",
+  "unk_token": "[UNK]"
+}

tokenizer.py

@@ -0,0 +1,446 @@
+# """
+# Improved Chess Tokenizer (Structured) for the Chess Challenge.
+
+# Key idea:
+# - Decompose each move into sub-tokens:
+#   SIDE_W / SIDE_B
+#   piece (P,N,B,R,Q,K)
+#   from-square (e2)
+#   to-square (e4)
+#   optional flags: CAPTURE, CHECK, MATE, CASTLE_SHORT, CASTLE_LONG
+
+# Important implementation detail:
+# - We MUST avoid token-string collisions. In particular, "B" is both:
+#   - Black side ("B")
+#   - Bishop piece ("B")
+#   If we used raw "W"/"B" for side, the vocab dict would overwrite one of them,
+#   creating non-contiguous IDs and leading to embedding "index out of range".
+# """
+
+# from __future__ import annotations
+
+# import json
+# import os
+# import re
+# from typing import Dict, List, Optional
+
+# from transformers import PreTrainedTokenizer
+
+
+# class ChessTokenizer(PreTrainedTokenizer):
+#     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]"
+
+#     # Side tokens (avoid collision with piece "B" for Bishop)
+#     SIDE_W = "SIDE_W"
+#     SIDE_B = "SIDE_B"
+
+#     def __init__(
+#         self,
+#         vocab_file: Optional[str] = None,
+#         vocab: Optional[Dict[str, int]] = None,
+#         **kwargs,
+#     ):
+#         self._pad_token = self.PAD_TOKEN
+#         self._bos_token = self.BOS_TOKEN
+#         self._eos_token = self.EOS_TOKEN
+#         self._unk_token = self.UNK_TOKEN
+
+#         # Avoid duplicate kwargs when HF loads from disk
+#         kwargs.pop("pad_token", None)
+#         kwargs.pop("bos_token", None)
+#         kwargs.pop("eos_token", None)
+#         kwargs.pop("unk_token", None)
+
+#         if vocab is not None:
+#             self._vocab = {str(k): int(v) for k, v in vocab.items()}
+#         elif vocab_file is not None and os.path.exists(vocab_file):
+#             with open(vocab_file, "r", encoding="utf-8") as f:
+#                 loaded = json.load(f)
+#             self._vocab = {str(k): int(v) for k, v in loaded.items()}
+#         else:
+#             self._vocab = self._create_default_vocab()
+
+#         # Ensure IDs are contiguous 0..(len-1) (robust to any old saved vocabs)
+#         self._vocab = self._normalize_vocab_ids(self._vocab)
+#         self._ids_to_tokens = {v: k for k, v in self._vocab.items()}
+
+#         super().__init__(
+#             pad_token=self._pad_token,
+#             bos_token=self._bos_token,
+#             eos_token=self._eos_token,
+#             unk_token=self._unk_token,
+#             **kwargs,
+#         )
+
+#     @staticmethod
+#     def _normalize_vocab_ids(vocab: Dict[str, int]) -> Dict[str, int]:
+#         """
+#         Re-map token IDs to be contiguous and deterministic.
+#         Sort by old id then by token string.
+#         """
+#         items = sorted(vocab.items(), key=lambda kv: (kv[1], kv[0]))
+#         return {tok: new_id for new_id, (tok, _) in enumerate(items)}
+
+#     # ------------------------------------------------------------------
+#     # REQUIRED compatibility method (train.py expects this to exist)
+#     # ------------------------------------------------------------------
+#     @classmethod
+#     def build_vocab_from_dataset(
+#         cls,
+#         dataset_name: str = "dlouapre/lichess_2025-01_1M",
+#         split: str = "train",
+#         column: str = "text",
+#         min_frequency: int = 1,
+#         max_samples: Optional[int] = None,
+#     ) -> "ChessTokenizer":
+#         """
+#         Compatibility hook.
+
+#         For the structured tokenizer, the vocabulary is fixed and does not
+#         depend on dataset statistics. We keep this method so src/train.py
+#         (template code) does not need to change.
+#         """
+#         return cls()
+
+#     # ------------------------------------------------------------------
+#     # Vocabulary construction
+#     # ------------------------------------------------------------------
+#     def _create_default_vocab(self) -> Dict[str, int]:
+#         special = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]
+
+#         sides = [self.SIDE_W, self.SIDE_B]  # no collision with piece tokens
+#         pieces = ["P", "N", "B", "R", "Q", "K"]
+
+#         files = list("abcdefgh")
+#         ranks = list("12345678")
+#         squares = [f + r for f in files for r in ranks]  # 64 tokens
+
+#         flags = ["CAPTURE", "CHECK", "MATE", "CASTLE_SHORT", "CASTLE_LONG"]
+
+#         tokens = special + sides + pieces + squares + flags
+#         return {tok: i for i, tok in enumerate(tokens)}  # contiguous by construction
+
+#     @property
+#     def vocab_size(self) -> int:
+#         return len(self._vocab)
+
+#     def get_vocab(self) -> Dict[str, int]:
+#         return dict(self._vocab)
+
+#     # ------------------------------------------------------------------
+#     # Tokenization logic
+#     # ------------------------------------------------------------------
+#     MOVE_REGEX = re.compile(
+#         r"""
+#         (?P<side>[WB])
+#         (?P<piece>[PNBRQK])
+#         (?P<from>[a-h][1-8])
+#         (?P<to>[a-h][1-8])
+#         (?P<suffix>.*)?
+#         """,
+#         re.VERBOSE,
+#     )
+
+#     def _tokenize(self, text: str) -> List[str]:
+#         out: List[str] = []
+#         for move in text.strip().split():
+#             out.extend(self._decompose_move(move))
+#         return out
+
+#     def _decompose_move(self, move: str) -> List[str]:
+#         m = self.MOVE_REGEX.match(move)
+#         if not m:
+#             return [self.UNK_TOKEN]
+
+#         side_raw = m.group("side")
+#         side_tok = self.SIDE_W if side_raw == "W" else self.SIDE_B
+
+#         tokens = [
+#             side_tok,
+#             m.group("piece"),
+#             m.group("from"),
+#             m.group("to"),
+#         ]
+
+#         suffix = m.group("suffix") or ""
+
+#         if "(x)" in suffix:
+#             tokens.append("CAPTURE")
+#         if "(+*)" in suffix:
+#             tokens.append("MATE")
+#         elif "(+)" in suffix:
+#             tokens.append("CHECK")
+#         if "(o)" in suffix:
+#             tokens.append("CASTLE_SHORT")
+#         if "(O)" in suffix:
+#             tokens.append("CASTLE_LONG")
+
+#         return tokens
+
+#     # ------------------------------------------------------------------
+#     # ID conversion
+#     # ------------------------------------------------------------------
+#     def _convert_token_to_id(self, token: str) -> int:
+#         return self._vocab.get(token, self._vocab[self.UNK_TOKEN])
+
+#     def _convert_id_to_token(self, index: int) -> str:
+#         return self._ids_to_tokens.get(index, self.UNK_TOKEN)
+
+#     def convert_tokens_to_string(self, tokens: List[str]) -> str:
+#         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)
+
+#     # ------------------------------------------------------------------
+#     # Saving
+#     # ------------------------------------------------------------------
+#     def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> tuple:
+#         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, 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 after structured tokenization.
+#     (Editor warning about 'datasets' can be ignored if terminal run works.)
+#     """
+#     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))))
+
+#     tok = ChessTokenizer()
+#     counts = Counter()
+
+#     for ex in dataset:
+#         counts.update(tok._tokenize(ex[column]))
+
+#     return dict(counts)
+
+"""
+Final Structured Chess Tokenizer for the Chess Challenge.
+
+Design goals:
+- Strong legality bias
+- Fixed, collision-free vocabulary
+- HF-compatible (Trainer, save/load, Hub)
+- Evaluator-friendly (square extraction still works)
+
+Move decomposition:
+  PIECE
+  FROM_<square>
+  TO_<square>
+  optional FLAGS
+
+Example:
+  P FROM_e2 TO_e4
+  N FROM_g1 TO_f3 CHECK
+  K FROM_e1 TO_g1 CASTLE_SHORT
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+from typing import Dict, List, Optional
+
+from transformers import PreTrainedTokenizer
+
+
+class ChessTokenizer(PreTrainedTokenizer):
+    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]"
+
+    # Fixed role prefixes
+    FROM_PREFIX = "FROM_"
+    TO_PREFIX = "TO_"
+
+    def __init__(
+        self,
+        vocab_file: Optional[str] = None,
+        vocab: Optional[Dict[str, int]] = None,
+        **kwargs,
+    ):
+        self._pad_token = self.PAD_TOKEN
+        self._bos_token = self.BOS_TOKEN
+        self._eos_token = self.EOS_TOKEN
+        self._unk_token = self.UNK_TOKEN
+
+        # Avoid duplicate kwargs when loading
+        for k in ("pad_token", "bos_token", "eos_token", "unk_token"):
+            kwargs.pop(k, None)
+
+        if vocab is not None:
+            self._vocab = {str(k): int(v) for k, v in vocab.items()}
+        elif vocab_file and os.path.exists(vocab_file):
+            with open(vocab_file, "r", encoding="utf-8") as f:
+                self._vocab = {str(k): int(v) for k, v in json.load(f).items()}
+        else:
+            self._vocab = self._create_default_vocab()
+
+        # Ensure contiguous IDs
+        self._vocab = self._normalize_vocab(self._vocab)
+        self._ids_to_tokens = {v: k for k, v in self._vocab.items()}
+
+        super().__init__(
+            pad_token=self._pad_token,
+            bos_token=self._bos_token,
+            eos_token=self._eos_token,
+            unk_token=self._unk_token,
+            **kwargs,
+        )
+
+    @staticmethod
+    def _normalize_vocab(vocab: Dict[str, int]) -> Dict[str, int]:
+        items = sorted(vocab.items(), key=lambda kv: (kv[1], kv[0]))
+        return {tok: i for i, (tok, _) in enumerate(items)}
+
+    # ------------------------------------------------------------
+    # Required by train.py (kept for compatibility)
+    # ------------------------------------------------------------
+    @classmethod
+    def build_vocab_from_dataset(
+        cls,
+        *args,
+        **kwargs,
+    ) -> "ChessTokenizer":
+        return cls()
+
+    # ------------------------------------------------------------
+    # Vocabulary
+    # ------------------------------------------------------------
+    def _create_default_vocab(self) -> Dict[str, int]:
+        special = [
+            self.PAD_TOKEN,
+            self.BOS_TOKEN,
+            self.EOS_TOKEN,
+            self.UNK_TOKEN,
+        ]
+
+        pieces = ["P", "N", "B", "R", "Q", "K"]
+
+        files = "abcdefgh"
+        ranks = "12345678"
+        squares = [f + r for f in files for r in ranks]
+
+        from_tokens = [self.FROM_PREFIX + sq for sq in squares]
+        to_tokens = [self.TO_PREFIX + sq for sq in squares]
+
+        flags = [
+            "CAPTURE",
+            "CHECK",
+            "MATE",
+            "CASTLE_SHORT",
+            "CASTLE_LONG",
+        ]
+
+        tokens = special + pieces + from_tokens + to_tokens + flags
+        return {tok: i for i, tok in enumerate(tokens)}
+
+    @property
+    def vocab_size(self) -> int:
+        return len(self._vocab)
+
+    def get_vocab(self) -> Dict[str, int]:
+        return dict(self._vocab)
+
+    # ------------------------------------------------------------
+    # Tokenization
+    # ------------------------------------------------------------
+    MOVE_REGEX = re.compile(
+        r"""
+        (?P<piece>[PNBRQK])
+        (?P<from>[a-h][1-8])
+        (?P<to>[a-h][1-8])
+        (?P<suffix>.*)?
+        """,
+        re.VERBOSE,
+    )
+
+    def _tokenize(self, text: str) -> List[str]:
+        out: List[str] = []
+        for move in text.strip().split():
+            out.extend(self._decompose_move(move))
+        return out
+
+    def _decompose_move(self, move: str) -> List[str]:
+        m = self.MOVE_REGEX.search(move)
+        if not m:
+            return [self.UNK_TOKEN]
+
+        tokens = [
+            m.group("piece"),
+            self.FROM_PREFIX + m.group("from"),
+            self.TO_PREFIX + m.group("to"),
+        ]
+
+        suffix = m.group("suffix") or ""
+
+        if "(x)" in suffix:
+            tokens.append("CAPTURE")
+        if "(+*)" in suffix:
+            tokens.append("MATE")
+        elif "(+)" in suffix:
+            tokens.append("CHECK")
+        if "(o)" in suffix:
+            tokens.append("CASTLE_SHORT")
+        if "(O)" in suffix:
+            tokens.append("CASTLE_LONG")
+
+        return tokens
+
+    # ------------------------------------------------------------
+    # ID conversion
+    # ------------------------------------------------------------
+    def _convert_token_to_id(self, token: str) -> int:
+        return self._vocab.get(token, self._vocab[self.UNK_TOKEN])
+
+    def _convert_id_to_token(self, index: int) -> str:
+        return self._ids_to_tokens.get(index, self.UNK_TOKEN)
+
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        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)
+
+    # ------------------------------------------------------------
+    # Saving
+    # ------------------------------------------------------------
+    def save_vocabulary(
+        self,
+        save_directory: str,
+        filename_prefix: Optional[str] = None,
+    ) -> tuple:
+        os.makedirs(save_directory, exist_ok=True)
+        path = os.path.join(
+            save_directory,
+            (filename_prefix + "-" if filename_prefix else "") + "vocab.json",
+        )
+        with open(path, "w", encoding="utf-8") as f:
+            json.dump(self._vocab, f, indent=2)
+        return (path,)

tokenizer_config.json

@@ -0,0 +1,50 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "[BOS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "[EOS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenizer.ChessTokenizer",
+      null
+    ]
+  },
+  "bos_token": "[BOS]",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "[EOS]",
+  "extra_special_tokens": {},
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "[PAD]",
+  "tokenizer_class": "ChessTokenizer",
+  "unk_token": "[UNK]"
+}

vocab.json

@@ -0,0 +1,145 @@
+{
+  "[PAD]": 0,
+  "[BOS]": 1,
+  "[EOS]": 2,
+  "[UNK]": 3,
+  "P": 4,
+  "N": 5,
+  "B": 6,
+  "R": 7,
+  "Q": 8,
+  "K": 9,
+  "FROM_a1": 10,
+  "FROM_a2": 11,
+  "FROM_a3": 12,
+  "FROM_a4": 13,
+  "FROM_a5": 14,
+  "FROM_a6": 15,
+  "FROM_a7": 16,
+  "FROM_a8": 17,
+  "FROM_b1": 18,
+  "FROM_b2": 19,
+  "FROM_b3": 20,
+  "FROM_b4": 21,
+  "FROM_b5": 22,
+  "FROM_b6": 23,
+  "FROM_b7": 24,
+  "FROM_b8": 25,
+  "FROM_c1": 26,
+  "FROM_c2": 27,
+  "FROM_c3": 28,
+  "FROM_c4": 29,
+  "FROM_c5": 30,
+  "FROM_c6": 31,
+  "FROM_c7": 32,
+  "FROM_c8": 33,
+  "FROM_d1": 34,
+  "FROM_d2": 35,
+  "FROM_d3": 36,
+  "FROM_d4": 37,
+  "FROM_d5": 38,
+  "FROM_d6": 39,
+  "FROM_d7": 40,
+  "FROM_d8": 41,
+  "FROM_e1": 42,
+  "FROM_e2": 43,
+  "FROM_e3": 44,
+  "FROM_e4": 45,
+  "FROM_e5": 46,
+  "FROM_e6": 47,
+  "FROM_e7": 48,
+  "FROM_e8": 49,
+  "FROM_f1": 50,
+  "FROM_f2": 51,
+  "FROM_f3": 52,
+  "FROM_f4": 53,
+  "FROM_f5": 54,
+  "FROM_f6": 55,
+  "FROM_f7": 56,
+  "FROM_f8": 57,
+  "FROM_g1": 58,
+  "FROM_g2": 59,
+  "FROM_g3": 60,
+  "FROM_g4": 61,
+  "FROM_g5": 62,
+  "FROM_g6": 63,
+  "FROM_g7": 64,
+  "FROM_g8": 65,
+  "FROM_h1": 66,
+  "FROM_h2": 67,
+  "FROM_h3": 68,
+  "FROM_h4": 69,
+  "FROM_h5": 70,
+  "FROM_h6": 71,
+  "FROM_h7": 72,
+  "FROM_h8": 73,
+  "TO_a1": 74,
+  "TO_a2": 75,
+  "TO_a3": 76,
+  "TO_a4": 77,
+  "TO_a5": 78,
+  "TO_a6": 79,
+  "TO_a7": 80,
+  "TO_a8": 81,
+  "TO_b1": 82,
+  "TO_b2": 83,
+  "TO_b3": 84,
+  "TO_b4": 85,
+  "TO_b5": 86,
+  "TO_b6": 87,
+  "TO_b7": 88,
+  "TO_b8": 89,
+  "TO_c1": 90,
+  "TO_c2": 91,
+  "TO_c3": 92,
+  "TO_c4": 93,
+  "TO_c5": 94,
+  "TO_c6": 95,
+  "TO_c7": 96,
+  "TO_c8": 97,
+  "TO_d1": 98,
+  "TO_d2": 99,
+  "TO_d3": 100,
+  "TO_d4": 101,
+  "TO_d5": 102,
+  "TO_d6": 103,
+  "TO_d7": 104,
+  "TO_d8": 105,
+  "TO_e1": 106,
+  "TO_e2": 107,
+  "TO_e3": 108,
+  "TO_e4": 109,
+  "TO_e5": 110,
+  "TO_e6": 111,
+  "TO_e7": 112,
+  "TO_e8": 113,
+  "TO_f1": 114,
+  "TO_f2": 115,
+  "TO_f3": 116,
+  "TO_f4": 117,
+  "TO_f5": 118,
+  "TO_f6": 119,
+  "TO_f7": 120,
+  "TO_f8": 121,
+  "TO_g1": 122,
+  "TO_g2": 123,
+  "TO_g3": 124,
+  "TO_g4": 125,
+  "TO_g5": 126,
+  "TO_g6": 127,
+  "TO_g7": 128,
+  "TO_g8": 129,
+  "TO_h1": 130,
+  "TO_h2": 131,
+  "TO_h3": 132,
+  "TO_h4": 133,
+  "TO_h5": 134,
+  "TO_h6": 135,
+  "TO_h7": 136,
+  "TO_h8": 137,
+  "CAPTURE": 138,
+  "CHECK": 139,
+  "MATE": 140,
+  "CASTLE_SHORT": 141,
+  "CASTLE_LONG": 142
+}