Chess Challenge submission by Sammy972

README.md

@@ -0,0 +1,26 @@
+---
+library_name: transformers
+tags:
+- chess
+- llm-course
+- chess-challenge
+license: mit
+---
+
+# chess-sam
+
+Chess model submitted to the LLM Course Chess Challenge.
+
+## Submission Info
+
+- **Submitted by**: [Sammy972](https://huggingface.co/Sammy972)
+- **Parameters**: 952,224
+- **Organization**: LLM-course
+
+## Model Details
+
+- **Architecture**: Chess Transformer (GPT-style)
+- **Vocab size**: 85
+- **Embedding dim**: 96
+- **Layers**: 8
+- **Heads**: 4

config.json

@@ -0,0 +1,21 @@
+{
+  "_name_or_path": "output/final_model",
+  "architectures": [
+    "ChessForCausalLM"
+  ],
+  "bos_token_id": 1,
+  "dropout": 0.1,
+  "eos_token_id": 2,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "chess_transformer",
+  "n_ctx": 512,
+  "n_embd": 96,
+  "n_head": 4,
+  "n_inner": 384,
+  "n_layer": 8,
+  "pad_token_id": 0,
+  "tie_weights": true,
+  "torch_dtype": "float32",
+  "transformers_version": "4.46.3",
+  "vocab_size": 85
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12b6170fa69d1033fe9b7f80029916d8bf495340311a739fd6c21a5954079e10
+size 3817288

special_tokens_map.json

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

tokenizer.py

@@ -0,0 +1,261 @@
+# =========================
+# CELLULE 1 — src/tokenizer.py
+# ➜ Remplace TOUT le contenu du fichier par ce code
+# =========================
+
+"""
+Factorized Chess Tokenizer for the Chess Challenge.
+
+Instead of "1 move = 1 token", we represent a move as multiple tokens:
+- Side: [W] / [B]
+- Piece: [P], [N], [BISHOP], [R], [Q], [K]
+- Squares: [e2], [e4], ...
+- Optional suffix: [x], [+], [#], [O-O], [O-O-O]
+- Optional promotion: [prom_Q], [prom_R], [prom_B], [prom_N]
+
+Important:
+- We KEEP squares as tokens so evaluation (regex [a-h][1-8]) can extract UCI moves.
+- We decode squares to plain "e2" etc, and promotions to "q/r/b/n" so evaluate.py can detect promotions.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+from typing import Dict, List, Optional
+
+from transformers import PreTrainedTokenizer
+
+
+MOVE_RE = re.compile(
+    r"^(?P<side>[WB])"
+    r"(?P<piece>[PNBRQK])"
+    r"(?P<src>[a-h][1-8])"
+    r"(?P<dst>[a-h][1-8])"
+    r"(?P<suffix>.*)$"
+)
+
+SQUARE_TOKEN_RE = re.compile(r"^\[[a-h][1-8]\]$")
+
+
+
+class ChessTokenizer(PreTrainedTokenizer):
+    model_input_names = ["input_ids", "attention_mask"]
+    vocab_files_names = {"vocab_file": "vocab.json"}
+
+    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,
+    ):
+        # 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
+
+        # Avoid duplicate kwargs
+        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 = 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:
+            self._vocab = self._create_default_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,
+        )
+
+    def _create_default_vocab(self) -> Dict[str, int]:
+        special_tokens = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]
+
+        side_tokens = ["[W]", "[B]"]
+        piece_tokens = ["[P]", "[N]", "[BISHOP]", "[R]", "[Q]", "[K]"]
+
+        square_tokens = [f"[{file}{rank}]" for rank in "12345678" for file in "abcdefgh"]
+
+        suffix_tokens = [
+            "[x]", "[+]", "[#]",
+            "[O-O]", "[O-O-O]",
+            "[prom_Q]", "[prom_R]", "[prom_B]", "[prom_N]",
+        ]
+
+        vocab_list = special_tokens + side_tokens + piece_tokens + square_tokens + suffix_tokens
+        return {tok: i for i, tok in enumerate(vocab_list)}
+
+    # IMPORTANT: prevent HF from auto-adding BOS/EOS on top of your text
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        if token_ids_1 is None:
+            return token_ids_0
+        return token_ids_0 + token_ids_1
+
+    @classmethod
+    def build_vocab_from_iterator(cls, iterator, min_frequency: int = 1) -> "ChessTokenizer":
+        # Fixed vocab (we ignore dataset frequency)
+        return cls()
+
+    @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":
+        # Fixed vocab (we ignore dataset frequency)
+        return cls()
+
+    @property
+    def vocab_size(self) -> int:
+        return len(self._vocab)
+
+    def get_vocab(self) -> Dict[str, int]:
+        return dict(self._vocab)
+
+    def _tokenize(self, text: str) -> List[str]:
+        tokens: List[str] = []
+        parts = str(text).strip().split()
+
+        specials = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}
+
+        for p in parts:
+            if p in specials:
+                tokens.append(p)
+                continue
+
+            m = MOVE_RE.match(p)
+            if not m:
+                tokens.append(self.UNK_TOKEN)
+                continue
+
+            side = m.group("side")
+            piece = m.group("piece")
+            src = m.group("src")
+            dst = m.group("dst")
+            suffix = m.group("suffix") or ""
+
+            tokens.append("[W]" if side == "W" else "[B]")
+
+            if piece == "B":
+                tokens.append("[BISHOP]")
+            else:
+                tokens.append(f"[{piece}]")
+
+            tokens.append(f"[{src}]")
+            tokens.append(f"[{dst}]")
+
+            # capture/check/checkmate
+            if "x" in suffix:
+                tokens.append("[x]")
+            if "*" in suffix:
+                tokens.append("[#]")
+            elif "+" in suffix:
+                tokens.append("[+]")
+
+            # castling annotation (optional, squares already encode it)
+            if piece == "K":
+                if (src, dst) in (("e1", "g1"), ("e8", "g8")) or "(o)" in suffix:
+                    tokens.append("[O-O]")
+                elif (src, dst) in (("e1", "c1"), ("e8", "c8")) or "(O)" in suffix:
+                    tokens.append("[O-O-O]")
+
+            # promotion
+            if "=" in suffix:
+                i = suffix.find("=")
+                if i != -1 and i + 1 < len(suffix):
+                    promo = suffix[i + 1].upper()
+                    if promo in ("Q", "R", "B", "N"):
+                        tokens.append(f"[prom_{promo}]")
+
+        return tokens
+
+    def _convert_token_to_id(self, token: str) -> int:
+        return self._vocab.get(token, self._vocab.get(self.UNK_TOKEN, 3))
+
+    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:
+        """
+        Decode tokens into a compact string so evaluate.py can extract squares easily.
+
+        Examples:
+          [W] [P] [e2] [e4]          -> "WPe2e4"
+          ... [e7] [e8] [prom_Q]     -> "WPe7e8q"   (promotion detectable)
+        """
+        out: List[str] = []
+        for t in tokens:
+            if t in (self.PAD_TOKEN,):
+                continue
+
+            # keep these literal so evaluator can compare EOS if needed
+            if t in (self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN):
+                out.append(t)
+                continue
+
+            if t == "[W]":
+                out.append("W")
+            elif t == "[B]":
+                out.append("B")
+            elif t == "[BISHOP]":
+                out.append("B")
+            elif t in ("[P]", "[N]", "[R]", "[Q]", "[K]"):
+                out.append(t.strip("[]"))
+            elif SQUARE_TOKEN_RE.match(t):
+                out.append(t[1:-1])  # "[e2]" -> "e2"
+            elif t == "[x]":
+                out.append("(x)")
+            elif t == "[+]":
+                out.append("(+)")
+            elif t == "[#]":
+                out.append("(+*)")
+            elif t == "[O-O]":
+                out.append("(o)")
+            elif t == "[O-O-O]":
+                out.append("(O)")
+            elif t == "[prom_Q]":
+                out.append("q")
+            elif t == "[prom_R]":
+                out.append("r")
+            elif t == "[prom_B]":
+                out.append("b")
+            elif t == "[prom_N]":
+                out.append("n")
+            else:
+                out.append(t)
+
+        return "".join(out)
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> tuple:
+        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,)

tokenizer_config.json

@@ -0,0 +1,49 @@
+{
+  "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]",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "[PAD]",
+  "tokenizer_class": "ChessTokenizer",
+  "unk_token": "[UNK]"
+}

vocab.json

@@ -0,0 +1,87 @@
+{
+  "[PAD]": 0,
+  "[BOS]": 1,
+  "[EOS]": 2,
+  "[UNK]": 3,
+  "[W]": 4,
+  "[B]": 5,
+  "[P]": 6,
+  "[N]": 7,
+  "[BISHOP]": 8,
+  "[R]": 9,
+  "[Q]": 10,
+  "[K]": 11,
+  "[a1]": 12,
+  "[b1]": 13,
+  "[c1]": 14,
+  "[d1]": 15,
+  "[e1]": 16,
+  "[f1]": 17,
+  "[g1]": 18,
+  "[h1]": 19,
+  "[a2]": 20,
+  "[b2]": 21,
+  "[c2]": 22,
+  "[d2]": 23,
+  "[e2]": 24,
+  "[f2]": 25,
+  "[g2]": 26,
+  "[h2]": 27,
+  "[a3]": 28,
+  "[b3]": 29,
+  "[c3]": 30,
+  "[d3]": 31,
+  "[e3]": 32,
+  "[f3]": 33,
+  "[g3]": 34,
+  "[h3]": 35,
+  "[a4]": 36,
+  "[b4]": 37,
+  "[c4]": 38,
+  "[d4]": 39,
+  "[e4]": 40,
+  "[f4]": 41,
+  "[g4]": 42,
+  "[h4]": 43,
+  "[a5]": 44,
+  "[b5]": 45,
+  "[c5]": 46,
+  "[d5]": 47,
+  "[e5]": 48,
+  "[f5]": 49,
+  "[g5]": 50,
+  "[h5]": 51,
+  "[a6]": 52,
+  "[b6]": 53,
+  "[c6]": 54,
+  "[d6]": 55,
+  "[e6]": 56,
+  "[f6]": 57,
+  "[g6]": 58,
+  "[h6]": 59,
+  "[a7]": 60,
+  "[b7]": 61,
+  "[c7]": 62,
+  "[d7]": 63,
+  "[e7]": 64,
+  "[f7]": 65,
+  "[g7]": 66,
+  "[h7]": 67,
+  "[a8]": 68,
+  "[b8]": 69,
+  "[c8]": 70,
+  "[d8]": 71,
+  "[e8]": 72,
+  "[f8]": 73,
+  "[g8]": 74,
+  "[h8]": 75,
+  "[x]": 76,
+  "[+]": 77,
+  "[#]": 78,
+  "[O-O]": 79,
+  "[O-O-O]": 80,
+  "[prom_Q]": 81,
+  "[prom_R]": 82,
+  "[prom_B]": 83,
+  "[prom_N]": 84
+}