Uploaded dataset from Kaggle

.virtual_documents/__notebook_source__.ipynb

@@ -0,0 +1,990 @@
+"""
+modeling_lfm2_bi.py
+-------------------
+Bidirectional LFM2 model for encoder tasks (MNTP, NER, RE).
+
+components:
+  - Lfm2BiModel:         backbone with bidirectional attention + symmetric convolutions
+
+Bug fixes over original submission:
+  1. CUDA fast path (causal_conv1d_fn) bypassed via forward() override
+  2. from_pretrained loads causal LM weights cleanly (no double conversion)
+  3. Proper 4D attention mask construction for GQA layers
+"""
+
+import types
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import AutoConfig
+from transformers.modeling_outputs import BaseModelOutput, MaskedLMOutput
+from transformers.models.lfm2.modeling_lfm2 import (
+    Lfm2ForCausalLM,
+    Lfm2Model,
+    Lfm2PreTrainedModel,
+    Lfm2ShortConv,
+)
+
+
+# ---------------------------------------------------------------------------
+# Core conversion: causal conv → bidirectional conv
+# ---------------------------------------------------------------------------
+
+def _make_conv_bidirectional(conv_module: Lfm2ShortConv) -> None:
+    """
+    Convert a single Lfm2ShortConv layer to bidirectional in-place.
+
+    How the original causal conv works (confirmed from source):
+        self.conv = nn.Conv1d(..., padding=L_cache-1)   # left-only padding
+        conv_out  = self.conv(Bx)[..., :seqlen]         # trim excess right tokens
+
+    For kernel_size=4 (LFM2-350M), original padding=3:
+        output_len = T + 2*3 - 4 + 1 = T + 3  →  [:T] = T  ✓  (causal left-only)
+
+    Bidirectional fix — replace with symmetric padding=kernel_size//2=2:
+        output_len = T + 2*2 - 4 + 1 = T + 1  →  [:T] = T  ✓  (center-padded)
+
+    Critical: override forward() to always call slow_forward().
+    Without this, CUDA path calls causal_conv1d_fn and ignores self.conv entirely.
+    """
+    kernel_size = conv_module.L_cache  # == config.conv_L_cache (4 for LFM2-350M)
+    old_conv    = conv_module.conv
+
+    # Replace conv with symmetric padding version, copy pretrained weights
+    new_conv = nn.Conv1d(
+        in_channels=old_conv.in_channels,
+        out_channels=old_conv.out_channels,
+        kernel_size=kernel_size,
+        groups=old_conv.groups,
+        bias=old_conv.bias is not None,
+        padding=kernel_size // 2,        # symmetric: equal left and right context
+    )
+    new_conv.weight.data.copy_(old_conv.weight.data)
+    if old_conv.bias is not None:
+        new_conv.bias.data.copy_(old_conv.bias.data)
+
+    conv_module.conv = new_conv
+
+    # Override forward to bypass causal CUDA kernel (causal_conv1d_fn).
+    # slow_forward calls self.conv which is now the symmetric version.
+    def _bidirectional_forward(
+        self,
+        hidden_states: torch.Tensor,
+        past_key_values=None,
+        cache_position=None,
+        attention_mask=None,
+        **kwargs,
+    ) -> torch.Tensor:
+        # Force slow_forward — never causal_conv1d_fn
+        return self.slow_forward(
+            hidden_states,
+            past_key_values=None,       # no caching during encoding
+            cache_position=cache_position,
+            attention_mask=attention_mask,
+        )
+
+    conv_module.forward = types.MethodType(_bidirectional_forward, conv_module)
+
+
+# ---------------------------------------------------------------------------
+# Bidirectional backbone
+# ---------------------------------------------------------------------------
+
+class Lfm2BiModel(Lfm2Model):
+    """
+    LFM2 backbone converted to bidirectional operation.
+
+    Changes vs Lfm2Model:
+      - GQA layers: full 4D attention mask (no causal triangular mask)
+      - Conv layers: symmetric padding + slow_forward forced (via _make_conv_bidirectional)
+
+    State dict keys are identical to Lfm2Model for clean weight loading.
+    """
+
+    def __init__(self, config):
+        super().__init__(config)
+        # Do NOT call _make_bidirectional here — weights are random at init.
+        # It is called in from_pretrained after weights are loaded.
+
+    def _make_bidirectional(self) -> None:
+        """Convert all conv and attention layers in-place."""
+        converted_conv = 0
+        for layer in self.layers:
+            # 1. Convert convolutions
+            if not layer.is_attention_layer and hasattr(layer, "conv"):
+                _make_conv_bidirectional(layer.conv)
+                converted_conv += 1
+                
+            # 2. Safety check: Ensure attention blocks don't force causal behavior
+            elif layer.is_attention_layer and hasattr(layer, "attn"):
+                # If your inspection shows an internal 'is_causal' attribute, kill it here:
+                if hasattr(layer.attn, "is_causal"):
+                    layer.attn.is_causal = False
+
+        print(f"[Lfm2BiModel] Converted {converted_conv} conv layers. Attention maps unconstrained.")
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        """
+        Load pretrained Lfm2Model weights then apply bidirectional conversion.
+
+        Clean flow:
+          1. Load into standard Lfm2Model (causal) to get pretrained weights
+          2. Transfer weights to Lfm2BiModel instance
+          3. Apply _make_bidirectional() exactly once
+        """
+        config = kwargs.pop("config", None)
+        if config is None:
+            config = AutoConfig.from_pretrained(
+                pretrained_model_name_or_path, **{
+                    k: v for k, v in kwargs.items()
+                    if k in ("trust_remote_code", "revision", "cache_dir")
+                }
+            )
+
+        # Load causal model with pretrained weights
+        base = Lfm2Model.from_pretrained(
+            pretrained_model_name_or_path, *model_args, config=config, **kwargs
+        )
+
+        # Build bidirectional instance and transfer weights
+        # Use __new__ + parent __init__ to avoid calling _make_bidirectional prematurely
+        instance = cls.__new__(cls)
+        Lfm2Model.__init__(instance, config)
+        instance.load_state_dict(base.state_dict(), strict=True)
+
+        del base  # free causal model memory
+
+        # Apply bidirectional conversion once with pretrained weights in place
+        instance._make_bidirectional()
+
+        return instance
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: torch.Tensor = None,
+        position_ids: torch.LongTensor = None,
+        inputs_embeds: torch.Tensor = None,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        **kwargs,
+    ) -> BaseModelOutput:
+        # Drop generation-only kwargs
+        kwargs.pop("past_key_values", None)
+        kwargs.pop("use_cache", None)
+        kwargs.pop("cache_position", None)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        batch_size, seq_len, _ = inputs_embeds.shape
+        device = inputs_embeds.device
+
+        # Position ids: full sequence, no offset
+        cache_position = torch.arange(seq_len, device=device)
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0).expand(batch_size, -1)
+
+        # --- Attention mask for GQA layers (4D additive) ---
+        # Shape: (B, 1, T, T)  —  0 = attend, -inf = ignore
+        if attention_mask is not None and attention_mask.dim() == 2:
+            # (B, T) binary mask → (B, 1, T, T) additive mask
+            # Key mask: which keys to ignore
+            key_mask = attention_mask[:, None, None, :].expand(
+                batch_size, 1, seq_len, seq_len
+            )  # 1 = keep, 0 = pad
+            # Query mask: ignore rows where query is padding
+            query_mask = attention_mask[:, :, None, None].expand(
+                batch_size, seq_len, 1, seq_len  # will broadcast
+            )
+            # Combined: only attend where both query and key are real tokens
+            combined = (key_mask * query_mask.transpose(1, 2)).to(inputs_embeds.dtype)
+            # Convert to additive: 0 → 0.0 (attend), 1 → min_val (ignore)... wait
+            # 1 means real token (attend), 0 means pad (ignore)
+            # Additive: real=0.0, pad=very_negative
+            bi_attn_mask = (1.0 - combined) * torch.finfo(inputs_embeds.dtype).min
+
+            # Linear mask for conv layers: 2D binary (B, T)
+            linear_mask = attention_mask
+        elif attention_mask is not None:
+            # Already in 4D or other format, use as-is
+            bi_attn_mask = attention_mask
+            linear_mask = None
+        else:
+            bi_attn_mask = None
+            linear_mask = None
+
+        # --- Forward through layers ---
+        hidden_states = inputs_embeds
+        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)
+
+        all_hidden_states = (hidden_states,) if output_hidden_states else None
+
+        for layer in self.layers[:self.config.num_hidden_layers]:
+            # GQA layers get 4D mask, conv layers get 2D binary mask
+            layer_mask = bi_attn_mask if layer.is_attention_layer else linear_mask
+
+            hidden_states = layer(
+                hidden_states,
+                attention_mask=layer_mask,
+                position_embeddings=position_embeddings,
+                position_ids=position_ids,
+                past_key_values=None,
+                cache_position=cache_position,
+            )
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+        hidden_states = self.embedding_norm(hidden_states)
+
+        return BaseModelOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+        )
+
+
+
+class Lfm2BiForCausalLM(Lfm2ForCausalLM):
+    """
+    Causal LM wrapper that patches the underlying Lfm2Model backbone 
+    with your bidirectional Lfm2BiModel for MNTP training.
+    """
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        # 1. Load the original causal model (brings in backbone + lm_head weights)
+        model = super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        
+        # 2. Dynamically change the backbone class to your bidirectional version
+        model.model.__class__ = Lfm2BiModel
+        
+        # 3. Execute your in-place conv transformations 
+        model.model._make_bidirectional()
+        
+        return model
+
+
+
+
+# Spoof the module paths so Hugging Face can find a physical __file__
+Lfm2BiModel.__module__ = Lfm2Model.__module__
+
+# Do the same for your CausalLM wrapper if you're using Phase 1
+if 'Lfm2BiForCausalLM' in globals():
+    Lfm2BiForCausalLM.__module__ = Lfm2ForCausalLM.__module__
+model = Lfm2BiModel.from_pretrained("LiquidAI/LFM2.5-350M")
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+# Load config to grab its class mapping
+config = AutoConfig.from_pretrained("LiquidAI/LFM2.5-350M") # or your specific model path
+config_class = type(config)
+
+
+AutoModel.register(config_class, Lfm2BiModel, exist_ok=True)
+AutoModelForCausalLM.register(config_class, Lfm2BiForCausalLM, exist_ok=True)
+
+
+
+
+
+from transformers import Lfm2ForCausalLM
+
+
+
+
+config = 
+
+
+
+
+
+import json
+
+mntp_config = {
+    "model_name_or_path": "LiquidAI/LFM2.5-350M",
+    "dataset_name": "wikitext",
+    "dataset_config_name": "wikitext-103-raw-v1",
+    "per_device_train_batch_size": 8,
+    "per_device_eval_batch_size": 8,
+    "gradient_accumulation_steps": 4,
+    "do_train": True,
+    "do_eval": True,
+    "max_seq_length": 512,
+    "mask_token_type": "blank",
+    "data_collator_type": "all_mask",
+    "mlm_probability": 0.15,
+    "overwrite_output_dir": True,
+    "output_dir": "output/LFM2.5-350M-MNTP",
+    "evaluation_strategy": "steps",
+    "eval_steps": 100,
+    "save_steps": 200,
+    "stop_after_n_steps": 1000,
+    "lora_r": 16,
+    "gradient_checkpointing": True,
+    "torch_dtype": "bfloat16",
+    "trust_remote_code": True
+}
+
+with open("mntp_config.json", "w") as f:
+    json.dump(mntp_config, f, indent=4)
+
+print("Successfully generated mntp_config.json in workspace.")
+
+
+script_content = """import types
+import sys
+import os
+import torch
+import torch.nn as nn
+
+# ===========================================================================
+# 0. BACKWARD COMPATIBILITY MONKEYPATCH (Fixes llm2vec TPU import error)
+# ===========================================================================
+import transformers
+if not hasattr(transformers, "is_torch_tpu_available"):
+    # Injects a dummy fallback function since newer transformers removed it from root.
+    # Safe for GPU/CPU environments.
+    transformers.is_torch_tpu_available = lambda *args, **kwargs: False
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+from transformers.modeling_outputs import BaseModelOutput
+
+# Dynamically locate the experiments directory based on your workspace layout
+if os.path.exists("./experiments"):
+    sys.path.append(os.path.abspath("./experiments"))
+elif os.path.exists("./llm2vec/experiments"):
+    sys.path.append(os.path.abspath("./llm2vec/experiments"))
+else:
+    raise FileNotFoundError("Could not find the llm2vec 'experiments' directory in your workspace.")
+
+import run_mntp 
+
+# ===========================================================================
+# 1. DYNAMIC CLASS EXTRACTION 
+# ===========================================================================
+print("[Launcher] Fetching base model blueprint to dynamically extract remote code architectures...")
+try:
+    base_dummy = AutoModelForCausalLM.from_pretrained(
+        "LiquidAI/LFM2.5-350M", 
+        trust_remote_code=True,
+        torch_dtype=torch.bfloat16
+    )
+    Lfm2ForCausalLM = base_dummy.__class__
+    Lfm2Model = base_dummy.model.__class__
+    
+    Lfm2ShortConv = None
+    for layer in base_dummy.model.layers:
+        if hasattr(layer, "conv"):
+            Lfm2ShortConv = layer.conv.__class__
+            break
+            
+    if Lfm2ShortConv is None:
+        raise AttributeError("Could not dynamically trace Lfm2ShortConv from architecture layers.")
+        
+    print(f"[Launcher] Safely extracted classes: {Lfm2ForCausalLM.__name__}, {Lfm2Model.__name__}")
+    
+    del base_dummy
+    torch.cuda.empty_cache()
+except Exception as e:
+    print(f"[Launcher] Fallback failed during runtime reflection setup: {e}")
+    raise e
+
+# ===========================================================================
+# 2. CORE BIDIRECTIONAL CONVERSION ENGINE
+# ===========================================================================
+
+def _make_conv_bidirectional(conv_module: Lfm2ShortConv) -> None:
+    kernel_size = conv_module.L_cache  
+    old_conv = conv_module.conv
+
+    new_conv = nn.Conv1d(
+        in_channels=old_conv.in_channels,
+        out_channels=old_conv.out_channels,
+        kernel_size=kernel_size,
+        groups=old_conv.groups,
+        bias=old_conv.bias is not None,
+        padding=kernel_size // 2,        
+    )
+    new_conv.weight.data.copy_(old_conv.weight.data)
+    if old_conv.bias is not None:
+        new_conv.bias.data.copy_(old_conv.bias.data)
+
+    conv_module.conv = new_conv
+
+    def _bidirectional_forward(
+        self,
+        hidden_states: torch.Tensor,
+        past_key_values=None,
+        cache_position=None,
+        attention_mask=None,
+        **kwargs,
+    ) -> torch.Tensor:
+        return self.slow_forward(
+            hidden_states,
+            past_key_values=None,       
+            cache_position=cache_position,
+            attention_mask=attention_mask,
+        )
+
+    conv_module.forward = types.MethodType(_bidirectional_forward, conv_module)
+
+
+# ===========================================================================
+# 3. BIDIRECTIONAL MODEL BACKBONE DEFINITIONS
+# ===========================================================================
+
+class Lfm2BiModel(Lfm2Model):
+    def __init__(self, config):
+        super().__init__(config)
+
+    def _make_bidirectional(self) -> None:
+        converted_conv = 0
+        for layer in self.layers:
+            if not layer.is_attention_layer and hasattr(layer, "conv"):
+                _make_conv_bidirectional(layer.conv)
+                converted_conv += 1
+            elif layer.is_attention_layer and hasattr(layer, "attn"):
+                if hasattr(layer.attn, "is_causal"):
+                    layer.attn.is_causal = False
+        print(f"[Lfm2BiModel] Applied modifications across {converted_conv} conv layers.")
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        config = kwargs.pop("config", None)
+        if config is None:
+            config = AutoConfig.from_pretrained(
+                pretrained_model_name_or_path, **{
+                    k: v for k, v in kwargs.items()
+                    if k in ("trust_remote_code", "revision", "cache_dir")
+                }
+            )
+
+        base = Lfm2Model.from_pretrained(
+            pretrained_model_name_or_path, *model_args, config=config, **kwargs
+        )
+
+        instance = cls.__new__(cls)
+        Lfm2Model.__init__(instance, config)
+        instance.load_state_dict(base.state_dict(), strict=True)
+        del base  
+
+        instance._make_bidirectional()
+        return instance
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: torch.Tensor = None,
+        position_ids: torch.LongTensor = None,
+        inputs_embeds: torch.Tensor = None,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        **kwargs,
+    ) -> BaseModelOutput:
+        kwargs.pop("past_key_values", None)
+        kwargs.pop("use_cache", None)
+        kwargs.pop("cache_position", None)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        batch_size, seq_len, _ = inputs_embeds.shape
+        device = inputs_embeds.device
+
+        cache_position = torch.arange(seq_len, device=device)
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0).expand(batch_size, -1)
+
+        if attention_mask is not None and attention_mask.dim() == 2:
+            key_mask = attention_mask[:, None, None, :].expand(batch_size, 1, seq_len, seq_len)
+            query_mask = attention_mask[:, :, None, None].expand(batch_size, seq_len, 1, seq_len)
+            combined = (key_mask * query_mask.transpose(1, 2)).to(inputs_embeds.dtype)
+            bi_attn_mask = (1.0 - combined) * torch.finfo(inputs_embeds.dtype).min
+            linear_mask = attention_mask
+        else:
+            bi_attn_mask = attention_mask
+            linear_mask = None
+
+        hidden_states = inputs_embeds
+        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)
+        all_hidden_states = (hidden_states,) if output_hidden_states else None
+
+        for layer in self.layers[:self.config.num_hidden_layers]:
+            layer_mask = bi_attn_mask if layer.is_attention_layer else linear_mask
+            hidden_states = layer(
+                hidden_states,
+                attention_mask=layer_mask,
+                position_embeddings=position_embeddings,
+                position_ids=position_ids,
+                past_key_values=None,
+                cache_position=cache_position,
+            )
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+        hidden_states = self.embedding_norm(hidden_states)
+        return BaseModelOutput(last_hidden_state=hidden_states, hidden_states=all_hidden_states)
+
+
+class Lfm2BiForCausalLM(Lfm2ForCausalLM):
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        model = super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        model.model.__class__ = Lfm2BiModel
+        model.model._make_bidirectional()
+        return model
+
+
+# ===========================================================================
+# 4. ENVIRONMENT MODULE PATH SPOOFING & FACTORY INJECTION
+# ===========================================================================
+
+Lfm2BiModel.__module__ = Lfm2Model.__module__
+Lfm2BiForCausalLM.__module__ = Lfm2ForCausalLM.__module__
+
+config = AutoConfig.from_pretrained("LiquidAI/LFM2.5-350M", trust_remote_code=True)
+config_class = type(config)
+
+AutoModel.register(config_class, Lfm2BiModel, exist_ok=True)
+AutoModelForCausalLM.register(config_class, Lfm2BiForCausalLM, exist_ok=True)
+print("[Launcher] Architectures forced into Hugging Face registry map successfully.")
+
+
+# ===========================================================================
+# 5. RUNTIME SYSTEM ARGUMENT HIJACK & EXECUTION
+# ===========================================================================
+if __name__ == "__main__":
+    config_file_name = "mntp_config.json"
+    sys.argv = [sys.argv[0], config_file_name]
+    
+    print(f"[Launcher] Relaying pipeline configurations over to LLM2Vec engine using: {config_file_name}")
+    run_mntp.main()
+"""
+
+with open("run_lfm2_mntp.py", "w") as f:
+    f.write(script_content)
+
+print("Successfully patched run_lfm2_mntp.py for backward-compatibility wrapper support.")
+
+
+get_ipython().getoutput("python run_lfm2_mntp.py")
+
+
+import os
+
+file_path = "/kaggle/working/llm2vec/experiments/run_mntp.py"
+
+if os.path.exists(file_path):
+    with open(file_path, "r") as f:
+        content = f.read()
+
+    # 1. Strip the telemetry import statement entirely
+    content = content.replace(
+        "from transformers.utils import send_example_telemetry", 
+        "# Telemetry stripped for modern transformers compatibility"
+    )
+
+    # 2. Clean up any remnants of the old TPU utility if still present
+    content = content.replace("is_torch_tpu_available,", "")
+    content = content.replace("is_torch_tpu_available", "")
+
+    # 3. Inject safe local dummy stubs right at the top of the file
+    fallback_stubs = (
+        "\n# Compatibility fallbacks for deprecated HF utilities\n"
+        "is_torch_tpu_available = lambda *args, **kwargs: False\n"
+        "send_example_telemetry = lambda *args, **kwargs: None\n\n"
+    )
+    patched_content = fallback_stubs + content
+
+    with open(file_path, "w") as f:
+        f.write(patched_content)
+
+    print("[Success] Cleaned up telemetry and TPU dependencies in run_mntp.py.")
+else:
+    print(f"[Error] Could not find script at {file_path}.")
+
+
+import urllib.request
+import os
+
+file_path = "/kaggle/working/llm2vec/experiments/run_mntp.py"
+
+print("[1/3] Restoring a clean, pristine copy of run_mntp.py...")
+try:
+    # Try fetching a fresh copy directly from the repository source
+    url = "https://raw.githubusercontent.com/McGill-NLP/llm2vec/main/experiments/run_mntp.py"
+    urllib.request.urlretrieve(url, file_path)
+    print(" -> Successfully restored original file via source download.")
+except Exception:
+    # Fallback to local git repository rollback if offline
+    os.system(f"git checkout -- {file_path}")
+    print(" -> Successfully restored original file via git checkout.")
+
+print("[2/3] Applying precise line-by-line patches...")
+with open(file_path, "r") as f:
+    lines = f.readlines()
+
+patched_lines = []
+for line in lines:
+    # Safely strip out the deprecated TPU check item from imports
+    if "is_torch_tpu_available" in line:
+        line = line.replace("is_torch_tpu_available,", "").replace("is_torch_tpu_available", "")
+    
+    # Safely neutralize the telemetry logging import line
+    if "send_example_telemetry" in line:
+        line = "    # Telemetry logging removed for modern transformers compatibility\n"
+        
+    patched_lines.append(line)
+
+# Prepend explicit safe dummy definitions right at the absolute top of the file
+fallback_headers = [
+    "is_torch_tpu_available = lambda *args, **kwargs: False\n",
+    "send_example_telemetry = lambda *args, **kwargs: None\n\n"
+]
+
+with open(file_path, "w") as f:
+    f.writelines(fallback_headers + patched_lines)
+
+print("[3/3] Done! The file is clean, uncorrupted, and syntactically sound.")
+
+
+# Clone the repository locally to access the experiment files
+get_ipython().getoutput("git clone https://github.com/McGill-NLP/llm2vec.git")
+get_ipython().getoutput("pip install -e ./llm2vec")
+get_ipython().getoutput("pip install flash-attn --no-build-isolation")
+
+
+
+
+
+# ── Cell 1: Numbers ───────────────────────────────────────────────────────
+
+# Training config (must match mntp_config.json)
+STEPS            = 2000
+BATCH_PER_DEVICE = 16
+GRAD_ACCUM       = 4
+SEQ_LEN          = 512
+BUFFER_FACTOR    = 10     # 10x safety buffer over minimum
+
+# Split
+TRAIN_RATIO      = 0.98
+VAL_RATIO        = 0.02
+
+# Sampling
+DOCS_PER_DOMAIN  = 100_000   # equal for every domain in AraMix
+NUM_DOMAINS      = 26         # will verify after scan
+ARAMIX_TOTAL     = DOCS_PER_DOMAIN * NUM_DOMAINS   # 2,600,000
+
+# Wikipedia: solve wiki / (wiki + aramix) = 0.20
+WIKI_TOTAL       = int(ARAMIX_TOTAL * 0.20 / 0.80)  # 650,000
+
+GRAND_TOTAL      = ARAMIX_TOTAL + WIKI_TOTAL          # 3,250,000
+TRAIN_SIZE       = int(GRAND_TOTAL * TRAIN_RATIO)     # 3,185,000
+VAL_SIZE         = GRAND_TOTAL - TRAIN_SIZE           #    65,000
+
+# Misc
+ARAMIX_CONFIG    = 'minhash_deduped'   # or 'sentence_deduped'
+MIN_WORDS        = 30                  # skip very short documents
+SEED             = 42
+
+print('=' * 50)
+print(f'AraMix docs   (80%) : {ARAMIX_TOTAL:>10,}  ({DOCS_PER_DOMAIN:,} × {NUM_DOMAINS} domains)')
+print(f'Wikipedia docs(20%) : {WIKI_TOTAL:>10,}')
+print(f'Grand total         : {GRAND_TOTAL:>10,}')
+print(f'Train (98%)         : {TRAIN_SIZE:>10,}')
+print(f'Val   ( 2%)         : {VAL_SIZE:>10,}')
+print('=' * 50)
+
+
+# ── Cell 2: Install & imports ─────────────────────────────────────────────
+
+# !pip uninstall -y pyarrow datasets
+# !pip install --no-cache datasets
+import os
+import json
+import random
+from pathlib import Path
+from collections import defaultdict, Counter
+
+import pandas as pd
+# import pyarrow as pa
+# import pyarrow.parquet as pq
+from datasets import load_dataset
+
+SAVE_DIR = Path('/kaggle/working')
+random.seed(SEED)
+
+print('Ready.')
+
+
+# ── Cell 3: Discover all AraMix domains (quick scan) ─────────────────────
+# Scan first 500K docs to find every unique domain name.
+# Takes ~5 min. Avoids hardcoding domain names.
+
+SCAN_LIMIT = 500_000
+print(f'Scanning {SCAN_LIMIT:,} docs to discover domains...')
+
+ds_scan = load_dataset(
+    'AdaMLLab/AraMix-domain-classified',
+    ARAMIX_CONFIG,
+    split='train',
+    streaming=True,
+    trust_remote_code=True,
+)
+
+domain_counter = Counter()
+for i, doc in enumerate(ds_scan):
+    domain_counter[doc.get('domain', 'unknown')] += 1
+    if i + 1 >= SCAN_LIMIT:
+        break
+
+ALL_DOMAINS = sorted(domain_counter.keys())
+NUM_DOMAINS = len(ALL_DOMAINS)
+
+# Recalculate totals if domain count differs from estimate
+ARAMIX_TOTAL = DOCS_PER_DOMAIN * NUM_DOMAINS
+WIKI_TOTAL   = int(ARAMIX_TOTAL * 0.20 / 0.80)
+GRAND_TOTAL  = ARAMIX_TOTAL + WIKI_TOTAL
+TRAIN_SIZE   = int(GRAND_TOTAL * TRAIN_RATIO)
+VAL_SIZE     = GRAND_TOTAL - TRAIN_SIZE
+
+print(f'\nFound {NUM_DOMAINS} domains:\n')
+print(f'{"Domain":<40} {"In scan":>8}  {"% of scan":>9}')
+print('-' * 62)
+for domain, count in domain_counter.most_common():
+    print(f'{domain:<40} {count:>8,}  {count/SCAN_LIMIT*100:>8.1f}%')
+
+print(f'\nUpdated totals:')
+print(f'  AraMix  : {ARAMIX_TOTAL:,}')
+print(f'  Wikipedia: {WIKI_TOTAL:,}')
+print(f'  Train   : {TRAIN_SIZE:,}')
+print(f'  Val     : {VAL_SIZE:,}')
+
+
+# ── Cell 4: Stream & sample AraMix equally by domain ─────────────────────
+# Each domain gets exactly DOCS_PER_DOMAIN documents.
+# Streaming + shuffle buffer gives a random sample — not just the first N.
+
+bucket_counts = Counter({d: 0 for d in ALL_DOMAINS})
+aramix_rows   = []
+
+skipped_short = 0
+skipped_full  = 0
+collected     = 0
+
+print(f'Sampling {DOCS_PER_DOMAIN:,} docs × {NUM_DOMAINS} domains = {ARAMIX_TOTAL:,} total...')
+
+ds = load_dataset(
+    'AdaMLLab/AraMix-domain-classified',
+    ARAMIX_CONFIG,
+    split='train',
+    streaming=True,
+    trust_remote_code=True,
+).shuffle(seed=SEED, buffer_size=100_000)
+
+for doc in ds:
+    domain = doc.get('domain', 'unknown')
+    text   = doc.get('text', '')
+
+    # Unknown domain — skip
+    if domain not in bucket_counts:
+        continue
+
+    # Bucket full — skip
+    if bucket_counts[domain] >= DOCS_PER_DOMAIN:
+        skipped_full += 1
+        if collected >= ARAMIX_TOTAL:
+            break
+        continue
+
+    # Too short — skip
+    if len(text.split()) < MIN_WORDS:
+        skipped_short += 1
+        continue
+
+    aramix_rows.append({
+        'text':   text,
+        'domain': domain,
+        'source': doc.get('source', 'unknown'),
+    })
+    bucket_counts[domain] += 1
+    collected += 1
+
+    if collected % 250_000 == 0:
+        filled   = sum(1 for d in ALL_DOMAINS if bucket_counts[d] >= DOCS_PER_DOMAIN)
+        pct      = collected / ARAMIX_TOTAL * 100
+        print(f'  [{pct:5.1f}%] {collected:>8,} collected | {filled}/{NUM_DOMAINS} domains full | {skipped_short:,} short skipped')
+
+    if collected >= ARAMIX_TOTAL:
+        break
+
+print(f'\nAraMix done: {collected:,} documents collected.')
+print(f'Skipped (too short) : {skipped_short:,}')
+print(f'Skipped (bucket full): {skipped_full:,}')
+
+
+# ── Cell 5: Verify AraMix bucket counts ───────────────────────────────────
+
+print(f'{"Domain":<40} {"Collected":>10}  {"Status":>8}')
+print('-' * 62)
+
+incomplete = []
+for domain in ALL_DOMAINS:
+    count  = bucket_counts[domain]
+    status = '✓' if count >= DOCS_PER_DOMAIN else f'⚠ {count:,}'
+    print(f'{domain:<40} {count:>10,}  {status:>8}')
+    if count < DOCS_PER_DOMAIN:
+        incomplete.append((domain, count))
+
+print()
+if incomplete:
+    print(f'⚠  {len(incomplete)} domain(s) below target (rare domains — expected):')
+    for d, c in incomplete:
+        print(f'   {d}: {c:,} / {DOCS_PER_DOMAIN:,}')
+else:
+    print(f'✓  All {NUM_DOMAINS} domains have exactly {DOCS_PER_DOMAIN:,} documents.')
+
+
+# ── Cell 6: Sample Arabic Wikipedia (650K articles) ───────────────────────
+# Wikipedia Arabic 20231101.ar has ~1.2M articles.
+# We sample WIKI_TOTAL (~650K = 53% of it).
+# Wikipedia is entity-dense and relation-rich — perfect RE signal for MNTP.
+
+print(f'Loading Arabic Wikipedia (wikimedia/wikipedia, 20231101.ar)...')
+print(f'Target: {WIKI_TOTAL:,} articles')
+
+wiki_ds = load_dataset(
+    'wikimedia/wikipedia',
+    '20231101.ar',
+    split='train',
+    streaming=True,
+    trust_remote_code=True,
+).shuffle(seed=SEED, buffer_size=50_000)
+
+wiki_rows      = []
+wiki_skipped   = 0
+
+for doc in wiki_ds:
+    text = doc.get('text', '').strip()
+
+    # Wikipedia articles can have very short stubs — skip them
+    if len(text.split()) < MIN_WORDS:
+        wiki_skipped += 1
+        continue
+
+    wiki_rows.append({
+        'text':   text,
+        'domain': 'Wikipedia',
+        'source': 'wikimedia/wikipedia',
+    })
+
+    if len(wiki_rows) % 100_000 == 0:
+        pct = len(wiki_rows) / WIKI_TOTAL * 100
+        print(f'  [{pct:5.1f}%] {len(wiki_rows):>8,} collected | {wiki_skipped:,} stubs skipped')
+
+    if len(wiki_rows) >= WIKI_TOTAL:
+        break
+
+print(f'\nWikipedia done: {len(wiki_rows):,} articles collected.')
+print(f'Skipped (stubs < {MIN_WORDS} words): {wiki_skipped:,}')
+
+
+# ── Cell 7: Combine → shuffle → train/val split ───────────────────────────
+
+print('Combining AraMix + Wikipedia...')
+all_rows = aramix_rows + wiki_rows
+print(f'  AraMix    : {len(aramix_rows):,}')
+print(f'  Wikipedia : {len(wiki_rows):,}')
+print(f'  Total     : {len(all_rows):,}')
+
+# Shuffle before splitting so train/val are not domain-sorted
+print('\nShuffling...')
+random.shuffle(all_rows)
+
+# Split
+actual_train_size = int(len(all_rows) * TRAIN_RATIO)
+train_rows = all_rows[:actual_train_size]
+val_rows   = all_rows[actual_train_size:]
+
+print(f'\nSplit:')
+print(f'  Train : {len(train_rows):,}  ({len(train_rows)/len(all_rows)*100:.1f}%)')
+print(f'  Val   : {len(val_rows):,}  ({len(val_rows)/len(all_rows)*100:.1f}%)')
+
+
+# ── Cell 8: Save train.parquet + val.parquet ──────────────────────────────
+
+def save_parquet(rows, path):
+    df = pd.DataFrame(rows)
+    df.to_parquet(path, index=False, engine='pyarrow', compression='snappy')
+    size_gb = Path(path).stat().st_size / 1e9
+    print(f'  Saved {len(rows):,} rows → {path}  ({size_gb:.2f} GB)')
+    return df
+
+print('Saving...')
+train_path = SAVE_DIR / 'train.parquet'
+val_path   = SAVE_DIR / 'val.parquet'
+
+train_df = save_parquet(train_rows, train_path)
+val_df   = save_parquet(val_rows,   val_path)
+
+# Save metadata alongside
+meta = {
+    'aramix_config':    ARAMIX_CONFIG,
+    'wiki_subset':      '20231101.ar',
+    'docs_per_domain':  DOCS_PER_DOMAIN,
+    'num_domains':      NUM_DOMAINS,
+    'aramix_total':     len(aramix_rows),
+    'wiki_total':       len(wiki_rows),
+    'grand_total':      len(all_rows),
+    'train_size':       len(train_rows),
+    'val_size':         len(val_rows),
+    'train_ratio':      TRAIN_RATIO,
+    'min_words':        MIN_WORDS,
+    'seed':             SEED,
+    'domains':          ALL_DOMAINS,
+}
+with open(SAVE_DIR / 'meta.json', 'w', encoding='utf-8') as f:
+    json.dump(meta, f, ensure_ascii=False, indent=2)
+
+print('\nDone. Files saved:')
+for p in [train_path, val_path, SAVE_DIR / 'meta.json']:
+    print(f'  {p}')
+
+
+# ── Cell 9: Sanity check ──────────────────────────────────────────────────
+
+print('=== TRAIN ===')
+print(f'Rows   : {len(train_df):,}')
+print(f'Columns: {list(train_df.columns)}')
+
+print('\nDomain distribution in train:')
+domain_dist = train_df['domain'].value_counts()
+for domain, count in domain_dist.items():
+    bar = '█' * int(count / domain_dist.max() * 30)
+    print(f'  {domain:<35} {count:>8,}  {bar}')
+
+print('\nSource distribution in train:')
+print(train_df['source'].value_counts().to_string())
+
+train_df['word_count'] = train_df['text'].str.split().str.len()
+print(f'\nWord count stats (train):')
+print(f'  Mean  : {train_df["word_count"].mean():.0f}')
+print(f'  Median: {train_df["word_count"].median():.0f}')
+print(f'  Min   : {train_df["word_count"].min()}')
+print(f'  Max   : {train_df["word_count"].max():,}')
+print(f'\nEstimated total tokens (×1.3 tok/word):')
+print(f'  Train : {train_df["word_count"].sum() * 1.3 / 1e6:.0f}M')
+
+val_df['word_count'] = val_df['text'].str.split().str.len()
+print(f'  Val   : {val_df["word_count"].sum() * 1.3 / 1e6:.0f}M')
+
+print('\n=== VAL ===')
+print(f'Rows   : {len(val_df):,}')
+print('Domain distribution in val:')
+print(val_df['domain'].value_counts().to_string())
+
+print('\n✓ All done. Upload /kaggle/working/train.parquet and val.parquet to your dataset.')
+
+
+

__notebook__.ipynb

@@ -0,0 +1 @@
+{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.12.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"none","dataSources":[],"dockerImageVersionId":28755,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"\"\"\"\nmodeling_lfm2_bi.py\n-------------------\nBidirectional LFM2 model for encoder tasks (MNTP, NER, RE).\n\ncomponents:\n  - Lfm2BiModel:         backbone with bidirectional attention + symmetric convolutions\n\nBug fixes over original submission:\n  1. CUDA fast path (causal_conv1d_fn) bypassed via forward() override\n  2. from_pretrained loads causal LM weights cleanly (no double conversion)\n  3. Proper 4D attention mask construction for GQA layers\n\"\"\"\n\nimport types\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom transformers import AutoConfig\nfrom transformers.modeling_outputs import BaseModelOutput, MaskedLMOutput\nfrom transformers.models.lfm2.modeling_lfm2 import (\n    Lfm2ForCausalLM,\n    Lfm2Model,\n    Lfm2PreTrainedModel,\n    Lfm2ShortConv,\n)\n\n\n# ---------------------------------------------------------------------------\n# Core conversion: causal conv → bidirectional conv\n# ---------------------------------------------------------------------------\n\ndef _make_conv_bidirectional(conv_module: Lfm2ShortConv) -> None:\n    \"\"\"\n    Convert a single Lfm2ShortConv layer to bidirectional in-place.\n\n    How the original causal conv works (confirmed from source):\n        self.conv = nn.Conv1d(..., padding=L_cache-1)   # left-only padding\n        conv_out  = self.conv(Bx)[..., :seqlen]         # trim excess right tokens\n\n    For kernel_size=4 (LFM2-350M), original padding=3:\n        output_len = T + 2*3 - 4 + 1 = T + 3  →  [:T] = T  ✓  (causal left-only)\n\n    Bidirectional fix — replace with symmetric padding=kernel_size//2=2:\n        output_len = T + 2*2 - 4 + 1 = T + 1  →  [:T] = T  ✓  (center-padded)\n\n    Critical: override forward() to always call slow_forward().\n    Without this, CUDA path calls causal_conv1d_fn and ignores self.conv entirely.\n    \"\"\"\n    kernel_size = conv_module.L_cache  # == config.conv_L_cache (4 for LFM2-350M)\n    old_conv    = conv_module.conv\n\n    # Replace conv with symmetric padding version, copy pretrained weights\n    new_conv = nn.Conv1d(\n        in_channels=old_conv.in_channels,\n        out_channels=old_conv.out_channels,\n        kernel_size=kernel_size,\n        groups=old_conv.groups,\n        bias=old_conv.bias is not None,\n        padding=kernel_size // 2,        # symmetric: equal left and right context\n    )\n    new_conv.weight.data.copy_(old_conv.weight.data)\n    if old_conv.bias is not None:\n        new_conv.bias.data.copy_(old_conv.bias.data)\n\n    conv_module.conv = new_conv\n\n    # Override forward to bypass causal CUDA kernel (causal_conv1d_fn).\n    # slow_forward calls self.conv which is now the symmetric version.\n    def _bidirectional_forward(\n        self,\n        hidden_states: torch.Tensor,\n        past_key_values=None,\n        cache_position=None,\n        attention_mask=None,\n        **kwargs,\n    ) -> torch.Tensor:\n        # Force slow_forward — never causal_conv1d_fn\n        return self.slow_forward(\n            hidden_states,\n            past_key_values=None,       # no caching during encoding\n            cache_position=cache_position,\n            attention_mask=attention_mask,\n        )\n\n    conv_module.forward = types.MethodType(_bidirectional_forward, conv_module)\n\n\n# ---------------------------------------------------------------------------\n# Bidirectional backbone\n# ---------------------------------------------------------------------------\n\nclass Lfm2BiModel(Lfm2Model):\n    \"\"\"\n    LFM2 backbone converted to bidirectional operation.\n\n    Changes vs Lfm2Model:\n      - GQA layers: full 4D attention mask (no causal triangular mask)\n      - Conv layers: symmetric padding + slow_forward forced (via _make_conv_bidirectional)\n\n    State dict keys are identical to Lfm2Model for clean weight loading.\n    \"\"\"\n\n    def __init__(self, config):\n        super().__init__(config)\n        # Do NOT call _make_bidirectional here — weights are random at init.\n        # It is called in from_pretrained after weights are loaded.\n\n    def _make_bidirectional(self) -> None:\n        \"\"\"Convert all conv and attention layers in-place.\"\"\"\n        converted_conv = 0\n        for layer in self.layers:\n            # 1. Convert convolutions\n            if not layer.is_attention_layer and hasattr(layer, \"conv\"):\n                _make_conv_bidirectional(layer.conv)\n                converted_conv += 1\n                \n            # 2. Safety check: Ensure attention blocks don't force causal behavior\n            elif layer.is_attention_layer and hasattr(layer, \"attn\"):\n                # If your inspection shows an internal 'is_causal' attribute, kill it here:\n                if hasattr(layer.attn, \"is_causal\"):\n                    layer.attn.is_causal = False\n\n        print(f\"[Lfm2BiModel] Converted {converted_conv} conv layers. Attention maps unconstrained.\")\n\n    @classmethod\n    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):\n        \"\"\"\n        Load pretrained Lfm2Model weights then apply bidirectional conversion.\n\n        Clean flow:\n          1. Load into standard Lfm2Model (causal) to get pretrained weights\n          2. Transfer weights to Lfm2BiModel instance\n          3. Apply _make_bidirectional() exactly once\n        \"\"\"\n        config = kwargs.pop(\"config\", None)\n        if config is None:\n            config = AutoConfig.from_pretrained(\n                pretrained_model_name_or_path, **{\n                    k: v for k, v in kwargs.items()\n                    if k in (\"trust_remote_code\", \"revision\", \"cache_dir\")\n                }\n            )\n\n        # Load causal model with pretrained weights\n        base = Lfm2Model.from_pretrained(\n            pretrained_model_name_or_path, *model_args, config=config, **kwargs\n        )\n\n        # Build bidirectional instance and transfer weights\n        # Use __new__ + parent __init__ to avoid calling _make_bidirectional prematurely\n        instance = cls.__new__(cls)\n        Lfm2Model.__init__(instance, config)\n        instance.load_state_dict(base.state_dict(), strict=True)\n\n        del base  # free causal model memory\n\n        # Apply bidirectional conversion once with pretrained weights in place\n        instance._make_bidirectional()\n\n        return instance\n\n    def forward(\n        self,\n        input_ids: torch.LongTensor = None,\n        attention_mask: torch.Tensor = None,\n        position_ids: torch.LongTensor = None,\n        inputs_embeds: torch.Tensor = None,\n        output_hidden_states: bool = False,\n        return_dict: bool = True,\n        **kwargs,\n    ) -> BaseModelOutput:\n        # Drop generation-only kwargs\n        kwargs.pop(\"past_key_values\", None)\n        kwargs.pop(\"use_cache\", None)\n        kwargs.pop(\"cache_position\", None)\n\n        if inputs_embeds is None:\n            inputs_embeds = self.embed_tokens(input_ids)\n\n        batch_size, seq_len, _ = inputs_embeds.shape\n        device = inputs_embeds.device\n\n        # Position ids: full sequence, no offset\n        cache_position = torch.arange(seq_len, device=device)\n        if position_ids is None:\n            position_ids = cache_position.unsqueeze(0).expand(batch_size, -1)\n\n        # --- Attention mask for GQA layers (4D additive) ---\n        # Shape: (B, 1, T, T)  —  0 = attend, -inf = ignore\n        if attention_mask is not None and attention_mask.dim() == 2:\n            # (B, T) binary mask → (B, 1, T, T) additive mask\n            # Key mask: which keys to ignore\n            key_mask = attention_mask[:, None, None, :].expand(\n                batch_size, 1, seq_len, seq_len\n            )  # 1 = keep, 0 = pad\n            # Query mask: ignore rows where query is padding\n            query_mask = attention_mask[:, :, None, None].expand(\n                batch_size, seq_len, 1, seq_len  # will broadcast\n            )\n            # Combined: only attend where both query and key are real tokens\n            combined = (key_mask * query_mask.transpose(1, 2)).to(inputs_embeds.dtype)\n            # Convert to additive: 0 → 0.0 (attend), 1 → min_val (ignore)... wait\n            # 1 means real token (attend), 0 means pad (ignore)\n            # Additive: real=0.0, pad=very_negative\n            bi_attn_mask = (1.0 - combined) * torch.finfo(inputs_embeds.dtype).min\n\n            # Linear mask for conv layers: 2D binary (B, T)\n            linear_mask = attention_mask\n        elif attention_mask is not None:\n            # Already in 4D or other format, use as-is\n            bi_attn_mask = attention_mask\n            linear_mask = None\n        else:\n            bi_attn_mask = None\n            linear_mask = None\n\n        # --- Forward through layers ---\n        hidden_states = inputs_embeds\n        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)\n\n        all_hidden_states = (hidden_states,) if output_hidden_states else None\n\n        for layer in self.layers[:self.config.num_hidden_layers]:\n            # GQA layers get 4D mask, conv layers get 2D binary mask\n            layer_mask = bi_attn_mask if layer.is_attention_layer else linear_mask\n\n            hidden_states = layer(\n                hidden_states,\n                attention_mask=layer_mask,\n                position_embeddings=position_embeddings,\n                position_ids=position_ids,\n                past_key_values=None,\n                cache_position=cache_position,\n            )\n\n            if output_hidden_states:\n                all_hidden_states = all_hidden_states + (hidden_states,)\n\n        hidden_states = self.embedding_norm(hidden_states)\n\n        return BaseModelOutput(\n            last_hidden_state=hidden_states,\n            hidden_states=all_hidden_states,\n        )\n\n\n\nclass Lfm2BiForCausalLM(Lfm2ForCausalLM):\n    \"\"\"\n    Causal LM wrapper that patches the underlying Lfm2Model backbone \n    with your bidirectional Lfm2BiModel for MNTP training.\n    \"\"\"\n    @classmethod\n    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):\n        # 1. Load the original causal model (brings in backbone + lm_head weights)\n        model = super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)\n        \n        # 2. Dynamically change the backbone class to your bidirectional version\n        model.model.__class__ = Lfm2BiModel\n        \n        # 3. Execute your in-place conv transformations \n        model.model._make_bidirectional()\n        \n        return model\n\n\n\n\n# Spoof the module paths so Hugging Face can find a physical __file__\nLfm2BiModel.__module__ = Lfm2Model.__module__\n\n# Do the same for your CausalLM wrapper if you're using Phase 1\nif 'Lfm2BiForCausalLM' in globals():\n    Lfm2BiForCausalLM.__module__ = Lfm2ForCausalLM.__module__\nmodel = Lfm2BiModel.from_pretrained(\"LiquidAI/LFM2.5-350M\")\nfrom transformers import AutoConfig, AutoModel, AutoModelForCausalLM\n\n# Load config to grab its class mapping\nconfig = AutoConfig.from_pretrained(\"LiquidAI/LFM2.5-350M\") # or your specific model path\nconfig_class = type(config)\n\n\nAutoModel.register(config_class, Lfm2BiModel, exist_ok=True)\nAutoModelForCausalLM.register(config_class, Lfm2BiForCausalLM, exist_ok=True)","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from transformers import Lfm2ForCausalLM\n\n","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"config = ","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import json\n\nmntp_config = {\n    \"model_name_or_path\": \"LiquidAI/LFM2.5-350M\",\n    \"dataset_name\": \"wikitext\",\n    \"dataset_config_name\": \"wikitext-103-raw-v1\",\n    \"per_device_train_batch_size\": 8,\n    \"per_device_eval_batch_size\": 8,\n    \"gradient_accumulation_steps\": 4,\n    \"do_train\": True,\n    \"do_eval\": True,\n    \"max_seq_length\": 512,\n    \"mask_token_type\": \"blank\",\n    \"data_collator_type\": \"all_mask\",\n    \"mlm_probability\": 0.15,\n    \"overwrite_output_dir\": True,\n    \"output_dir\": \"output/LFM2.5-350M-MNTP\",\n    \"evaluation_strategy\": \"steps\",\n    \"eval_steps\": 100,\n    \"save_steps\": 200,\n    \"stop_after_n_steps\": 1000,\n    \"lora_r\": 16,\n    \"gradient_checkpointing\": True,\n    \"torch_dtype\": \"bfloat16\",\n    \"trust_remote_code\": True\n}\n\nwith open(\"mntp_config.json\", \"w\") as f:\n    json.dump(mntp_config, f, indent=4)\n\nprint(\"Successfully generated mntp_config.json in workspace.\")","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"script_content = \"\"\"import types\nimport sys\nimport os\nimport torch\nimport torch.nn as nn\n\n# ===========================================================================\n# 0. BACKWARD COMPATIBILITY MONKEYPATCH (Fixes llm2vec TPU import error)\n# ===========================================================================\nimport transformers\nif not hasattr(transformers, \"is_torch_tpu_available\"):\n    # Injects a dummy fallback function since newer transformers removed it from root.\n    # Safe for GPU/CPU environments.\n    transformers.is_torch_tpu_available = lambda *args, **kwargs: False\n\nfrom transformers import AutoConfig, AutoModel, AutoModelForCausalLM\nfrom transformers.modeling_outputs import BaseModelOutput\n\n# Dynamically locate the experiments directory based on your workspace layout\nif os.path.exists(\"./experiments\"):\n    sys.path.append(os.path.abspath(\"./experiments\"))\nelif os.path.exists(\"./llm2vec/experiments\"):\n    sys.path.append(os.path.abspath(\"./llm2vec/experiments\"))\nelse:\n    raise FileNotFoundError(\"Could not find the llm2vec 'experiments' directory in your workspace.\")\n\nimport run_mntp \n\n# ===========================================================================\n# 1. DYNAMIC CLASS EXTRACTION \n# ===========================================================================\nprint(\"[Launcher] Fetching base model blueprint to dynamically extract remote code architectures...\")\ntry:\n    base_dummy = AutoModelForCausalLM.from_pretrained(\n        \"LiquidAI/LFM2.5-350M\", \n        trust_remote_code=True,\n        torch_dtype=torch.bfloat16\n    )\n    Lfm2ForCausalLM = base_dummy.__class__\n    Lfm2Model = base_dummy.model.__class__\n    \n    Lfm2ShortConv = None\n    for layer in base_dummy.model.layers:\n        if hasattr(layer, \"conv\"):\n            Lfm2ShortConv = layer.conv.__class__\n            break\n            \n    if Lfm2ShortConv is None:\n        raise AttributeError(\"Could not dynamically trace Lfm2ShortConv from architecture layers.\")\n        \n    print(f\"[Launcher] Safely extracted classes: {Lfm2ForCausalLM.__name__}, {Lfm2Model.__name__}\")\n    \n    del base_dummy\n    torch.cuda.empty_cache()\nexcept Exception as e:\n    print(f\"[Launcher] Fallback failed during runtime reflection setup: {e}\")\n    raise e\n\n# ===========================================================================\n# 2. CORE BIDIRECTIONAL CONVERSION ENGINE\n# ===========================================================================\n\ndef _make_conv_bidirectional(conv_module: Lfm2ShortConv) -> None:\n    kernel_size = conv_module.L_cache  \n    old_conv = conv_module.conv\n\n    new_conv = nn.Conv1d(\n        in_channels=old_conv.in_channels,\n        out_channels=old_conv.out_channels,\n        kernel_size=kernel_size,\n        groups=old_conv.groups,\n        bias=old_conv.bias is not None,\n        padding=kernel_size // 2,        \n    )\n    new_conv.weight.data.copy_(old_conv.weight.data)\n    if old_conv.bias is not None:\n        new_conv.bias.data.copy_(old_conv.bias.data)\n\n    conv_module.conv = new_conv\n\n    def _bidirectional_forward(\n        self,\n        hidden_states: torch.Tensor,\n        past_key_values=None,\n        cache_position=None,\n        attention_mask=None,\n        **kwargs,\n    ) -> torch.Tensor:\n        return self.slow_forward(\n            hidden_states,\n            past_key_values=None,       \n            cache_position=cache_position,\n            attention_mask=attention_mask,\n        )\n\n    conv_module.forward = types.MethodType(_bidirectional_forward, conv_module)\n\n\n# ===========================================================================\n# 3. BIDIRECTIONAL MODEL BACKBONE DEFINITIONS\n# ===========================================================================\n\nclass Lfm2BiModel(Lfm2Model):\n    def __init__(self, config):\n        super().__init__(config)\n\n    def _make_bidirectional(self) -> None:\n        converted_conv = 0\n        for layer in self.layers:\n            if not layer.is_attention_layer and hasattr(layer, \"conv\"):\n                _make_conv_bidirectional(layer.conv)\n                converted_conv += 1\n            elif layer.is_attention_layer and hasattr(layer, \"attn\"):\n                if hasattr(layer.attn, \"is_causal\"):\n                    layer.attn.is_causal = False\n        print(f\"[Lfm2BiModel] Applied modifications across {converted_conv} conv layers.\")\n\n    @classmethod\n    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):\n        config = kwargs.pop(\"config\", None)\n        if config is None:\n            config = AutoConfig.from_pretrained(\n                pretrained_model_name_or_path, **{\n                    k: v for k, v in kwargs.items()\n                    if k in (\"trust_remote_code\", \"revision\", \"cache_dir\")\n                }\n            )\n\n        base = Lfm2Model.from_pretrained(\n            pretrained_model_name_or_path, *model_args, config=config, **kwargs\n        )\n\n        instance = cls.__new__(cls)\n        Lfm2Model.__init__(instance, config)\n        instance.load_state_dict(base.state_dict(), strict=True)\n        del base  \n\n        instance._make_bidirectional()\n        return instance\n\n    def forward(\n        self,\n        input_ids: torch.LongTensor = None,\n        attention_mask: torch.Tensor = None,\n        position_ids: torch.LongTensor = None,\n        inputs_embeds: torch.Tensor = None,\n        output_hidden_states: bool = False,\n        return_dict: bool = True,\n        **kwargs,\n    ) -> BaseModelOutput:\n        kwargs.pop(\"past_key_values\", None)\n        kwargs.pop(\"use_cache\", None)\n        kwargs.pop(\"cache_position\", None)\n\n        if inputs_embeds is None:\n            inputs_embeds = self.embed_tokens(input_ids)\n\n        batch_size, seq_len, _ = inputs_embeds.shape\n        device = inputs_embeds.device\n\n        cache_position = torch.arange(seq_len, device=device)\n        if position_ids is None:\n            position_ids = cache_position.unsqueeze(0).expand(batch_size, -1)\n\n        if attention_mask is not None and attention_mask.dim() == 2:\n            key_mask = attention_mask[:, None, None, :].expand(batch_size, 1, seq_len, seq_len)\n            query_mask = attention_mask[:, :, None, None].expand(batch_size, seq_len, 1, seq_len)\n            combined = (key_mask * query_mask.transpose(1, 2)).to(inputs_embeds.dtype)\n            bi_attn_mask = (1.0 - combined) * torch.finfo(inputs_embeds.dtype).min\n            linear_mask = attention_mask\n        else:\n            bi_attn_mask = attention_mask\n            linear_mask = None\n\n        hidden_states = inputs_embeds\n        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)\n        all_hidden_states = (hidden_states,) if output_hidden_states else None\n\n        for layer in self.layers[:self.config.num_hidden_layers]:\n            layer_mask = bi_attn_mask if layer.is_attention_layer else linear_mask\n            hidden_states = layer(\n                hidden_states,\n                attention_mask=layer_mask,\n                position_embeddings=position_embeddings,\n                position_ids=position_ids,\n                past_key_values=None,\n                cache_position=cache_position,\n            )\n            if output_hidden_states:\n                all_hidden_states = all_hidden_states + (hidden_states,)\n\n        hidden_states = self.embedding_norm(hidden_states)\n        return BaseModelOutput(last_hidden_state=hidden_states, hidden_states=all_hidden_states)\n\n\nclass Lfm2BiForCausalLM(Lfm2ForCausalLM):\n    @classmethod\n    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):\n        model = super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)\n        model.model.__class__ = Lfm2BiModel\n        model.model._make_bidirectional()\n        return model\n\n\n# ===========================================================================\n# 4. ENVIRONMENT MODULE PATH SPOOFING & FACTORY INJECTION\n# ===========================================================================\n\nLfm2BiModel.__module__ = Lfm2Model.__module__\nLfm2BiForCausalLM.__module__ = Lfm2ForCausalLM.__module__\n\nconfig = AutoConfig.from_pretrained(\"LiquidAI/LFM2.5-350M\", trust_remote_code=True)\nconfig_class = type(config)\n\nAutoModel.register(config_class, Lfm2BiModel, exist_ok=True)\nAutoModelForCausalLM.register(config_class, Lfm2BiForCausalLM, exist_ok=True)\nprint(\"[Launcher] Architectures forced into Hugging Face registry map successfully.\")\n\n\n# ===========================================================================\n# 5. RUNTIME SYSTEM ARGUMENT HIJACK & EXECUTION\n# ===========================================================================\nif __name__ == \"__main__\":\n    config_file_name = \"mntp_config.json\"\n    sys.argv = [sys.argv[0], config_file_name]\n    \n    print(f\"[Launcher] Relaying pipeline configurations over to LLM2Vec engine using: {config_file_name}\")\n    run_mntp.main()\n\"\"\"\n\nwith open(\"run_lfm2_mntp.py\", \"w\") as f:\n    f.write(script_content)\n\nprint(\"Successfully patched run_lfm2_mntp.py for backward-compatibility wrapper support.\")","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"!python run_lfm2_mntp.py","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\n\nfile_path = \"/kaggle/working/llm2vec/experiments/run_mntp.py\"\n\nif os.path.exists(file_path):\n    with open(file_path, \"r\") as f:\n        content = f.read()\n\n    # 1. Strip the telemetry import statement entirely\n    content = content.replace(\n        \"from transformers.utils import send_example_telemetry\", \n        \"# Telemetry stripped for modern transformers compatibility\"\n    )\n\n    # 2. Clean up any remnants of the old TPU utility if still present\n    content = content.replace(\"is_torch_tpu_available,\", \"\")\n    content = content.replace(\"is_torch_tpu_available\", \"\")\n\n    # 3. Inject safe local dummy stubs right at the top of the file\n    fallback_stubs = (\n        \"\\n# Compatibility fallbacks for deprecated HF utilities\\n\"\n        \"is_torch_tpu_available = lambda *args, **kwargs: False\\n\"\n        \"send_example_telemetry = lambda *args, **kwargs: None\\n\\n\"\n    )\n    patched_content = fallback_stubs + content\n\n    with open(file_path, \"w\") as f:\n        f.write(patched_content)\n\n    print(\"[Success] Cleaned up telemetry and TPU dependencies in run_mntp.py.\")\nelse:\n    print(f\"[Error] Could not find script at {file_path}.\")","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import urllib.request\nimport os\n\nfile_path = \"/kaggle/working/llm2vec/experiments/run_mntp.py\"\n\nprint(\"[1/3] Restoring a clean, pristine copy of run_mntp.py...\")\ntry:\n    # Try fetching a fresh copy directly from the repository source\n    url = \"https://raw.githubusercontent.com/McGill-NLP/llm2vec/main/experiments/run_mntp.py\"\n    urllib.request.urlretrieve(url, file_path)\n    print(\" -> Successfully restored original file via source download.\")\nexcept Exception:\n    # Fallback to local git repository rollback if offline\n    os.system(f\"git checkout -- {file_path}\")\n    print(\" -> Successfully restored original file via git checkout.\")\n\nprint(\"[2/3] Applying precise line-by-line patches...\")\nwith open(file_path, \"r\") as f:\n    lines = f.readlines()\n\npatched_lines = []\nfor line in lines:\n    # Safely strip out the deprecated TPU check item from imports\n    if \"is_torch_tpu_available\" in line:\n        line = line.replace(\"is_torch_tpu_available,\", \"\").replace(\"is_torch_tpu_available\", \"\")\n    \n    # Safely neutralize the telemetry logging import line\n    if \"send_example_telemetry\" in line:\n        line = \"    # Telemetry logging removed for modern transformers compatibility\\n\"\n        \n    patched_lines.append(line)\n\n# Prepend explicit safe dummy definitions right at the absolute top of the file\nfallback_headers = [\n    \"is_torch_tpu_available = lambda *args, **kwargs: False\\n\",\n    \"send_example_telemetry = lambda *args, **kwargs: None\\n\\n\"\n]\n\nwith open(file_path, \"w\") as f:\n    f.writelines(fallback_headers + patched_lines)\n\nprint(\"[3/3] Done! The file is clean, uncorrupted, and syntactically sound.\")","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Clone the repository locally to access the experiment files\n!git clone https://github.com/McGill-NLP/llm2vec.git\n!pip install -e ./llm2vec\n!pip install flash-attn --no-build-isolation","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ── Cell 1: Numbers ───────────────────────────────────────────────────────\n\n# Training config (must match mntp_config.json)\nSTEPS            = 2000\nBATCH_PER_DEVICE = 16\nGRAD_ACCUM       = 4\nSEQ_LEN          = 512\nBUFFER_FACTOR    = 10     # 10x safety buffer over minimum\n\n# Split\nTRAIN_RATIO      = 0.98\nVAL_RATIO        = 0.02\n\n# Sampling\nDOCS_PER_DOMAIN  = 100_000   # equal for every domain in AraMix\nNUM_DOMAINS      = 26         # will verify after scan\nARAMIX_TOTAL     = DOCS_PER_DOMAIN * NUM_DOMAINS   # 2,600,000\n\n# Wikipedia: solve wiki / (wiki + aramix) = 0.20\nWIKI_TOTAL       = int(ARAMIX_TOTAL * 0.20 / 0.80)  # 650,000\n\nGRAND_TOTAL      = ARAMIX_TOTAL + WIKI_TOTAL          # 3,250,000\nTRAIN_SIZE       = int(GRAND_TOTAL * TRAIN_RATIO)     # 3,185,000\nVAL_SIZE         = GRAND_TOTAL - TRAIN_SIZE           #    65,000\n\n# Misc\nARAMIX_CONFIG    = 'minhash_deduped'   # or 'sentence_deduped'\nMIN_WORDS        = 30                  # skip very short documents\nSEED             = 42\n\nprint('=' * 50)\nprint(f'AraMix docs   (80%) : {ARAMIX_TOTAL:>10,}  ({DOCS_PER_DOMAIN:,} × {NUM_DOMAINS} domains)')\nprint(f'Wikipedia docs(20%) : {WIKI_TOTAL:>10,}')\nprint(f'Grand total         : {GRAND_TOTAL:>10,}')\nprint(f'Train (98%)         : {TRAIN_SIZE:>10,}')\nprint(f'Val   ( 2%)         : {VAL_SIZE:>10,}')\nprint('=' * 50)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-06-15T23:12:23.476945Z","iopub.execute_input":"2026-06-15T23:12:23.477327Z","iopub.status.idle":"2026-06-15T23:12:23.483177Z","shell.execute_reply.started":"2026-06-15T23:12:23.477310Z","shell.execute_reply":"2026-06-15T23:12:23.482241Z"}},"outputs":[{"name":"stdout","text":"==================================================\nAraMix docs   (80%) :  2,600,000  (100,000 × 26 domains)\nWikipedia docs(20%) :    650,000\nGrand total         :  3,250,000\nTrain (98%)         :  3,185,000\nVal   ( 2%)         :     65,000\n==================================================\n","output_type":"stream"}],"execution_count":2},{"cell_type":"code","source":"# ── Cell 2: Install & imports ─────────────────────────────────────────────\n\n# !pip uninstall -y pyarrow datasets\n# !pip install --no-cache datasets\nimport os\nimport json\nimport random\nfrom pathlib import Path\nfrom collections import defaultdict, Counter\n\nimport pandas as pd\n# import pyarrow as pa\n# import pyarrow.parquet as pq\nfrom datasets import load_dataset\n\nSAVE_DIR = Path('/kaggle/working')\nrandom.seed(SEED)\n\nprint('Ready.')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-06-15T23:12:25.150192Z","iopub.execute_input":"2026-06-15T23:12:25.150430Z","iopub.status.idle":"2026-06-15T23:12:25.154693Z","shell.execute_reply.started":"2026-06-15T23:12:25.150413Z","shell.execute_reply":"2026-06-15T23:12:25.153822Z"}},"outputs":[{"name":"stdout","text":"Ready.\n","output_type":"stream"}],"execution_count":3},{"cell_type":"code","source":"# ── Cell 3: Discover all AraMix domains (quick scan) ─────────────────────\n# Scan first 500K docs to find every unique domain name.\n# Takes ~5 min. Avoids hardcoding domain names.\n\nSCAN_LIMIT = 500_000\nprint(f'Scanning {SCAN_LIMIT:,} docs to discover domains...')\n\nds_scan = load_dataset(\n    'AdaMLLab/AraMix-domain-classified',\n    ARAMIX_CONFIG,\n    split='train',\n    streaming=True,\n    trust_remote_code=True,\n)\n\ndomain_counter = Counter()\nfor i, doc in enumerate(ds_scan):\n    domain_counter[doc.get('domain', 'unknown')] += 1\n    if i + 1 >= SCAN_LIMIT:\n        break\n\nALL_DOMAINS = sorted(domain_counter.keys())\nNUM_DOMAINS = len(ALL_DOMAINS)\n\n# Recalculate totals if domain count differs from estimate\nARAMIX_TOTAL = DOCS_PER_DOMAIN * NUM_DOMAINS\nWIKI_TOTAL   = int(ARAMIX_TOTAL * 0.20 / 0.80)\nGRAND_TOTAL  = ARAMIX_TOTAL + WIKI_TOTAL\nTRAIN_SIZE   = int(GRAND_TOTAL * TRAIN_RATIO)\nVAL_SIZE     = GRAND_TOTAL - TRAIN_SIZE\n\nprint(f'\\nFound {NUM_DOMAINS} domains:\\n')\nprint(f'{\"Domain\":<40} {\"In scan\":>8}  {\"% of scan\":>9}')\nprint('-' * 62)\nfor domain, count in domain_counter.most_common():\n    print(f'{domain:<40} {count:>8,}  {count/SCAN_LIMIT*100:>8.1f}%')\n\nprint(f'\\nUpdated totals:')\nprint(f'  AraMix  : {ARAMIX_TOTAL:,}')\nprint(f'  Wikipedia: {WIKI_TOTAL:,}')\nprint(f'  Train   : {TRAIN_SIZE:,}')\nprint(f'  Val     : {VAL_SIZE:,}')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-06-15T23:12:28.059408Z","iopub.execute_input":"2026-06-15T23:12:28.059668Z","iopub.status.idle":"2026-06-15T23:14:11.575042Z","shell.execute_reply.started":"2026-06-15T23:12:28.059647Z","shell.execute_reply":"2026-06-15T23:14:11.573944Z"}},"outputs":[{"name":"stderr","text":"`trust_remote_code` is not supported anymore.\nPlease check that the Hugging Face dataset 'AdaMLLab/AraMix-domain-classified' isn't based on a loading script and remove `trust_remote_code`.\nIf the dataset is based on a loading script, please ask the dataset author to remove it and convert it to a standard format like Parquet.\n","output_type":"stream"},{"name":"stdout","text":"Scanning 500,000 docs to discover domains...\n","output_type":"stream"},{"output_type":"display_data","data":{"text/plain":"README.md:   0%|          | 0.00/3.09k [00:00<?, ?B/s]","application/vnd.jupyter.widget-view+json":{"version_major":2,"version_minor":0,"model_id":"19a51db3e76f4e4397dc628cab592712"}},"metadata":{}},{"name":"stderr","text":"Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.\n","output_type":"stream"},{"output_type":"display_data","data":{"text/plain":"Resolving data files:   0%|          | 0/1683 [00:00<?, ?it/s]","application/vnd.jupyter.widget-view+json":{"version_major":2,"version_minor":0,"model_id":"435f3d286b1f4fd386e8e654f9c5c6d4"}},"metadata":{}},{"output_type":"display_data","data":{"text/plain":"Resolving data files:   0%|          | 0/1683 [00:00<?, ?it/s]","application/vnd.jupyter.widget-view+json":{"version_major":2,"version_minor":0,"model_id":"a0369e1fbb2249699c962a56b1ef08df"}},"metadata":{}},{"name":"stdout","text":"\nFound 26 domains:\n\nDomain                                    In scan  % of scan\n--------------------------------------------------------------\nNews                                       69,204      13.8%\nSensitive_Subjects                         61,761      12.4%\nBusiness_and_Industrial                    49,011       9.8%\nPeople_and_Society                         48,491       9.7%\nSports                                     37,880       7.6%\nHealth                                     29,972       6.0%\nJobs_and_Education                         23,279       4.7%\nArts_and_Entertainment                     22,013       4.4%\nLaw_and_Government                         20,597       4.1%\nFood_and_Drink                             16,142       3.2%\nHome_and_Garden                            14,249       2.8%\nFinance                                    13,643       2.7%\nBeauty_and_Fitness                         13,127       2.6%\nTravel_and_Transportation                  10,661       2.1%\nComputers_and_Electronics                  10,549       2.1%\nBooks_and_Literature                       10,405       2.1%\nInternet_and_Telecom                        9,967       2.0%\nAutos_and_Vehicles                          7,052       1.4%\nScience                                     5,498       1.1%\nGames                                       5,068       1.0%\nAdult                                       4,971       1.0%\nShopping                                    4,521       0.9%\nReal_Estate                                 3,917       0.8%\nHobbies_and_Leisure                         3,606       0.7%\nPets_and_Animals                            3,106       0.6%\nOnline_Communities                          1,310       0.3%\n\nUpdated totals:\n  AraMix  : 2,600,000\n  Wikipedia: 650,000\n  Train   : 3,185,000\n  Val     : 65,000\n","output_type":"stream"}],"execution_count":4},{"cell_type":"code","source":"# ── Cell 4: Stream & sample AraMix equally by domain ─────────────────────\n# Each domain gets exactly DOCS_PER_DOMAIN documents.\n# Streaming + shuffle buffer gives a random sample — not just the first N.\n\nbucket_counts = Counter({d: 0 for d in ALL_DOMAINS})\naramix_rows   = []\n\nskipped_short = 0\nskipped_full  = 0\ncollected     = 0\n\nprint(f'Sampling {DOCS_PER_DOMAIN:,} docs × {NUM_DOMAINS} domains = {ARAMIX_TOTAL:,} total...')\n\nds = load_dataset(\n    'AdaMLLab/AraMix-domain-classified',\n    ARAMIX_CONFIG,\n    split='train',\n    streaming=True,\n    trust_remote_code=True,\n).shuffle(seed=SEED, buffer_size=100_000)\n\nfor doc in ds:\n    domain = doc.get('domain', 'unknown')\n    text   = doc.get('text', '')\n\n    # Unknown domain — skip\n    if domain not in bucket_counts:\n        continue\n\n    # Bucket full — skip\n    if bucket_counts[domain] >= DOCS_PER_DOMAIN:\n        skipped_full += 1\n        if collected >= ARAMIX_TOTAL:\n            break\n        continue\n\n    # Too short — skip\n    if len(text.split()) < MIN_WORDS:\n        skipped_short += 1\n        continue\n\n    aramix_rows.append({\n        'text':   text,\n        'domain': domain,\n        'source': doc.get('source', 'unknown'),\n    })\n    bucket_counts[domain] += 1\n    collected += 1\n\n    if collected % 250_000 == 0:\n        filled   = sum(1 for d in ALL_DOMAINS if bucket_counts[d] >= DOCS_PER_DOMAIN)\n        pct      = collected / ARAMIX_TOTAL * 100\n        print(f'  [{pct:5.1f}%] {collected:>8,} collected | {filled}/{NUM_DOMAINS} domains full | {skipped_short:,} short skipped')\n\n    if collected >= ARAMIX_TOTAL:\n        break\n\nprint(f'\\nAraMix done: {collected:,} documents collected.')\nprint(f'Skipped (too short) : {skipped_short:,}')\nprint(f'Skipped (bucket full): {skipped_full:,}')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-06-15T23:14:11.575713Z","iopub.execute_input":"2026-06-15T23:14:11.575993Z","iopub.status.idle":"2026-06-16T00:47:46.981827Z","shell.execute_reply.started":"2026-06-15T23:14:11.575977Z","shell.execute_reply":"2026-06-16T00:47:46.975558Z"}},"outputs":[{"name":"stderr","text":"`trust_remote_code` is not supported anymore.\nPlease check that the Hugging Face dataset 'AdaMLLab/AraMix-domain-classified' isn't based on a loading script and remove `trust_remote_code`.\nIf the dataset is based on a loading script, please ask the dataset author to remove it and convert it to a standard format like Parquet.\n","output_type":"stream"},{"name":"stdout","text":"Sampling 100,000 docs × 26 domains = 2,600,000 total...\n","output_type":"stream"},{"output_type":"display_data","data":{"text/plain":"Resolving data files:   0%|          | 0/1683 [00:00<?, ?it/s]","application/vnd.jupyter.widget-view+json":{"version_major":2,"version_minor":0,"model_id":"468af2e446c548939c0273afd063a978"}},"metadata":{}},{"output_type":"display_data","data":{"text/plain":"Resolving data files:   0%|          | 0/1683 [00:00<?, ?it/s]","application/vnd.jupyter.widget-view+json":{"version_major":2,"version_minor":0,"model_id":"07596fe0bdd847abb87419687511bb32"}},"metadata":{}},{"name":"stderr","text":"Got disconnected from remote data host. Retrying in 5sec [1/20]\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00513-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n","output_type":"stream"},{"name":"stdout","text":"  [  9.6%]  250,000 collected | 0/26 domains full | 854 short skipped\n  [ 19.2%]  500,000 collected | 0/26 domains full | 1,730 short skipped\n  [ 28.8%]  750,000 collected | 2/26 domains full | 2,583 short skipped\n","output_type":"stream"},{"name":"stderr","text":"'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00656-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n","output_type":"stream"},{"name":"stdout","text":"  [ 38.5%] 1,000,000 collected | 3/26 domains full | 3,897 short skipped\n  [ 48.1%] 1,250,000 collected | 5/26 domains full | 6,214 short skipped\n","output_type":"stream"},{"name":"stderr","text":"'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00364-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n","output_type":"stream"},{"name":"stdout","text":"  [ 57.7%] 1,500,000 collected | 8/26 domains full | 9,150 short skipped\n  [ 67.3%] 1,750,000 collected | 9/26 domains full | 13,300 short skipped\n","output_type":"stream"},{"name":"stderr","text":"'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01150-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n","output_type":"stream"},{"name":"stdout","text":"  [ 76.9%] 2,000,000 collected | 14/26 domains full | 21,604 short skipped\n","output_type":"stream"},{"name":"stderr","text":"'The read operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00123-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'The read operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01299-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01300-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01575-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00023-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'The read operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00113-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n","output_type":"stream"},{"name":"stdout","text":"  [ 86.5%] 2,250,000 collected | 17/26 domains full | 30,188 short skipped\n","output_type":"stream"},{"name":"stderr","text":"'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01567-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01033-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01251-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00728-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01310-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01062-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01099-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n","output_type":"stream"},{"name":"stdout","text":"  [ 96.2%] 2,500,000 collected | 21/26 domains full | 36,081 short skipped\n","output_type":"stream"},{"name":"stderr","text":"'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00931-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01276-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'The read operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00452-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01276-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01276-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00081-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00240-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'The read operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01068-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00278-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00424-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'The read operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01307-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00596-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'The read operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00064-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'The read operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-00883-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01394-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01020-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01020-of-01683.parquet\nRetrying in 2s [Retry 2/5].\n'_ssl.c:993: The handshake operation timed out' thrown while requesting GET https://huggingface.co/datasets/AdaMLLab/AraMix-domain-classified/resolve/8a88f490e33735f64dd63fa103da1a82c06238a3/minhash_deduped/train-01020-of-01683.parquet\nRetrying in 1s [Retry 1/5].\n","output_type":"stream"},{"name":"stdout","text":"\nAraMix done: 2,600,000 documents collected.\nSkipped (too short) : 38,802\nSkipped (bucket full): 17,818,396\n","output_type":"stream"}],"execution_count":5},{"cell_type":"code","source":"# ── Cell 5: Verify AraMix bucket counts ───────────────────────────────────\n\nprint(f'{\"Domain\":<40} {\"Collected\":>10}  {\"Status\":>8}')\nprint('-' * 62)\n\nincomplete = []\nfor domain in ALL_DOMAINS:\n    count  = bucket_counts[domain]\n    status = '✓' if count >= DOCS_PER_DOMAIN else f'⚠ {count:,}'\n    print(f'{domain:<40} {count:>10,}  {status:>8}')\n    if count < DOCS_PER_DOMAIN:\n        incomplete.append((domain, count))\n\nprint()\nif incomplete:\n    print(f'⚠  {len(incomplete)} domain(s) below target (rare domains — expected):')\n    for d, c in incomplete:\n        print(f'   {d}: {c:,} / {DOCS_PER_DOMAIN:,}')\nelse:\n    print(f'✓  All {NUM_DOMAINS} domains have exactly {DOCS_PER_DOMAIN:,} documents.')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-06-16T00:47:46.982883Z","iopub.execute_input":"2026-06-16T00:47:46.983068Z","iopub.status.idle":"2026-06-16T00:47:46.989215Z","shell.execute_reply.started":"2026-06-16T00:47:46.983052Z","shell.execute_reply":"2026-06-16T00:47:46.988326Z"}},"outputs":[{"name":"stdout","text":"Domain                                    Collected    Status\n--------------------------------------------------------------\nAdult                                       100,000         ✓\nArts_and_Entertainment                      100,000         ✓\nAutos_and_Vehicles                          100,000         ✓\nBeauty_and_Fitness                          100,000         ✓\nBooks_and_Literature                        100,000         ✓\nBusiness_and_Industrial                     100,000         ✓\nComputers_and_Electronics                   100,000         ✓\nFinance                                     100,000         ✓\nFood_and_Drink                              100,000         ✓\nGames                                       100,000         ✓\nHealth                                      100,000         ✓\nHobbies_and_Leisure                         100,000         ✓\nHome_and_Garden                             100,000         ✓\nInternet_and_Telecom                        100,000         ✓\nJobs_and_Education                          100,000         ✓\nLaw_and_Government                          100,000         ✓\nNews                                        100,000         ✓\nOnline_Communities                          100,000         ✓\nPeople_and_Society                          100,000         ✓\nPets_and_Animals                            100,000         ✓\nReal_Estate                                 100,000         ✓\nScience                                     100,000         ✓\nSensitive_Subjects                          100,000         ✓\nShopping                                    100,000         ✓\nSports                                      100,000         ✓\nTravel_and_Transportation                   100,000         ✓\n\n✓  All 26 domains have exactly 100,000 documents.\n","output_type":"stream"}],"execution_count":6},{"cell_type":"code","source":"# ── Cell 6: Sample Arabic Wikipedia (650K articles) ───────────────────────\n# Wikipedia Arabic 20231101.ar has ~1.2M articles.\n# We sample WIKI_TOTAL (~650K = 53% of it).\n# Wikipedia is entity-dense and relation-rich — perfect RE signal for MNTP.\n\nprint(f'Loading Arabic Wikipedia (wikimedia/wikipedia, 20231101.ar)...')\nprint(f'Target: {WIKI_TOTAL:,} articles')\n\nwiki_ds = load_dataset(\n    'wikimedia/wikipedia',\n    '20231101.ar',\n    split='train',\n    streaming=True,\n    trust_remote_code=True,\n).shuffle(seed=SEED, buffer_size=50_000)\n\nwiki_rows      = []\nwiki_skipped   = 0\n\nfor doc in wiki_ds:\n    text = doc.get('text', '').strip()\n\n    # Wikipedia articles can have very short stubs — skip them\n    if len(text.split()) < MIN_WORDS:\n        wiki_skipped += 1\n        continue\n\n    wiki_rows.append({\n        'text':   text,\n        'domain': 'Wikipedia',\n        'source': 'wikimedia/wikipedia',\n    })\n\n    if len(wiki_rows) % 100_000 == 0:\n        pct = len(wiki_rows) / WIKI_TOTAL * 100\n        print(f'  [{pct:5.1f}%] {len(wiki_rows):>8,} collected | {wiki_skipped:,} stubs skipped')\n\n    if len(wiki_rows) >= WIKI_TOTAL:\n        break\n\nprint(f'\\nWikipedia done: {len(wiki_rows):,} articles collected.')\nprint(f'Skipped (stubs < {MIN_WORDS} words): {wiki_skipped:,}')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-06-16T00:48:47.403449Z","iopub.execute_input":"2026-06-16T00:48:47.403697Z","iopub.status.idle":"2026-06-16T00:51:49.906698Z","shell.execute_reply.started":"2026-06-16T00:48:47.403682Z","shell.execute_reply":"2026-06-16T00:51:49.900474Z"}},"outputs":[{"name":"stderr","text":"`trust_remote_code` is not supported anymore.\nPlease check that the Hugging Face dataset 'wikimedia/wikipedia' isn't based on a loading script and remove `trust_remote_code`.\nIf the dataset is based on a loading script, please ask the dataset author to remove it and convert it to a standard format like Parquet.\n","output_type":"stream"},{"name":"stdout","text":"Loading Arabic Wikipedia (wikimedia/wikipedia, 20231101.ar)...\nTarget: 650,000 articles\n","output_type":"stream"},{"output_type":"display_data","data":{"text/plain":"README.md:   0%|          | 0.00/131k [00:00<?, ?B/s]","application/vnd.jupyter.widget-view+json":{"version_major":2,"version_minor":0,"model_id":"4a21d72456664496be3911213a3df510"}},"metadata":{}},{"name":"stdout","text":"  [ 15.4%]  100,000 collected | 11,126 stubs skipped\n  [ 30.8%]  200,000 collected | 21,681 stubs skipped\n  [ 46.2%]  300,000 collected | 31,703 stubs skipped\n  [ 61.5%]  400,000 collected | 42,870 stubs skipped\n  [ 76.9%]  500,000 collected | 53,969 stubs skipped\n  [ 92.3%]  600,000 collected | 64,942 stubs skipped\n\nWikipedia done: 650,000 articles collected.\nSkipped (stubs < 30 words): 69,894\n","output_type":"stream"}],"execution_count":7},{"cell_type":"code","source":"# ── Cell 7: Combine → shuffle → train/val split ───────────────────────────\n\nprint('Combining AraMix + Wikipedia...')\nall_rows = aramix_rows + wiki_rows\nprint(f'  AraMix    : {len(aramix_rows):,}')\nprint(f'  Wikipedia : {len(wiki_rows):,}')\nprint(f'  Total     : {len(all_rows):,}')\n\n# Shuffle before splitting so train/val are not domain-sorted\nprint('\\nShuffling...')\nrandom.shuffle(all_rows)\n\n# Split\nactual_train_size = int(len(all_rows) * TRAIN_RATIO)\ntrain_rows = all_rows[:actual_train_size]\nval_rows   = all_rows[actual_train_size:]\n\nprint(f'\\nSplit:')\nprint(f'  Train : {len(train_rows):,}  ({len(train_rows)/len(all_rows)*100:.1f}%)')\nprint(f'  Val   : {len(val_rows):,}  ({len(val_rows)/len(all_rows)*100:.1f}%)')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-06-16T00:52:00.895149Z","iopub.execute_input":"2026-06-16T00:52:00.895433Z","iopub.status.idle":"2026-06-16T00:52:02.565252Z","shell.execute_reply.started":"2026-06-16T00:52:00.895417Z","shell.execute_reply":"2026-06-16T00:52:02.564246Z"}},"outputs":[{"name":"stdout","text":"Combining AraMix + Wikipedia...\n  AraMix    : 2,600,000\n  Wikipedia : 650,000\n  Total     : 3,250,000\n\nShuffling...\n\nSplit:\n  Train : 3,185,000  (98.0%)\n  Val   : 65,000  (2.0%)\n","output_type":"stream"}],"execution_count":8},{"cell_type":"code","source":"# ── Cell 8: Save train.parquet + val.parquet ──────────────────────────────\n\ndef save_parquet(rows, path):\n    df = pd.DataFrame(rows)\n    df.to_parquet(path, index=False, engine='pyarrow', compression='snappy')\n    size_gb = Path(path).stat().st_size / 1e9\n    print(f'  Saved {len(rows):,} rows → {path}  ({size_gb:.2f} GB)')\n    return df\n\nprint('Saving...')\ntrain_path = SAVE_DIR / 'train.parquet'\nval_path   = SAVE_DIR / 'val.parquet'\n\ntrain_df = save_parquet(train_rows, train_path)\nval_df   = save_parquet(val_rows,   val_path)\n\n# Save metadata alongside\nmeta = {\n    'aramix_config':    ARAMIX_CONFIG,\n    'wiki_subset':      '20231101.ar',\n    'docs_per_domain':  DOCS_PER_DOMAIN,\n    'num_domains':      NUM_DOMAINS,\n    'aramix_total':     len(aramix_rows),\n    'wiki_total':       len(wiki_rows),\n    'grand_total':      len(all_rows),\n    'train_size':       len(train_rows),\n    'val_size':         len(val_rows),\n    'train_ratio':      TRAIN_RATIO,\n    'min_words':        MIN_WORDS,\n    'seed':             SEED,\n    'domains':          ALL_DOMAINS,\n}\nwith open(SAVE_DIR / 'meta.json', 'w', encoding='utf-8') as f:\n    json.dump(meta, f, ensure_ascii=False, indent=2)\n\nprint('\\nDone. Files saved:')\nfor p in [train_path, val_path, SAVE_DIR / 'meta.json']:\n    print(f'  {p}')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-06-16T00:52:08.374502Z","iopub.execute_input":"2026-06-16T00:52:08.374736Z","iopub.status.idle":"2026-06-16T00:53:34.199214Z","shell.execute_reply.started":"2026-06-16T00:52:08.374721Z","shell.execute_reply":"2026-06-16T00:53:34.198031Z"}},"outputs":[{"name":"stdout","text":"Saving...\n  Saved 3,185,000 rows → /kaggle/working/train.parquet  (6.51 GB)\n  Saved 65,000 rows → /kaggle/working/val.parquet  (0.13 GB)\n\nDone. Files saved:\n  /kaggle/working/train.parquet\n  /kaggle/working/val.parquet\n  /kaggle/working/meta.json\n","output_type":"stream"}],"execution_count":9},{"cell_type":"code","source":"# ── Cell 9: Sanity check ──────────────────────────────────────────────────\n\nprint('=== TRAIN ===')\nprint(f'Rows   : {len(train_df):,}')\nprint(f'Columns: {list(train_df.columns)}')\n\nprint('\\nDomain distribution in train:')\ndomain_dist = train_df['domain'].value_counts()\nfor domain, count in domain_dist.items():\n    bar = '█' * int(count / domain_dist.max() * 30)\n    print(f'  {domain:<35} {count:>8,}  {bar}')\n\nprint('\\nSource distribution in train:')\nprint(train_df['source'].value_counts().to_string())\n\ntrain_df['word_count'] = train_df['text'].str.split().str.len()\nprint(f'\\nWord count stats (train):')\nprint(f'  Mean  : {train_df[\"word_count\"].mean():.0f}')\nprint(f'  Median: {train_df[\"word_count\"].median():.0f}')\nprint(f'  Min   : {train_df[\"word_count\"].min()}')\nprint(f'  Max   : {train_df[\"word_count\"].max():,}')\nprint(f'\\nEstimated total tokens (×1.3 tok/word):')\nprint(f'  Train : {train_df[\"word_count\"].sum() * 1.3 / 1e6:.0f}M')\n\nval_df['word_count'] = val_df['text'].str.split().str.len()\nprint(f'  Val   : {val_df[\"word_count\"].sum() * 1.3 / 1e6:.0f}M')\n\nprint('\\n=== VAL ===')\nprint(f'Rows   : {len(val_df):,}')\nprint('Domain distribution in val:')\nprint(val_df['domain'].value_counts().to_string())\n\nprint('\\n✓ All done. Upload /kaggle/working/train.parquet and val.parquet to your dataset.')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-06-16T00:53:34.200044Z","iopub.execute_input":"2026-06-16T00:53:34.200260Z"}},"outputs":[{"name":"stdout","text":"=== TRAIN ===\nRows   : 3,185,000\nColumns: ['text', 'domain', 'source']\n\nDomain distribution in train:\n  Wikipedia                            636,898  ██████████████████████████████\n  Sensitive_Subjects                    98,086  ████\n  Hobbies_and_Leisure                   98,082  ████\n  Law_and_Government                    98,066  ████\n  Science                               98,060  ████\n  People_and_Society                    98,045  ████\n  Sports                                98,021  ████\n  Real_Estate                           98,018  ████\n  Shopping                              98,018  ████\n  Business_and_Industrial               98,017  ████\n  Beauty_and_Fitness                    98,017  ████\n  News                                  98,002  ████\n  Adult                                 98,001  ████\n  Food_and_Drink                        98,000  ████\n  Health                                97,998  ████\n  Computers_and_Electronics             97,996  ████\n  Autos_and_Vehicles                    97,996  ████\n  Pets_and_Animals                      97,992  ████\n  Games                                 97,992  ████\n  Finance                               97,992  ████\n  Home_and_Garden                       97,986  ████\n  Travel_and_Transportation             97,968  ████\n  Arts_and_Entertainment                97,968  ████\n  Internet_and_Telecom                  97,964  ████\n  Jobs_and_Education                    97,960  ████\n  Online_Communities                    97,934  ████\n  Books_and_Literature                  97,923  ████\n\nSource distribution in train:\nsource\nlightonai/ArabicWeb24                            708468\nwikimedia/wikipedia                              636898\nHPLT/HPLT2.0_cleaned                             499454\nHuggingFaceFW/fineweb-2                          465344\nuonlp/CulturaX                                   454111\nallenai/c4                                       305080\nClusterlabAi/101_billion_arabic_words_dataset    113686\nHuggingFaceFW/finepdfs                             1959\n","output_type":"stream"}],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}

__output__.json

@@ -0,0 +1,7 @@
+[{"stream_name":"stderr","time":4.24407107,"data":"/usr/local/lib/python3.12/dist-packages/mistune.py:435: SyntaxWarning: invalid escape sequence '\\|'\n"}
+,{"stream_name":"stderr","time":4.24415726,"data":"  cells[i][c] = re.sub('\\\\\\\\\\|', '|', cell)\n"}
+,{"stream_name":"stderr","time":4.775275103,"data":"/usr/local/lib/python3.12/dist-packages/nbconvert/filters/filter_links.py:36: SyntaxWarning: invalid escape sequence '\\_'\n"}
+,{"stream_name":"stderr","time":4.775366133,"data":"  text = re.sub(r'_', '\\_', text) # Escape underscores in display text\n"}
+,{"stream_name":"stderr","time":6.159286222,"data":"[NbConvertApp] Converting notebook __notebook__.ipynb to html\n"}
+,{"stream_name":"stderr","time":8.080042824,"data":"[NbConvertApp] Writing 421881 bytes to __results__.html\n"}
+]

meta.json

@@ -0,0 +1,42 @@
+{
+  "aramix_config": "minhash_deduped",
+  "wiki_subset": "20231101.ar",
+  "docs_per_domain": 100000,
+  "num_domains": 26,
+  "aramix_total": 2600000,
+  "wiki_total": 650000,
+  "grand_total": 3250000,
+  "train_size": 3185000,
+  "val_size": 65000,
+  "train_ratio": 0.98,
+  "min_words": 30,
+  "seed": 42,
+  "domains": [
+    "Adult",
+    "Arts_and_Entertainment",
+    "Autos_and_Vehicles",
+    "Beauty_and_Fitness",
+    "Books_and_Literature",
+    "Business_and_Industrial",
+    "Computers_and_Electronics",
+    "Finance",
+    "Food_and_Drink",
+    "Games",
+    "Health",
+    "Hobbies_and_Leisure",
+    "Home_and_Garden",
+    "Internet_and_Telecom",
+    "Jobs_and_Education",
+    "Law_and_Government",
+    "News",
+    "Online_Communities",
+    "People_and_Society",
+    "Pets_and_Animals",
+    "Real_Estate",
+    "Science",
+    "Sensitive_Subjects",
+    "Shopping",
+    "Sports",
+    "Travel_and_Transportation"
+  ]
+}

train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17e718200cf6d26231e23215fdb87b22dacaca4badd2d9148c00aeab28a7dbda
+size 6510559457

val.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:43a777083cb210934beca3ae7076b645cfdf134fa348df6bf5373978b2208368
+size 132463329