Upload 4 files

configuration_mbert_greek_news_bert.py

@@ -0,0 +1,21 @@
+from transformers import BertConfig
+
+
+class MBertGreekNewsConfig(BertConfig):
+
+    model_type = "mbert_greek_news"
+
+    def __init__(
+        self,
+        num_labels_class: int = 19,
+        num_labels_ner:   int = 32,
+        ner_loss_weight:  float = 3.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.num_labels_class = num_labels_class
+        self.num_labels_ner   = num_labels_ner
+        self.ner_loss_weight  = ner_loss_weight
+
+
+MBertGreekNewsConfig.register_for_auto_class()

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:895f362839b1c433b65b0e56acc86aaead0e2033e7ea633953ea2b3ef88df0ea
+size 713956908

modeling_mbert_greek_news_bert.py

@@ -0,0 +1,90 @@
+import torch.nn as nn
+from transformers import BertModel, BertPreTrainedModel
+# relative import → required for remote code
+from .configuration_mbert_greek_news import MBertGreekNewsConfig
+
+
+class MBertGreekNews(BertPreTrainedModel):
+    config_class = MBertGreekNewsConfig
+    _auto_class  = "AutoModel"                # appears in auto_map
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config)
+
+        n_cls = config.num_labels_class
+        n_ner = config.num_labels_ner
+        self.ner_loss_weight = getattr(config, "ner_loss_weight", 3.0)
+
+        # ── classification head ─────────────────────────────
+        self.class_dropout = nn.Dropout(0.3)
+        self.class_fc      = nn.Linear(config.hidden_size, 768)
+        self.class_relu    = nn.ReLU()
+        self.classifier    = nn.Linear(768, n_cls)
+
+        # ── NER head ────────────────────────────────────────
+        self.ner_classifier = nn.Linear(config.hidden_size, n_ner)
+
+        # helpers for dynamic-normalised training
+        self.initial_cls_loss = None
+        self.initial_ner_loss = None
+
+        self.init_weights()
+
+    # ----------------------------------------------------------
+    def forward(
+        self,
+        input_ids,
+        attention_mask=None,
+        token_type_ids=None,
+        labels_class=None,
+        labels_ner=None,
+    ):
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            return_dict=True,
+        )
+        seq_out   = outputs.last_hidden_state          # (B, L, H)
+        pooled_out= outputs.pooler_output              # (B, H)
+
+        # ── classification branch ───────────────────────────
+        x = self.class_dropout(pooled_out)
+        x = self.class_fc(x)
+        x = self.class_relu(x)
+        logits_class = self.classifier(x)
+
+        # ── NER branch ──────────────────────────────────────
+        logits_ner = self.ner_classifier(seq_out)
+
+        # inference path
+        if labels_class is None or labels_ner is None:
+            return logits_class, logits_ner
+
+        # — classification loss
+        loss_cls = nn.CrossEntropyLoss()(logits_class, labels_class)
+
+        # — NER loss: summed, averaged over non-pad tokens
+        ner_loss_sum = nn.CrossEntropyLoss(ignore_index=-100, reduction="sum")(
+            logits_ner.view(-1, logits_ner.size(-1)),
+            labels_ner.view(-1),
+        )
+        mask = (labels_ner != -100).view(-1).float()
+        loss_ner = ner_loss_sum / (mask.sum() + 1e-9)
+
+        # — dynamic normalisation
+        if self.initial_cls_loss is None and self.training:
+            self.initial_cls_loss = loss_cls.item()
+        if self.initial_ner_loss is None and self.training:
+            self.initial_ner_loss = loss_ner.item()
+
+        if (self.initial_cls_loss is not None) and (self.initial_ner_loss is not None):
+            norm_cls = loss_cls / (self.initial_cls_loss + 1e-8)
+            norm_ner = loss_ner / (self.initial_ner_loss + 1e-8)
+        else:
+            norm_cls, norm_ner = loss_cls, loss_ner
+
+        loss = norm_cls + self.ner_loss_weight * norm_ner
+        return loss, logits_class, logits_ner

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5a6b23224aa3fb90e0be01245560f0a84d5b537f60de8dbd37a1dd790aacec7
+size 5304