Chess Challenge submission by MaximeMuhlethaler

README.md

@@ -0,0 +1,26 @@
+---
+library_name: transformers
+tags:
+- chess
+- llm-course
+- chess-challenge
+license: mit
+---
+
+# chess_MaximeMuh3
+
+Chess model submitted to the LLM Course Chess Challenge.
+
+## Submission Info
+
+- **Submitted by**: [MaximeMuhlethaler](https://huggingface.co/MaximeMuhlethaler)
+- **Parameters**: 997,376
+- **Organization**: LLM-course
+
+## Model Details
+
+- **Architecture**: Chess Transformer (GPT-style)
+- **Vocab size**: 72
+- **Embedding dim**: 128
+- **Layers**: 7
+- **Heads**: 8

config.json

@@ -0,0 +1,25 @@
+{
+  "architectures": [
+    "ChessForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "model.ChessConfig",
+    "AutoModelForCausalLM": "model.ChessForCausalLM"
+  },
+  "bias": false,
+  "bos_token_id": 1,
+  "dropout": 0.05,
+  "dtype": "float32",
+  "eos_token_id": 2,
+  "model_type": "chess_transformer",
+  "n_ctx": 256,
+  "n_embd": 128,
+  "n_head": 8,
+  "n_inner": 275,
+  "n_layer": 7,
+  "pad_token_id": 0,
+  "tie_weights": true,
+  "transformers_version": "4.57.6",
+  "unk_token_id": 3,
+  "vocab_size": 72
+}

model.py

@@ -0,0 +1,141 @@
+"""
+Optimized Chess Transformer
+Strategy: Deep & Narrow.
+Max parameters dedicated to reasoning (Layers), minimal for Vocab.
+"""
+from __future__ import annotations
+import math
+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):
+    model_type = "chess_transformer"
+    
+    def __init__(
+        self,
+        vocab_size=80,    
+        n_embd=128,       
+        n_layer=10,  
+        n_head=8,
+        n_ctx=256,    
+        n_inner=None,       
+        dropout=0.0,        
+        bias=False,     
+        tie_weights=True,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        unk_token_id=3,
+        **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 4 * n_embd
+        self.dropout = dropout
+        self.bias = bias
+        self.tie_weights = tie_weights
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, unk_token_id=unk_token_id, **kwargs)
+
+class MultiHeadAttention(nn.Module):
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.head_dim = config.n_embd // config.n_head
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
+        self.dropout = nn.Dropout(config.dropout)
+        self.register_buffer("bias_mask", 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, attention_mask=None):
+        B, T, C = x.size()
+        qkv = self.c_attn(x)
+        q, k, v = qkv.split(self.n_embd, dim=2)
+        q = q.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
+        k = k.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
+        v = v.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
+        att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(self.head_dim))
+        att = att.masked_fill(self.bias_mask[:,:,:T,:T] == 0, float('-inf'))
+        if attention_mask is not None: att = att.masked_fill(attention_mask.view(B, 1, 1, T) == 0, float('-inf'))
+        att = F.softmax(att, dim=-1)
+        att = self.dropout(att)
+        y = att @ v
+        y = y.transpose(1, 2).contiguous().view(B, T, C)
+        return self.c_proj(y)
+
+class FeedForward(nn.Module):
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        self.c_fc = nn.Linear(config.n_embd, config.n_inner, bias=config.bias)
+        self.c_proj = nn.Linear(config.n_inner, config.n_embd, bias=config.bias)
+        self.dropout = nn.Dropout(config.dropout)
+    def forward(self, x):
+        return self.dropout(self.c_proj(F.gelu(self.c_fc(x))))
+
+class TransformerBlock(nn.Module):
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        self.ln_1 = nn.LayerNorm(config.n_embd)
+        self.attn = MultiHeadAttention(config)
+        self.ln_2 = nn.LayerNorm(config.n_embd)
+        self.mlp = FeedForward(config)
+    def forward(self, x, attention_mask=None):
+        x = x + self.attn(self.ln_1(x), attention_mask)
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+class ChessForCausalLM(PreTrainedModel):
+    config_class = ChessConfig
+    base_model_prefix = "transformer"
+    def __init__(self, config: ChessConfig):
+        super().__init__(config)
+        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)
+        self.h = nn.ModuleList([TransformerBlock(config) for _ in range(config.n_layer)])
+        self.ln_f = nn.LayerNorm(config.n_embd)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+        if config.tie_weights: self.lm_head.weight = self.wte.weight
+        self.post_init()
+
+    def get_input_embeddings(self): return self.wte
+    def set_input_embeddings(self, new): self.wte = new
+    def get_output_embeddings(self): return self.lm_head
+    def set_output_embeddings(self, new): self.lm_head = new
+
+    def forward(self, input_ids, attention_mask=None, position_ids=None, labels=None, return_dict=None, **kwargs):
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        if return_dict is None: return_dict = True
+        device = input_ids.device
+        b, t = input_ids.size()
+        if position_ids is None: position_ids = torch.arange(t, device=device).unsqueeze(0)
+        x = self.wte(input_ids) + self.wpe(position_ids)
+        x = self.drop(x)
+        for block in self.h: x = block(x, attention_mask)
+        x = self.ln_f(x)
+        logits = self.lm_head(x)
+        
+        # Protection génération
+        if labels is None:
+             logits[:, :, [self.config.pad_token_id, self.config.bos_token_id, self.config.unk_token_id]] = float("-inf")
+
+        loss = None
+        if labels is not None:
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            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: return (loss, logits) if loss is not None else (logits,)
+        return CausalLMOutputWithPast(loss=loss, logits=logits)
+
+from transformers import AutoConfig, AutoModelForCausalLM
+AutoConfig.register("chess_transformer", ChessConfig)
+AutoModelForCausalLM.register(ChessConfig, ChessForCausalLM)

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19ac6a27df140ea95ed952ddce10f4a085b904e5e3d563cea00febeb8ef22fc2
+size 4007995

special_tokens_map.json

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

tokenizer.py

@@ -0,0 +1,195 @@
+from __future__ import annotations
+
+import json
+import os
+import re
+import shutil
+from typing import Dict, List, Optional
+
+from transformers import PreTrainedTokenizer
+
+
+REGEX_CASE = re.compile(r"([a-h][1-8])")
+
+REGEX_PROMO = re.compile(r"[=\(]?([qrbnQRBN])[\)]?$")
+
+class ChessTokenizer(PreTrainedTokenizer):
+    """
+    Tokenizer qui traite le jeu d'échecs case par case.
+    Vocabulaire déterministe : Spéciaux + Cases (a1..h8) + Promotions.
+    """
+    
+
+    vocab_files_names = {"vocab_file": "vocab.json"}
+    model_input_names = ["input_ids", "attention_mask"]
+
+    # 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,
+    ):
+        self._pad_token = self.PAD_TOKEN
+        self._bos_token = self.BOS_TOKEN
+        self._eos_token = self.EOS_TOKEN
+        self._unk_token = self.UNK_TOKEN
+
+        for cle in ["pad_token", "bos_token", "eos_token", "unk_token"]:
+            kwargs.pop(cle, None)
+
+        if vocab:
+            self.map_token_id = vocab
+        elif vocab_file and os.path.exists(vocab_file):
+            with open(vocab_file, "r", encoding="utf-8") as f:
+                self.map_token_id = json.load(f)
+        else:
+            self.map_token_id = self._generer_vocabulaire()
+
+
+        self.map_id_token = {i: t for t, i in self.map_token_id.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 _generer_vocabulaire(self) -> Dict[str, int]:
+        """Génère la liste fixe des tokens nécessaires."""
+        liste_tokens = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]
+        
+        colonnes = "abcdefgh"
+        lignes = "12345678"
+        cases = [f"{c}{l}" for c in colonnes for l in lignes]
+        liste_tokens.extend(cases)
+        
+        pieces_promo = ["q", "r", "b", "n"]
+        liste_tokens.extend(pieces_promo)
+        
+
+        return {t: i for i, t in enumerate(liste_tokens)}
+
+    @property
+    def vocab_size(self) -> int:
+        return len(self.map_token_id)
+
+    def get_vocab(self) -> Dict[str, int]:
+        return dict(self.map_token_id)
+
+    def _tokenize(self, text: str) -> List[str]:
+        """
+        Transforme une phrase de coups en liste de tokens.
+        """
+        resultat = []
+        
+
+        mouvements = text.strip().split()
+        
+        for mv in mouvements:
+
+            cases_trouvees = REGEX_CASE.findall(mv)
+            
+
+            if len(cases_trouvees) >= 2:
+
+                resultat.extend(cases_trouvees[:2])
+                
+
+                match_promo = REGEX_PROMO.search(mv)
+                if match_promo:
+
+                    resultat.append(match_promo.group(1).lower())
+            
+
+            elif mv in self.map_token_id:
+                resultat.append(mv)
+            else:
+
+                resultat.append(self.UNK_TOKEN)
+                
+        return resultat
+
+    def _convert_token_to_id(self, token: str) -> int:
+        return self.map_token_id.get(token, self.map_token_id[self.UNK_TOKEN])
+
+    def _convert_id_to_token(self, index: int) -> str:
+        return self.map_id_token.get(index, self.UNK_TOKEN)
+
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        """
+        Reconstruit la chaine de caractères depuis les tokens.
+        Logique : on assemble les paires de cases.
+        """
+        sortie = []
+        tampon_cases = []
+        
+        exclus = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}
+        promotions = {"q", "r", "b", "n"}
+        
+        for t in tokens:
+            if t in exclus:
+                continue
+                
+            if t in promotions:
+                if sortie:
+                    sortie[-1] += t
+            else:
+                tampon_cases.append(t)
+
+                if len(tampon_cases) == 2:
+                    coup_complet = "".join(tampon_cases)
+                    sortie.append(coup_complet)
+                    tampon_cases = []
+        
+        return " ".join(sortie)
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> tuple:
+        """Sauvegarde le vocabulaire sur le disque."""
+        if not os.path.exists(save_directory):
+            os.makedirs(save_directory)
+            
+        nom_fichier = "vocab.json"
+        if filename_prefix:
+            nom_fichier = f"{filename_prefix}-{nom_fichier}"
+            
+        chemin_complet = os.path.join(save_directory, nom_fichier)
+        
+        with open(chemin_complet, "w", encoding="utf-8") as f:
+            json.dump(self.map_token_id, f, ensure_ascii=False, indent=2)
+            
+        return (chemin_complet,)
+
+
+    def save_pretrained(self, save_directory: str, **kwargs):
+        """
+        Sauvegarde standard + Copie du script tokenizer.py pour Hugging Face.
+        """
+        super().save_pretrained(save_directory, **kwargs)
+        
+        source = os.path.abspath(__file__)
+        dest = os.path.join(save_directory, "tokenizer.py")
+        if source != dest:
+            shutil.copy(source, dest)
+            
+        chem_config = os.path.join(save_directory, "tokenizer_config.json")
+        if os.path.exists(chem_config):
+            with open(chem_config, "r") as f:
+                cfg = json.load(f)
+            cfg["auto_map"] = {"AutoTokenizer": "tokenizer.ChessTokenizer"}
+            with open(chem_config, "w") as f:
+                json.dump(cfg, f, indent=2)
+
+
+from transformers import AutoTokenizer
+try:
+    ChessTokenizer.register_for_auto_class("AutoTokenizer")
+except:
+    pass

tokenizer_config.json

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