Chess Challenge submission by Bnz94

README.md

@@ -0,0 +1,31 @@
+---
+library_name: transformers
+tags:
+- chess
+- llm-course
+- chess-challenge
+license: mit
+---
+
+# chess-model-bnz
+
+Chess model submitted to the LLM Course Chess Challenge.
+
+## Submission Info
+
+- **Submitted by**: [Bnz94](https://huggingface.co/Bnz94)
+- **Parameters**: 996,096
+- **Organization**: LLM-course
+
+## Usage
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model = AutoModelForCausalLM.from_pretrained("LLM-course/chess-model-bnz", trust_remote_code=True)
+tokenizer = AutoTokenizer.from_pretrained("LLM-course/chess-model-bnz", trust_remote_code=True)
+```
+
+## Evaluation
+
+This model is evaluated at the [Chess Challenge Arena](https://huggingface.co/spaces/LLM-course/Chess1MChallenge).

config.json

@@ -0,0 +1,28 @@
+{
+  "architectures": [
+    "ChessForCausalLM"
+  ],
+  "bos_token_id": 1,
+  "dropout": 0.1,
+  "dtype": "float32",
+  "eos_token_id": 2,
+  "layer_norm_epsilon": 1e-06,
+  "model_type": "chess_transformer",
+  "n_ctx": 256,
+  "n_embd": 128,
+  "n_head": 4,
+  "n_inner": 256,
+  "n_kv_head": 4,
+  "n_layer": 6,
+  "pad_token_id": 0,
+  "rms_norm_epsilon": 1e-06,
+  "rope_theta": 10000.0,
+  "tie_weights": true,
+  "transformers_version": "4.57.6",
+  "use_rope": true,
+  "vocab_size": 89,
+  "auto_map": {
+    "AutoConfig": "model.ChessConfig",
+    "AutoModelForCausalLM": "model.ChessForCausalLM"
+  }
+}

model.py

@@ -0,0 +1,484 @@
+"""
+Chess Transformer Model for the Chess Challenge.
+This module provides a modern transformer architecture with:
+- RoPE (Rotary Position Embeddings)
+- SwiGLU activation
+- RMSNorm
+
+Designed to fit within the 1M parameter constraint.
+"""
+
+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):
+    """
+    Configuration class for the Chess Transformer model.
+    
+    Uses modern architecture choices:
+    - RoPE: No learned position embeddings (saves n_ctx * n_embd params)
+    - SwiGLU: 3 matrices instead of 2, but more expressive
+    - RMSNorm: Simpler and faster than LayerNorm
+    """
+    
+    model_type = "chess_transformer"
+    
+    def __init__(
+        self,
+        vocab_size: int = 1200,
+        n_embd: int = 128,
+        n_layer: int = 6,
+        n_head: int = 4,
+        n_kv_head: Optional[int] = None,  # For GQA, None = MHA
+        n_ctx: int = 256,
+        n_inner: Optional[int] = None,
+        dropout: float = 0.1,
+        rms_norm_epsilon: float = 1e-6,
+        tie_weights: bool = True,
+        use_rope: bool = True,
+        rope_theta: float = 10000.0,
+        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_kv_head = n_kv_head if n_kv_head is not None else n_head
+        self.n_ctx = n_ctx
+        # SwiGLU typically uses 2/3 * 4 * n_embd, rounded to multiple of 64
+        self.n_inner = n_inner if n_inner is not None else self._compute_swiglu_dim(n_embd)
+        self.dropout = dropout
+        self.rms_norm_epsilon = rms_norm_epsilon
+        self.tie_weights = tie_weights
+        self.tie_word_embeddings = bool(tie_weights)
+        self.use_rope = use_rope
+        self.rope_theta = rope_theta
+        # For compatibility with src/utils.py parameter estimation
+        self.layer_norm_epsilon = rms_norm_epsilon
+    
+    @staticmethod
+    def _compute_swiglu_dim(n_embd: int) -> int:
+        """Compute SwiGLU hidden dimension (typically 8/3 * n_embd, rounded)."""
+        # Standard SwiGLU uses ~2.67x multiplier
+        hidden = int(8 * n_embd / 3)
+        # Round to multiple of 64 for efficiency (optional)
+        return ((hidden + 63) // 64) * 64
+
+
+class RMSNorm(nn.Module):
+    """
+    Root Mean Square Layer Normalization.
+    
+    Simpler and faster than LayerNorm - no mean centering, no bias.
+    """
+    
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(dim))
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        # RMSNorm: x * weight / sqrt(mean(x^2) + eps)
+        norm = torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+        return x * norm * self.weight
+
+
+class RotaryEmbedding(nn.Module):
+    """
+    Rotary Position Embeddings (RoPE).
+    
+    Encodes position information directly into attention computation
+    without learnable parameters.
+    """
+    
+    def __init__(self, dim: int, max_seq_len: int = 512, theta: float = 10000.0):
+        super().__init__()
+        self.dim = dim
+        self.max_seq_len = max_seq_len
+        self.theta = theta
+        
+        # Precompute frequencies
+        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        
+        # Precompute cos/sin cache
+        self._build_cache(max_seq_len)
+    
+    def _build_cache(self, seq_len: int):
+        """Build cos/sin cache for positions."""
+        t = torch.arange(seq_len, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
+        freqs = torch.outer(t, self.inv_freq)
+        # Concatenate to get full dim
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos(), persistent=False)
+        self.register_buffer("sin_cached", emb.sin(), persistent=False)
+    
+    def forward(self, seq_len: int, device: torch.device) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Return cos and sin for the given sequence length."""
+        if seq_len > self.max_seq_len:
+            self._build_cache(seq_len)
+            self.max_seq_len = seq_len
+        
+        return (
+            self.cos_cached[:seq_len].to(device),
+            self.sin_cached[:seq_len].to(device),
+        )
+
+
+def rotate_half(x: torch.Tensor) -> torch.Tensor:
+    """Rotate half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Apply rotary position embeddings to query and key tensors."""
+    # q, k: (batch, n_head, seq_len, head_dim)
+    # cos, sin: (seq_len, head_dim)
+    cos = cos.unsqueeze(0).unsqueeze(0)  # (1, 1, seq_len, head_dim)
+    sin = sin.unsqueeze(0).unsqueeze(0)
+    
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    
+    return q_embed, k_embed
+
+
+class MultiHeadAttention(nn.Module):
+    """
+    Multi-head self-attention with RoPE.
+    
+    Supports Grouped Query Attention (GQA) when n_kv_head < n_head.
+    """
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        
+        assert config.n_embd % config.n_head == 0
+        
+        self.n_head = config.n_head
+        self.n_kv_head = config.n_kv_head
+        self.n_embd = config.n_embd
+        self.head_dim = config.n_embd // config.n_head
+        self.n_rep = config.n_head // config.n_kv_head  # For GQA
+        
+        # Separate Q, K, V projections for clarity with GQA
+        self.q_proj = nn.Linear(config.n_embd, config.n_head * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(config.n_embd, config.n_kv_head * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(config.n_embd, config.n_kv_head * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(config.n_head * self.head_dim, config.n_embd, bias=False)
+        
+        self.dropout = nn.Dropout(config.dropout)
+        
+        # RoPE
+        self.rotary_emb = RotaryEmbedding(
+            self.head_dim,
+            max_seq_len=config.n_ctx,
+            theta=config.rope_theta,
+        )
+        
+        # Causal mask
+        self.register_buffer(
+            "causal_mask",
+            torch.tril(torch.ones(config.n_ctx, config.n_ctx)).view(
+                1, 1, config.n_ctx, config.n_ctx
+            ),
+            persistent=False,
+        )
+    
+    def _repeat_kv(self, x: torch.Tensor) -> torch.Tensor:
+        """Repeat KV heads for GQA."""
+        if self.n_rep == 1:
+            return x
+        batch, n_kv_head, seq_len, head_dim = x.shape
+        x = x[:, :, None, :, :].expand(batch, n_kv_head, self.n_rep, seq_len, head_dim)
+        return x.reshape(batch, n_kv_head * self.n_rep, seq_len, head_dim)
+    
+    def forward(
+        self,
+        x: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, seq_len, _ = x.size()
+        
+        # Project Q, K, V
+        q = self.q_proj(x)
+        k = self.k_proj(x)
+        v = self.v_proj(x)
+        
+        # Reshape for 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_kv_head, self.head_dim).transpose(1, 2)
+        v = v.view(batch_size, seq_len, self.n_kv_head, self.head_dim).transpose(1, 2)
+        
+        # Apply RoPE
+        cos, sin = self.rotary_emb(seq_len, x.device)
+        q, k = apply_rotary_pos_emb(q, k, cos, sin)
+        
+        # Repeat KV for GQA
+        k = self._repeat_kv(k)
+        v = self._repeat_kv(v)
+        
+        # Scaled dot-product attention
+        attn_weights = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
+        
+        # Apply causal mask
+        causal_mask = self.causal_mask[:, :, :seq_len, :seq_len]
+        attn_weights = attn_weights.masked_fill(causal_mask == 0, float("-inf"))
+        
+        # Apply padding mask
+        if attention_mask is not None:
+            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 and project output
+        attn_output = attn_output.transpose(1, 2).contiguous().view(
+            batch_size, seq_len, self.n_embd
+        )
+        attn_output = self.o_proj(attn_output)
+        
+        return attn_output
+
+
+class SwiGLU(nn.Module):
+    """
+    SwiGLU Feed-Forward Network.
+    
+    SwiGLU(x) = (xW1 * SiLU(xW_gate)) @ W2
+    
+    More expressive than standard FFN with similar parameter count.
+    """
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        
+        hidden_dim = config.n_inner
+        
+        # Gate and up projections (can be fused for efficiency)
+        self.gate_proj = nn.Linear(config.n_embd, hidden_dim, bias=False)
+        self.up_proj = nn.Linear(config.n_embd, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, config.n_embd, bias=False)
+        self.dropout = nn.Dropout(config.dropout)
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        # SwiGLU: SiLU(gate) * up, then down
+        gate = F.silu(self.gate_proj(x))
+        up = self.up_proj(x)
+        x = gate * up
+        x = self.down_proj(x)
+        x = self.dropout(x)
+        return x
+
+
+class TransformerBlock(nn.Module):
+    """
+    Transformer block with RMSNorm, RoPE attention, and SwiGLU FFN.
+    
+    Uses pre-normalization for training stability.
+    """
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        
+        self.ln_1 = RMSNorm(config.n_embd, eps=config.rms_norm_epsilon)
+        self.attn = MultiHeadAttention(config)
+        self.ln_2 = RMSNorm(config.n_embd, eps=config.rms_norm_epsilon)
+        self.mlp = SwiGLU(config)
+    
+    def forward(
+        self,
+        x: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        # Pre-norm attention with residual
+        x = x + self.attn(self.ln_1(x), attention_mask=attention_mask)
+        # Pre-norm FFN with residual
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class ChessForCausalLM(PreTrainedModel):
+    """
+    Chess Transformer for Causal Language Modeling.
+    
+    Modern architecture with RoPE, SwiGLU, and RMSNorm.
+    """
+    
+    config_class = ChessConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    _tied_weights_keys = ["lm_head.weight"]
+    keys_to_ignore_on_load_missing = ["lm_head.weight"]
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__(config)
+        
+        # Token embeddings (no position embeddings - using RoPE)
+        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
+        
+        self.drop = nn.Dropout(config.dropout)
+        
+        # Transformer blocks
+        self.h = nn.ModuleList([
+            TransformerBlock(config) for _ in range(config.n_layer)
+        ])
+        
+        # Final RMSNorm
+        self.ln_f = RMSNorm(config.n_embd, eps=config.rms_norm_epsilon)
+        
+        # Output head
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+        
+        # 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):
+        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."""
+        std = 0.02
+        if isinstance(module, nn.Linear):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=std)
+            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=std)
+        elif isinstance(module, RMSNorm):
+            torch.nn.init.ones_(module.weight)
+    
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        
+        # Get token embeddings (no position embeddings - RoPE handles position)
+        hidden_states = self.wte(input_ids)
+        hidden_states = self.drop(hidden_states)
+        
+        # Pass through transformer blocks
+        for block in self.h:
+            hidden_states = block(hidden_states, attention_mask=attention_mask)
+        
+        # Final norm and head
+        hidden_states = self.ln_f(hidden_states)
+        logits = self.lm_head(hidden_states)
+        
+        # Compute loss if labels provided
+        loss = None
+        if labels is not None:
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
+            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 token."""
+        self.eval()
+        
+        outputs = self(input_ids)
+        logits = outputs.logits[:, -1, :] / temperature
+        
+        if top_k is not None:
+            indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+            logits[indices_to_remove] = float("-inf")
+        
+        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)
+            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")
+        
+        probs = F.softmax(logits, dim=-1)
+        next_token = torch.multinomial(probs, num_samples=1)
+        
+        return next_token.item()
+
+
+# Register with Auto classes
+from transformers import AutoConfig, AutoModelForCausalLM
+
+AutoConfig.register("chess_transformer", ChessConfig)
+AutoModelForCausalLM.register(ChessConfig, ChessForCausalLM)

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72beeff5f878a0adafff7eed5ce9f82349cdc8098d9d97a2d89a4d2221b97fa7
+size 3989408

special_tokens_map.json

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

tokenizer.py

@@ -0,0 +1,131 @@
+from __future__ import annotations
+
+import json, os, re
+from typing import Dict, List, Optional
+from transformers import PreTrainedTokenizer
+
+# UCI étendu: WPe2e4, BNg8f6(x+*), promotions "=Q", roque "(o)/(O)"
+_MOVE_RE = re.compile(r"^(?P<side>[WB])(?P<piece>[PNBRQK])(?P<src>[a-h][1-8])(?P<dst>[a-h][1-8])(?P<suffix>.*)$")
+_PROMO_RE = re.compile(r"=([QRBNqrbn])")
+
+def _parse_suffix(suffix: str):
+    s = (suffix or "").strip()
+    is_capture = "x" in s
+    is_check = "+" in s
+    is_mate = "*" in s
+    castle = "O-O-O" if "(O)" in s else ("O-O" if "(o)" in s else None)
+    promo = None
+    m = _PROMO_RE.search(s)
+    if m:
+        promo = m.group(1).lower()
+    return is_capture, is_check, is_mate, castle, promo
+
+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):
+        self._pad_token = self.PAD_TOKEN
+        self._bos_token = self.BOS_TOKEN
+        self._eos_token = self.EOS_TOKEN
+        self._unk_token = self.UNK_TOKEN
+
+        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 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,
+        )
+
+    @property
+    def vocab_size(self) -> int:
+        return len(self._vocab)
+
+    def get_vocab(self) -> Dict[str, int]:
+        return dict(self._vocab)
+
+    def _convert_token_to_id(self, token: str) -> int:
+        return self._vocab.get(token, self._vocab.get(self.UNK_TOKEN, 0))
+
+    def _convert_id_to_token(self, index: int) -> str:
+        return self._ids_to_tokens.get(index, self.UNK_TOKEN)
+
+    def _create_default_vocab(self) -> Dict[str, int]:
+        tokens: List[str] = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]
+        # Side+piece tokens (12)
+        tokens += [f"[W{p}]" for p in "PNBRQK"]
+        tokens += [f"[B{p}]" for p in "PNBRQK"]
+        # 64 squares
+        tokens += [f"[{f}{r}]" for f in "abcdefgh" for r in "12345678"]
+        # Flags / castles / promotions
+        tokens += ["[x]", "[+]", "[#]", "[O-O]", "[O-O-O]"]
+        tokens += [f"[={p}]" for p in "qrbn"]
+        return {tok: i for i, tok in enumerate(tokens)}
+
+    def _tokenize(self, text: str) -> List[str]:
+        out: List[str] = []
+        for move in (text or "").strip().split():
+            # Raw UCI like e2e4 / e7e8q (no side/piece available)
+            if re.fullmatch(r"[a-h][1-8][a-h][1-8][qrbn]?", move):
+                src, dst = move[:2], move[2:4]
+                out += [f"[{src}]", f"[{dst}]"]
+                if len(move) == 5:
+                    out += [f"[={move[4]}]"]
+                continue
+
+            m = _MOVE_RE.match(move)
+            if not m:
+                out.append(self.UNK_TOKEN)
+                continue
+
+            side = m.group("side")   # "W" or "B"
+            piece = m.group("piece") # P/N/B/R/Q/K
+            src = f"[{m.group('src')}]"
+            dst = f"[{m.group('dst')}]"
+            is_cap, is_chk, is_mate, castle, promo = _parse_suffix(m.group("suffix") or "")
+
+            out += [f"[{side}{piece}]", src, dst]
+            if castle:
+                out.append(f"[{castle}]")
+            if is_cap:
+                out.append("[x]")
+            if is_mate:
+                out.append("[#]")
+            elif is_chk:
+                out.append("[+]")
+            if promo:
+                out.append(f"[={promo}]")
+        return out
+
+    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)
+
+    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 +1,50 @@
+{
+  "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
+    }
+  },
+  "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]",
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenizer.ChessTokenizer",
+      null
+    ]
+  }
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f79be9a6b9ab0e70e9325e38d9ad07047c8c571ed7101f4bfd4b65b7cc784561
+size 5841

vocab.json

@@ -0,0 +1,91 @@
+{
+  "[PAD]": 0,
+  "[BOS]": 1,
+  "[EOS]": 2,
+  "[UNK]": 3,
+  "[WP]": 4,
+  "[WN]": 5,
+  "[WB]": 6,
+  "[WR]": 7,
+  "[WQ]": 8,
+  "[WK]": 9,
+  "[BP]": 10,
+  "[BN]": 11,
+  "[BB]": 12,
+  "[BR]": 13,
+  "[BQ]": 14,
+  "[BK]": 15,
+  "[a1]": 16,
+  "[a2]": 17,
+  "[a3]": 18,
+  "[a4]": 19,
+  "[a5]": 20,
+  "[a6]": 21,
+  "[a7]": 22,
+  "[a8]": 23,
+  "[b1]": 24,
+  "[b2]": 25,
+  "[b3]": 26,
+  "[b4]": 27,
+  "[b5]": 28,
+  "[b6]": 29,
+  "[b7]": 30,
+  "[b8]": 31,
+  "[c1]": 32,
+  "[c2]": 33,
+  "[c3]": 34,
+  "[c4]": 35,
+  "[c5]": 36,
+  "[c6]": 37,
+  "[c7]": 38,
+  "[c8]": 39,
+  "[d1]": 40,
+  "[d2]": 41,
+  "[d3]": 42,
+  "[d4]": 43,
+  "[d5]": 44,
+  "[d6]": 45,
+  "[d7]": 46,
+  "[d8]": 47,
+  "[e1]": 48,
+  "[e2]": 49,
+  "[e3]": 50,
+  "[e4]": 51,
+  "[e5]": 52,
+  "[e6]": 53,
+  "[e7]": 54,
+  "[e8]": 55,
+  "[f1]": 56,
+  "[f2]": 57,
+  "[f3]": 58,
+  "[f4]": 59,
+  "[f5]": 60,
+  "[f6]": 61,
+  "[f7]": 62,
+  "[f8]": 63,
+  "[g1]": 64,
+  "[g2]": 65,
+  "[g3]": 66,
+  "[g4]": 67,
+  "[g5]": 68,
+  "[g6]": 69,
+  "[g7]": 70,
+  "[g8]": 71,
+  "[h1]": 72,
+  "[h2]": 73,
+  "[h3]": 74,
+  "[h4]": 75,
+  "[h5]": 76,
+  "[h6]": 77,
+  "[h7]": 78,
+  "[h8]": 79,
+  "[x]": 80,
+  "[+]": 81,
+  "[#]": 82,
+  "[O-O]": 83,
+  "[O-O-O]": 84,
+  "[=q]": 85,
+  "[=r]": 86,
+  "[=b]": 87,
+  "[=n]": 88
+}