anounymos upload 3 files

config.json

@@ -0,0 +1,20 @@
+{
+  "architectures": [
+    "CAPModel"
+  ],
+  "base_model_name": "roberta-base",
+  "dropout": 0.1,
+  "dtype": "float32",
+  "id2label": {
+    "0": "LABEL_0",
+    "1": "LABEL_1",
+    "2": "LABEL_2"
+  },
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1,
+    "LABEL_2": 2
+  },
+  "model_type": "cap",
+  "transformers_version": "5.13.1"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4cf4506f60e3d3d300061b605b35c1aeb99e90d919d8a22707e0c052a85a7c4
+size 498616084

modeling_cap.py

@@ -0,0 +1,34 @@
+# modeling_cap.py
+import torch
+import torch.nn as nn
+from transformers import AutoModel, PreTrainedModel, PretrainedConfig
+
+class CAPConfig(PretrainedConfig):
+    model_type = "cap"
+    def __init__(self, base_model_name="roberta-base", num_labels=3, dropout=0.1, **kwargs):
+        super().__init__(**kwargs)
+        self.base_model_name = base_model_name
+        self.num_labels = num_labels
+        self.dropout = dropout
+
+class CAPModel(PreTrainedModel):
+    config_class = CAPConfig
+    base_model_prefix = "backbone"
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.backbone = AutoModel.from_pretrained(config.base_model_name)
+        hidden_size = self.backbone.config.hidden_size
+        self.num_labels = config.num_labels
+        self.dropout = nn.Dropout(config.dropout)
+        self.head = nn.Linear(hidden_size, config.num_labels)
+
+    def forward(self, input_ids, attention_mask, token_valid_mask):
+        outputs = self.backbone(input_ids=input_ids, attention_mask=attention_mask)
+        subword_states = self.dropout(outputs.last_hidden_state)
+        token_logits = self.head(subword_states)
+        valid_mask = token_valid_mask.unsqueeze(-1)
+        masked_token_logits = token_logits * valid_mask
+        valid_counts = token_valid_mask.sum(dim=1, keepdim=True).clamp(min=1e-9)
+        sequence_logits = masked_token_logits.sum(dim=1) / valid_counts
+        return sequence_logits, token_logits