End of training

README.md

@@ -10,6 +10,8 @@ metrics:
 - precision
 - recall
 - f1
+- accuracy
+- matthews_correlation
 model-index:
 - name: DisamBertCrossEncoder-base
   results: []
@@ -20,13 +22,14 @@ should probably proofread and complete it, then remove this comment. -->
 
 # DisamBertCrossEncoder-base
 
-This model is a fine-tuned version of [answerdotai/ModernBERT-base](https://huggingface.co/answerdotai/ModernBERT-base) on the semcor dataset.
+This model is a fine-tuned version of [answerdotai/ModernBERT-base](https://huggingface.co/answerdotai/ModernBERT-base) on the None dataset.
 It achieves the following results on the evaluation set:
-- Loss: 13.9274
-- Precision: 0.6274
-- Recall: 0.6398
-- F1: 0.6335
-- Matthews: 0.6392
+- Loss: 0.9841
+- Precision: 0.6896
+- Recall: 0.6396
+- F1: 0.6636
+- Accuracy: 0.9412
+- Matthews Correlation: 0.6320
 
 ## Model description
 
@@ -45,35 +48,36 @@ More information needed
 ### Training hyperparameters
 
 The following hyperparameters were used during training:
-- learning_rate: 0.0001
+- learning_rate: 1e-05
 - train_batch_size: 64
 - eval_batch_size: 64
 - seed: 42
+- gradient_accumulation_steps: 5
+- total_train_batch_size: 320
 - optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
-- lr_scheduler_type: inverse_sqrt
-- lr_scheduler_warmup_steps: 1000
+- lr_scheduler_type: cosine
 - num_epochs: 10
 
 ### Training results
 
-| Training Loss | Epoch | Step  | Validation Loss | Precision | Recall | F1     | Matthews |
-|:-------------:|:-----:|:-----:|:---------------:|:---------:|:------:|:------:|:--------:|
-| No log        | 0     | 0     | 427.8458        | 0.5014    | 0.4790 | 0.4899 | 0.4781   |
-| 9.0689        | 1.0   | 3504  | 15.5744         | 0.6010    | 0.6196 | 0.6102 | 0.6190   |
-| 8.5420        | 2.0   | 7008  | 15.4129         | 0.6088    | 0.6253 | 0.6170 | 0.6247   |
-| 7.8106        | 3.0   | 10512 | 14.3562         | 0.6138    | 0.6328 | 0.6232 | 0.6322   |
-| 7.6303        | 4.0   | 14016 | 13.9741         | 0.6157    | 0.6372 | 0.6262 | 0.6366   |
-| 7.6930        | 5.0   | 17520 | 13.8324         | 0.6262    | 0.6402 | 0.6331 | 0.6397   |
-| 7.4897        | 6.0   | 21024 | 13.9649         | 0.6144    | 0.6323 | 0.6232 | 0.6318   |
-| 7.3819        | 7.0   | 24528 | 13.4877         | 0.6273    | 0.6407 | 0.6339 | 0.6401   |
-| 7.4083        | 8.0   | 28032 | 13.7249         | 0.6321    | 0.6402 | 0.6362 | 0.6397   |
-| 7.0140        | 9.0   | 31536 | 13.5219         | 0.6168    | 0.6389 | 0.6277 | 0.6383   |
-| 7.7287        | 10.0  | 35040 | 13.9274         | 0.6274    | 0.6398 | 0.6335 | 0.6392   |
+| Training Loss | Epoch | Step   | Validation Loss | Precision | Recall | F1     | Accuracy | Matthews Correlation |
+|:-------------:|:-----:|:------:|:---------------:|:---------:|:------:|:------:|:--------:|:--------------------:|
+| No log        | 0     | 0      | 430.2531        | 0.0905    | 0.9978 | 0.1660 | 0.0911   | -0.0157              |
+| 0.0672        | 1.0   | 12551  | 0.1555          | 0.6786    | 0.5846 | 0.6281 | 0.9372   | 0.5960               |
+| 0.0550        | 2.0   | 25102  | 0.1447          | 0.7176    | 0.6813 | 0.6990 | 0.9468   | 0.6701               |
+| 0.0427        | 3.0   | 37653  | 0.1498          | 0.7690    | 0.6440 | 0.7010 | 0.9502   | 0.6772               |
+| 0.0309        | 4.0   | 50204  | 0.1779          | 0.6773    | 0.7011 | 0.6890 | 0.9426   | 0.6575               |
+| 0.0179        | 5.0   | 62755  | 0.2554          | 0.7021    | 0.6681 | 0.6847 | 0.9442   | 0.6543               |
+| 0.0092        | 6.0   | 75306  | 0.3257          | 0.6927    | 0.6637 | 0.6779 | 0.9428   | 0.6467               |
+| 0.0047        | 7.0   | 87857  | 0.4757          | 0.6674    | 0.6791 | 0.6732 | 0.9402   | 0.6403               |
+| 0.0022        | 8.0   | 100408 | 0.6664          | 0.6943    | 0.6440 | 0.6682 | 0.9420   | 0.6370               |
+| 0.0011        | 9.0   | 112959 | 0.8230          | 0.6872    | 0.6374 | 0.6613 | 0.9408   | 0.6295               |
+| 0.0009        | 10.0  | 125510 | 0.9841          | 0.6896    | 0.6396 | 0.6636 | 0.9412   | 0.6320               |
 
 
 ### Framework versions
 
-- Transformers 5.2.0
+- Transformers 5.3.0
 - Pytorch 2.10.0+cu128
 - Datasets 4.5.0
 - Tokenizers 0.22.2

config.json

@@ -1,12 +1,9 @@
 {
   "architectures": [
-    "DisamBertSingleSense"
+    "DisamBertCrossEncoder"
   ],
   "attention_bias": false,
   "attention_dropout": 0.0,
-  "auto_map": {
-    "AutoModel": "DisamBertSingleSense.DisamBertSingleSense"
-  },
   "bos_token_id": null,
   "classifier_activation": "gelu",
   "classifier_bias": false,
@@ -15,12 +12,10 @@
   "cls_token_id": 50281,
   "decoder_bias": true,
   "deterministic_flash_attn": false,
-  "dtype": "bfloat16",
+  "dtype": "float32",
   "embedding_dropout": 0.0,
-  "end_token": 50369,
   "eos_token_id": null,
   "global_attn_every_n_layers": 3,
-  "gloss_token": 50370,
   "gradient_checkpointing": false,
   "hidden_activation": "gelu",
   "hidden_size": 768,
@@ -77,9 +72,9 @@
   "sep_token_id": 50282,
   "sparse_pred_ignore_index": -100,
   "sparse_prediction": false,
-  "start_token": 50368,
   "tie_word_embeddings": true,
-  "transformers_version": "5.2.0",
+  "tokenizer_path": "answerdotai/ModernBERT-base",
+  "transformers_version": "5.3.0",
   "use_cache": false,
-  "vocab_size": 50371
+  "vocab_size": 50368
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:16fd64c635d8aa42aea9369b6196dc02bae7771edbb9dc4db1231ab774844f91
-size 298047648
+oid sha256:11ae98a5fdc6ef95a0c62701efd72e8656fa76263c9683a3dc6fd26a7b8e0df1
+size 596071480

modeling_deberta_v2.py

@@ -0,0 +1,1364 @@
+# Copyright 2020 Microsoft and the Hugging Face Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch DeBERTa-v2 model."""
+
+from collections.abc import Sequence
+
+import torch
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, LayerNorm, MSELoss
+
+from ... import initialization as init
+from ...activations import ACT2FN
+from ...modeling_layers import GradientCheckpointingLayer
+from ...modeling_outputs import (
+    BaseModelOutput,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...utils import auto_docstring, logging
+from .configuration_deberta_v2 import DebertaV2Config
+
+
+logger = logging.get_logger(__name__)
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaSelfOutput with DebertaLayerNorm->LayerNorm
+class DebertaV2SelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+@torch.jit.script
+def make_log_bucket_position(relative_pos, bucket_size: int, max_position: int):
+    sign = torch.sign(relative_pos)
+    mid = bucket_size // 2
+    abs_pos = torch.where(
+        (relative_pos < mid) & (relative_pos > -mid),
+        torch.tensor(mid - 1).type_as(relative_pos),
+        torch.abs(relative_pos),
+    )
+    log_pos = (
+        torch.ceil(torch.log(abs_pos / mid) / torch.log(torch.tensor((max_position - 1) / mid)) * (mid - 1)) + mid
+    )
+    bucket_pos = torch.where(abs_pos <= mid, relative_pos.type_as(log_pos), log_pos * sign)
+    return bucket_pos
+
+
+def build_relative_position(query_layer, key_layer, bucket_size: int = -1, max_position: int = -1):
+    """
+    Build relative position according to the query and key
+
+    We assume the absolute position of query \\(P_q\\) is range from (0, query_size) and the absolute position of key
+    \\(P_k\\) is range from (0, key_size), The relative positions from query to key is \\(R_{q \\rightarrow k} = P_q -
+    P_k\\)
+
+    Args:
+        query_size (int): the length of query
+        key_size (int): the length of key
+        bucket_size (int): the size of position bucket
+        max_position (int): the maximum allowed absolute position
+        device (`torch.device`): the device on which tensors will be created.
+
+    Return:
+        `torch.LongTensor`: A tensor with shape [1, query_size, key_size]
+    """
+    query_size = query_layer.size(-2)
+    key_size = key_layer.size(-2)
+
+    q_ids = torch.arange(query_size, dtype=torch.long, device=query_layer.device)
+    k_ids = torch.arange(key_size, dtype=torch.long, device=key_layer.device)
+    rel_pos_ids = q_ids[:, None] - k_ids[None, :]
+    if bucket_size > 0 and max_position > 0:
+        rel_pos_ids = make_log_bucket_position(rel_pos_ids, bucket_size, max_position)
+    rel_pos_ids = rel_pos_ids.to(torch.long)
+    rel_pos_ids = rel_pos_ids[:query_size, :]
+    rel_pos_ids = rel_pos_ids.unsqueeze(0)
+    return rel_pos_ids
+
+
+@torch.jit.script
+# Copied from transformers.models.deberta.modeling_deberta.c2p_dynamic_expand
+def c2p_dynamic_expand(c2p_pos, query_layer, relative_pos):
+    return c2p_pos.expand([query_layer.size(0), query_layer.size(1), query_layer.size(2), relative_pos.size(-1)])
+
+
+@torch.jit.script
+# Copied from transformers.models.deberta.modeling_deberta.p2c_dynamic_expand
+def p2c_dynamic_expand(c2p_pos, query_layer, key_layer):
+    return c2p_pos.expand([query_layer.size(0), query_layer.size(1), key_layer.size(-2), key_layer.size(-2)])
+
+
+@torch.jit.script
+# Copied from transformers.models.deberta.modeling_deberta.pos_dynamic_expand
+def pos_dynamic_expand(pos_index, p2c_att, key_layer):
+    return pos_index.expand(p2c_att.size()[:2] + (pos_index.size(-2), key_layer.size(-2)))
+
+
+@torch.jit.script
+def scaled_size_sqrt(query_layer: torch.Tensor, scale_factor: int):
+    return torch.sqrt(torch.tensor(query_layer.size(-1), dtype=torch.float) * scale_factor)
+
+
+@torch.jit.script
+def build_rpos(query_layer, key_layer, relative_pos, position_buckets: int, max_relative_positions: int):
+    if key_layer.size(-2) != query_layer.size(-2):
+        return build_relative_position(
+            key_layer,
+            key_layer,
+            bucket_size=position_buckets,
+            max_position=max_relative_positions,
+        )
+    else:
+        return relative_pos
+
+
+class DisentangledSelfAttention(nn.Module):
+    """
+    Disentangled self-attention module
+
+    Parameters:
+        config (`DebertaV2Config`):
+            A model config class instance with the configuration to build a new model. The schema is similar to
+            *BertConfig*, for more details, please refer [`DebertaV2Config`]
+
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0:
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads})"
+            )
+        self.num_attention_heads = config.num_attention_heads
+        _attention_head_size = config.hidden_size // config.num_attention_heads
+        self.attention_head_size = getattr(config, "attention_head_size", _attention_head_size)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+        self.query_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+        self.key_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+        self.value_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+
+        self.share_att_key = getattr(config, "share_att_key", False)
+        self.pos_att_type = config.pos_att_type if config.pos_att_type is not None else []
+        self.relative_attention = getattr(config, "relative_attention", False)
+
+        if self.relative_attention:
+            self.position_buckets = getattr(config, "position_buckets", -1)
+            self.max_relative_positions = getattr(config, "max_relative_positions", -1)
+            if self.max_relative_positions < 1:
+                self.max_relative_positions = config.max_position_embeddings
+            self.pos_ebd_size = self.max_relative_positions
+            if self.position_buckets > 0:
+                self.pos_ebd_size = self.position_buckets
+
+            self.pos_dropout = nn.Dropout(config.hidden_dropout_prob)
+
+            if not self.share_att_key:
+                if "c2p" in self.pos_att_type:
+                    self.pos_key_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+                if "p2c" in self.pos_att_type:
+                    self.pos_query_proj = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x, attention_heads) -> torch.Tensor:
+        new_x_shape = x.size()[:-1] + (attention_heads, -1)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3).contiguous().view(-1, x.size(1), x.size(-1))
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        output_attentions=False,
+        query_states=None,
+        relative_pos=None,
+        rel_embeddings=None,
+    ):
+        """
+        Call the module
+
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input states to the module usually the output from previous layer, it will be the Q,K and V in
+                *Attention(Q,K,V)*
+
+            attention_mask (`torch.BoolTensor`):
+                An attention mask matrix of shape [*B*, *N*, *N*] where *B* is the batch size, *N* is the maximum
+                sequence length in which element [i,j] = *1* means the *i* th token in the input can attend to the *j*
+                th token.
+
+            output_attentions (`bool`, *optional*):
+                Whether return the attention matrix.
+
+            query_states (`torch.FloatTensor`, *optional*):
+                The *Q* state in *Attention(Q,K,V)*.
+
+            relative_pos (`torch.LongTensor`):
+                The relative position encoding between the tokens in the sequence. It's of shape [*B*, *N*, *N*] with
+                values ranging in [*-max_relative_positions*, *max_relative_positions*].
+
+            rel_embeddings (`torch.FloatTensor`):
+                The embedding of relative distances. It's a tensor of shape [\\(2 \\times
+                \\text{max_relative_positions}\\), *hidden_size*].
+
+
+        """
+        if query_states is None:
+            query_states = hidden_states
+        query_layer = self.transpose_for_scores(self.query_proj(query_states), self.num_attention_heads)
+        key_layer = self.transpose_for_scores(self.key_proj(hidden_states), self.num_attention_heads)
+        value_layer = self.transpose_for_scores(self.value_proj(hidden_states), self.num_attention_heads)
+
+        rel_att = None
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        scale_factor = 1
+        if "c2p" in self.pos_att_type:
+            scale_factor += 1
+        if "p2c" in self.pos_att_type:
+            scale_factor += 1
+        scale = scaled_size_sqrt(query_layer, scale_factor)
+        attention_scores = torch.bmm(query_layer, key_layer.transpose(-1, -2) / scale.to(dtype=query_layer.dtype))
+        if self.relative_attention:
+            rel_embeddings = self.pos_dropout(rel_embeddings)
+            rel_att = self.disentangled_attention_bias(
+                query_layer, key_layer, relative_pos, rel_embeddings, scale_factor
+            )
+
+        if rel_att is not None:
+            attention_scores = attention_scores + rel_att
+        attention_scores = attention_scores.view(
+            -1, self.num_attention_heads, attention_scores.size(-2), attention_scores.size(-1)
+        )
+
+        attention_mask = attention_mask.bool()
+        attention_scores = attention_scores.masked_fill(~(attention_mask), torch.finfo(query_layer.dtype).min)
+        # bsz x height x length x dimension
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        attention_probs = self.dropout(attention_probs)
+        context_layer = torch.bmm(
+            attention_probs.view(-1, attention_probs.size(-2), attention_probs.size(-1)), value_layer
+        )
+        context_layer = (
+            context_layer.view(-1, self.num_attention_heads, context_layer.size(-2), context_layer.size(-1))
+            .permute(0, 2, 1, 3)
+            .contiguous()
+        )
+        new_context_layer_shape = context_layer.size()[:-2] + (-1,)
+        context_layer = context_layer.view(new_context_layer_shape)
+        if not output_attentions:
+            return (context_layer, None)
+        return (context_layer, attention_probs)
+
+    def disentangled_attention_bias(self, query_layer, key_layer, relative_pos, rel_embeddings, scale_factor):
+        if relative_pos is None:
+            relative_pos = build_relative_position(
+                query_layer,
+                key_layer,
+                bucket_size=self.position_buckets,
+                max_position=self.max_relative_positions,
+            )
+        if relative_pos.dim() == 2:
+            relative_pos = relative_pos.unsqueeze(0).unsqueeze(0)
+        elif relative_pos.dim() == 3:
+            relative_pos = relative_pos.unsqueeze(1)
+        # bsz x height x query x key
+        elif relative_pos.dim() != 4:
+            raise ValueError(f"Relative position ids must be of dim 2 or 3 or 4. {relative_pos.dim()}")
+
+        att_span = self.pos_ebd_size
+        relative_pos = relative_pos.to(device=query_layer.device, dtype=torch.long)
+
+        rel_embeddings = rel_embeddings[0 : att_span * 2, :].unsqueeze(0)
+        if self.share_att_key:
+            pos_query_layer = self.transpose_for_scores(
+                self.query_proj(rel_embeddings), self.num_attention_heads
+            ).repeat(query_layer.size(0) // self.num_attention_heads, 1, 1)
+            pos_key_layer = self.transpose_for_scores(self.key_proj(rel_embeddings), self.num_attention_heads).repeat(
+                query_layer.size(0) // self.num_attention_heads, 1, 1
+            )
+        else:
+            if "c2p" in self.pos_att_type:
+                pos_key_layer = self.transpose_for_scores(
+                    self.pos_key_proj(rel_embeddings), self.num_attention_heads
+                ).repeat(query_layer.size(0) // self.num_attention_heads, 1, 1)  # .split(self.all_head_size, dim=-1)
+            if "p2c" in self.pos_att_type:
+                pos_query_layer = self.transpose_for_scores(
+                    self.pos_query_proj(rel_embeddings), self.num_attention_heads
+                ).repeat(query_layer.size(0) // self.num_attention_heads, 1, 1)  # .split(self.all_head_size, dim=-1)
+
+        score = 0
+        # content->position
+        if "c2p" in self.pos_att_type:
+            scale = scaled_size_sqrt(pos_key_layer, scale_factor)
+            c2p_att = torch.bmm(query_layer, pos_key_layer.transpose(-1, -2))
+            c2p_pos = torch.clamp(relative_pos + att_span, 0, att_span * 2 - 1)
+            c2p_att = torch.gather(
+                c2p_att,
+                dim=-1,
+                index=c2p_pos.squeeze(0).expand([query_layer.size(0), query_layer.size(1), relative_pos.size(-1)]),
+            )
+            score += c2p_att / scale.to(dtype=c2p_att.dtype)
+
+        # position->content
+        if "p2c" in self.pos_att_type:
+            scale = scaled_size_sqrt(pos_query_layer, scale_factor)
+            r_pos = build_rpos(
+                query_layer,
+                key_layer,
+                relative_pos,
+                self.max_relative_positions,
+                self.position_buckets,
+            )
+            p2c_pos = torch.clamp(-r_pos + att_span, 0, att_span * 2 - 1)
+            p2c_att = torch.bmm(key_layer, pos_query_layer.transpose(-1, -2))
+            p2c_att = torch.gather(
+                p2c_att,
+                dim=-1,
+                index=p2c_pos.squeeze(0).expand([query_layer.size(0), key_layer.size(-2), key_layer.size(-2)]),
+            ).transpose(-1, -2)
+            score += p2c_att / scale.to(dtype=p2c_att.dtype)
+
+        return score
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaAttention with Deberta->DebertaV2
+class DebertaV2Attention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.self = DisentangledSelfAttention(config)
+        self.output = DebertaV2SelfOutput(config)
+        self.config = config
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        output_attentions: bool = False,
+        query_states=None,
+        relative_pos=None,
+        rel_embeddings=None,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        self_output, att_matrix = self.self(
+            hidden_states,
+            attention_mask,
+            output_attentions,
+            query_states=query_states,
+            relative_pos=relative_pos,
+            rel_embeddings=rel_embeddings,
+        )
+        if query_states is None:
+            query_states = hidden_states
+        attention_output = self.output(self_output, query_states)
+
+        if output_attentions:
+            return (attention_output, att_matrix)
+        else:
+            return (attention_output, None)
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate with Bert->DebertaV2
+class DebertaV2Intermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaOutput with DebertaLayerNorm->LayerNorm
+class DebertaV2Output(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.config = config
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaLayer with Deberta->DebertaV2
+class DebertaV2Layer(GradientCheckpointingLayer):
+    def __init__(self, config):
+        super().__init__()
+        self.attention = DebertaV2Attention(config)
+        self.intermediate = DebertaV2Intermediate(config)
+        self.output = DebertaV2Output(config)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        query_states=None,
+        relative_pos=None,
+        rel_embeddings=None,
+        output_attentions: bool = False,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        attention_output, att_matrix = self.attention(
+            hidden_states,
+            attention_mask,
+            output_attentions=output_attentions,
+            query_states=query_states,
+            relative_pos=relative_pos,
+            rel_embeddings=rel_embeddings,
+        )
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+
+        if output_attentions:
+            return (layer_output, att_matrix)
+        else:
+            return (layer_output, None)
+
+
+class ConvLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        kernel_size = getattr(config, "conv_kernel_size", 3)
+        groups = getattr(config, "conv_groups", 1)
+        self.conv_act = getattr(config, "conv_act", "tanh")
+        self.conv = nn.Conv1d(
+            config.hidden_size, config.hidden_size, kernel_size, padding=(kernel_size - 1) // 2, groups=groups
+        )
+        self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.config = config
+
+    def forward(self, hidden_states, residual_states, input_mask):
+        out = self.conv(hidden_states.permute(0, 2, 1).contiguous()).permute(0, 2, 1).contiguous()
+        rmask = (1 - input_mask).bool()
+        out.masked_fill_(rmask.unsqueeze(-1).expand(out.size()), 0)
+        out = ACT2FN[self.conv_act](self.dropout(out))
+
+        layer_norm_input = residual_states + out
+        output = self.LayerNorm(layer_norm_input).to(layer_norm_input)
+
+        if input_mask is None:
+            output_states = output
+        else:
+            if input_mask.dim() != layer_norm_input.dim():
+                if input_mask.dim() == 4:
+                    input_mask = input_mask.squeeze(1).squeeze(1)
+                input_mask = input_mask.unsqueeze(2)
+
+            input_mask = input_mask.to(output.dtype)
+            output_states = output * input_mask
+
+        return output_states
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaEmbeddings with DebertaLayerNorm->LayerNorm,Deberta->DebertaV2
+class DebertaV2Embeddings(nn.Module):
+    """Construct the embeddings from word, position and token_type embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+        pad_token_id = getattr(config, "pad_token_id", 0)
+        self.embedding_size = getattr(config, "embedding_size", config.hidden_size)
+        self.word_embeddings = nn.Embedding(config.vocab_size, self.embedding_size, padding_idx=pad_token_id)
+
+        self.position_biased_input = getattr(config, "position_biased_input", True)
+        if not self.position_biased_input:
+            self.position_embeddings = None
+        else:
+            self.position_embeddings = nn.Embedding(config.max_position_embeddings, self.embedding_size)
+
+        if config.type_vocab_size > 0:
+            self.token_type_embeddings = nn.Embedding(config.type_vocab_size, self.embedding_size)
+        else:
+            self.token_type_embeddings = None
+
+        if self.embedding_size != config.hidden_size:
+            self.embed_proj = nn.Linear(self.embedding_size, config.hidden_size, bias=False)
+        else:
+            self.embed_proj = None
+
+        self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.config = config
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer(
+            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
+        )
+
+    def forward(self, input_ids=None, token_type_ids=None, position_ids=None, mask=None, inputs_embeds=None):
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, :seq_length]
+
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+
+        if self.position_embeddings is not None:
+            position_embeddings = self.position_embeddings(position_ids.long())
+        else:
+            position_embeddings = torch.zeros_like(inputs_embeds)
+
+        embeddings = inputs_embeds
+        if self.position_biased_input:
+            embeddings = embeddings + position_embeddings
+        if self.token_type_embeddings is not None:
+            token_type_embeddings = self.token_type_embeddings(token_type_ids)
+            embeddings = embeddings + token_type_embeddings
+
+        if self.embed_proj is not None:
+            embeddings = self.embed_proj(embeddings)
+
+        embeddings = self.LayerNorm(embeddings)
+
+        if mask is not None:
+            if mask.dim() != embeddings.dim():
+                if mask.dim() == 4:
+                    mask = mask.squeeze(1).squeeze(1)
+                mask = mask.unsqueeze(2)
+            mask = mask.to(embeddings.dtype)
+
+            embeddings = embeddings * mask
+
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class DebertaV2Encoder(nn.Module):
+    """Modified BertEncoder with relative position bias support"""
+
+    def __init__(self, config):
+        super().__init__()
+
+        self.layer = nn.ModuleList([DebertaV2Layer(config) for _ in range(config.num_hidden_layers)])
+        self.relative_attention = getattr(config, "relative_attention", False)
+
+        if self.relative_attention:
+            self.max_relative_positions = getattr(config, "max_relative_positions", -1)
+            if self.max_relative_positions < 1:
+                self.max_relative_positions = config.max_position_embeddings
+
+            self.position_buckets = getattr(config, "position_buckets", -1)
+            pos_ebd_size = self.max_relative_positions * 2
+
+            if self.position_buckets > 0:
+                pos_ebd_size = self.position_buckets * 2
+
+            self.rel_embeddings = nn.Embedding(pos_ebd_size, config.hidden_size)
+
+        self.norm_rel_ebd = [x.strip() for x in getattr(config, "norm_rel_ebd", "none").lower().split("|")]
+
+        if "layer_norm" in self.norm_rel_ebd:
+            self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=True)
+
+        self.conv = ConvLayer(config) if getattr(config, "conv_kernel_size", 0) > 0 else None
+        self.gradient_checkpointing = False
+
+    def get_rel_embedding(self):
+        rel_embeddings = self.rel_embeddings.weight if self.relative_attention else None
+        if rel_embeddings is not None and ("layer_norm" in self.norm_rel_ebd):
+            rel_embeddings = self.LayerNorm(rel_embeddings)
+        return rel_embeddings
+
+    def get_attention_mask(self, attention_mask):
+        if attention_mask.dim() <= 2:
+            extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+            attention_mask = extended_attention_mask * extended_attention_mask.squeeze(-2).unsqueeze(-1)
+        elif attention_mask.dim() == 3:
+            attention_mask = attention_mask.unsqueeze(1)
+
+        return attention_mask
+
+    def get_rel_pos(self, hidden_states, query_states=None, relative_pos=None):
+        if self.relative_attention and relative_pos is None:
+            if query_states is not None:
+                relative_pos = build_relative_position(
+                    query_states,
+                    hidden_states,
+                    bucket_size=self.position_buckets,
+                    max_position=self.max_relative_positions,
+                )
+            else:
+                relative_pos = build_relative_position(
+                    hidden_states,
+                    hidden_states,
+                    bucket_size=self.position_buckets,
+                    max_position=self.max_relative_positions,
+                )
+        return relative_pos
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        output_hidden_states=True,
+        output_attentions=False,
+        query_states=None,
+        relative_pos=None,
+        return_dict=True,
+    ):
+        if attention_mask.dim() <= 2:
+            input_mask = attention_mask
+        else:
+            input_mask = attention_mask.sum(-2) > 0
+        attention_mask = self.get_attention_mask(attention_mask)
+        relative_pos = self.get_rel_pos(hidden_states, query_states, relative_pos)
+
+        all_hidden_states: tuple[torch.Tensor] | None = (hidden_states,) if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        next_kv = hidden_states
+        rel_embeddings = self.get_rel_embedding()
+        for i, layer_module in enumerate(self.layer):
+            output_states, attn_weights = layer_module(
+                next_kv,
+                attention_mask,
+                query_states=query_states,
+                relative_pos=relative_pos,
+                rel_embeddings=rel_embeddings,
+                output_attentions=output_attentions,
+            )
+
+            if output_attentions:
+                all_attentions = all_attentions + (attn_weights,)
+
+            if i == 0 and self.conv is not None:
+                output_states = self.conv(hidden_states, output_states, input_mask)
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (output_states,)
+
+            if query_states is not None:
+                query_states = output_states
+                if isinstance(hidden_states, Sequence):
+                    next_kv = hidden_states[i + 1] if i + 1 < len(self.layer) else None
+            else:
+                next_kv = output_states
+
+        if not return_dict:
+            return tuple(v for v in [output_states, all_hidden_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=output_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+
+@auto_docstring
+class DebertaV2PreTrainedModel(PreTrainedModel):
+    config: DebertaV2Config
+    base_model_prefix = "deberta"
+    _keys_to_ignore_on_load_unexpected = ["position_embeddings"]
+    supports_gradient_checkpointing = True
+
+    @torch.no_grad()
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        super()._init_weights(module)
+        if isinstance(module, (LegacyDebertaV2LMPredictionHead, DebertaV2LMPredictionHead)):
+            init.zeros_(module.bias)
+        elif isinstance(module, DebertaV2Embeddings):
+            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
+
+
+@auto_docstring
+# Copied from transformers.models.deberta.modeling_deberta.DebertaModel with Deberta->DebertaV2
+class DebertaV2Model(DebertaV2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embeddings = DebertaV2Embeddings(config)
+        self.encoder = DebertaV2Encoder(config)
+        self.z_steps = 0
+        self.config = config
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, new_embeddings):
+        self.embeddings.word_embeddings = new_embeddings
+
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.Tensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        token_type_ids: torch.Tensor | None = None,
+        position_ids: torch.Tensor | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+        **kwargs,
+    ) -> tuple | BaseModelOutput:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+        )
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask,
+            output_hidden_states=True,
+            output_attentions=output_attentions,
+            return_dict=return_dict,
+        )
+        encoded_layers = encoder_outputs[1]
+
+        if self.z_steps > 1:
+            hidden_states = encoded_layers[-2]
+            layers = [self.encoder.layer[-1] for _ in range(self.z_steps)]
+            query_states = encoded_layers[-1]
+            rel_embeddings = self.encoder.get_rel_embedding()
+            attention_mask = self.encoder.get_attention_mask(attention_mask)
+            rel_pos = self.encoder.get_rel_pos(embedding_output)
+            for layer in layers[1:]:
+                query_states = layer(
+                    hidden_states,
+                    attention_mask,
+                    output_attentions=False,
+                    query_states=query_states,
+                    relative_pos=rel_pos,
+                    rel_embeddings=rel_embeddings,
+                )
+                encoded_layers.append(query_states)
+
+        sequence_output = encoded_layers[-1]
+
+        if not return_dict:
+            return (sequence_output,) + encoder_outputs[(1 if output_hidden_states else 2) :]
+
+        return BaseModelOutput(
+            last_hidden_state=sequence_output,
+            hidden_states=encoder_outputs.hidden_states if output_hidden_states else None,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+# Copied from transformers.models.deberta.modeling_deberta.LegacyDebertaPredictionHeadTransform with Deberta->DebertaV2
+class LegacyDebertaV2PredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.embedding_size = getattr(config, "embedding_size", config.hidden_size)
+
+        self.dense = nn.Linear(config.hidden_size, self.embedding_size)
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        self.LayerNorm = nn.LayerNorm(self.embedding_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+class LegacyDebertaV2LMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = LegacyDebertaV2PredictionHeadTransform(config)
+
+        self.embedding_size = getattr(config, "embedding_size", config.hidden_size)
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(self.embedding_size, config.vocab_size)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+class LegacyDebertaV2OnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = LegacyDebertaV2LMPredictionHead(config)
+
+    def forward(self, sequence_output):
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+class DebertaV2LMPredictionHead(nn.Module):
+    """https://github.com/microsoft/DeBERTa/blob/master/DeBERTa/deberta/bert.py#L270"""
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, elementwise_affine=True)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+    # note that the input embeddings must be passed as an argument
+    def forward(self, hidden_states, word_embeddings):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        hidden_states = torch.matmul(hidden_states, word_embeddings.weight.t()) + self.bias
+        return hidden_states
+
+
+class DebertaV2OnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.lm_head = DebertaV2LMPredictionHead(config)
+
+    # note that the input embeddings must be passed as an argument
+    def forward(self, sequence_output, word_embeddings):
+        prediction_scores = self.lm_head(sequence_output, word_embeddings)
+        return prediction_scores
+
+
+@auto_docstring
+class DebertaV2ForMaskedLM(DebertaV2PreTrainedModel):
+    _tied_weights_keys = {
+        "cls.predictions.decoder.bias": "cls.predictions.bias",
+        "cls.predictions.decoder.weight": "deberta.embeddings.word_embeddings.weight",
+    }
+    _keys_to_ignore_on_load_unexpected = [r"mask_predictions.*"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.legacy = config.legacy
+        self.deberta = DebertaV2Model(config)
+        if self.legacy:
+            self.cls = LegacyDebertaV2OnlyMLMHead(config)
+        else:
+            self._tied_weights_keys = {
+                "lm_predictions.lm_head.weight": "deberta.embeddings.word_embeddings.weight",
+            }
+            self.lm_predictions = DebertaV2OnlyMLMHead(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        if self.legacy:
+            return self.cls.predictions.decoder
+        else:
+            return self.lm_predictions.lm_head.dense
+
+    def set_output_embeddings(self, new_embeddings):
+        if self.legacy:
+            self.cls.predictions.decoder = new_embeddings
+            self.cls.predictions.bias = new_embeddings.bias
+        else:
+            self.lm_predictions.lm_head.dense = new_embeddings
+            self.lm_predictions.lm_head.bias = new_embeddings.bias
+
+    @auto_docstring
+    # Copied from transformers.models.deberta.modeling_deberta.DebertaForMaskedLM.forward with Deberta->DebertaV2
+    def forward(
+        self,
+        input_ids: torch.Tensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        token_type_ids: torch.Tensor | None = None,
+        position_ids: torch.Tensor | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+        labels: torch.Tensor | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+        **kwargs,
+    ) -> tuple | MaskedLMOutput:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
+            config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
+            loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
+        """
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.deberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+        if self.legacy:
+            prediction_scores = self.cls(sequence_output)
+        else:
+            prediction_scores = self.lm_predictions(sequence_output, self.deberta.embeddings.word_embeddings)
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()  # -100 index = padding token
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[1:]
+            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+# Copied from transformers.models.deberta.modeling_deberta.ContextPooler
+class ContextPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.pooler_hidden_size, config.pooler_hidden_size)
+        self.dropout = nn.Dropout(config.pooler_dropout)
+        self.config = config
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+
+        context_token = hidden_states[:, 0]
+        context_token = self.dropout(context_token)
+        pooled_output = self.dense(context_token)
+        pooled_output = ACT2FN[self.config.pooler_hidden_act](pooled_output)
+        return pooled_output
+
+    @property
+    def output_dim(self):
+        return self.config.hidden_size
+
+
+@auto_docstring(
+    custom_intro="""
+    DeBERTa Model transformer with a sequence classification/regression head on top (a linear layer on top of the
+    pooled output) e.g. for GLUE tasks.
+    """
+)
+class DebertaV2ForSequenceClassification(DebertaV2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        num_labels = getattr(config, "num_labels", 2)
+        self.num_labels = num_labels
+
+        self.deberta = DebertaV2Model(config)
+        self.pooler = ContextPooler(config)
+        output_dim = self.pooler.output_dim
+
+        self.classifier = nn.Linear(output_dim, num_labels)
+        drop_out = getattr(config, "cls_dropout", None)
+        drop_out = self.config.hidden_dropout_prob if drop_out is None else drop_out
+        self.dropout = nn.Dropout(drop_out)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.deberta.get_input_embeddings()
+
+    def set_input_embeddings(self, new_embeddings):
+        self.deberta.set_input_embeddings(new_embeddings)
+
+    @auto_docstring
+    # Copied from transformers.models.deberta.modeling_deberta.DebertaForSequenceClassification.forward with Deberta->DebertaV2
+    def forward(
+        self,
+        input_ids: torch.Tensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        token_type_ids: torch.Tensor | None = None,
+        position_ids: torch.Tensor | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+        labels: torch.Tensor | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+        **kwargs,
+    ) -> tuple | SequenceClassifierOutput:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.deberta(
+            input_ids,
+            token_type_ids=token_type_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        encoder_layer = outputs[0]
+        pooled_output = self.pooler(encoder_layer)
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    # regression task
+                    loss_fn = nn.MSELoss()
+                    logits = logits.view(-1).to(labels.dtype)
+                    loss = loss_fn(logits, labels.view(-1))
+                elif labels.dim() == 1 or labels.size(-1) == 1:
+                    label_index = (labels >= 0).nonzero()
+                    labels = labels.long()
+                    if label_index.size(0) > 0:
+                        labeled_logits = torch.gather(
+                            logits, 0, label_index.expand(label_index.size(0), logits.size(1))
+                        )
+                        labels = torch.gather(labels, 0, label_index.view(-1))
+                        loss_fct = CrossEntropyLoss()
+                        loss = loss_fct(labeled_logits.view(-1, self.num_labels).float(), labels.view(-1))
+                    else:
+                        loss = torch.tensor(0).to(logits)
+                else:
+                    log_softmax = nn.LogSoftmax(-1)
+                    loss = -((log_softmax(logits) * labels).sum(-1)).mean()
+            elif self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions
+        )
+
+
+@auto_docstring
+# Copied from transformers.models.deberta.modeling_deberta.DebertaForTokenClassification with Deberta->DebertaV2
+class DebertaV2ForTokenClassification(DebertaV2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.deberta = DebertaV2Model(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.Tensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        token_type_ids: torch.Tensor | None = None,
+        position_ids: torch.Tensor | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+        labels: torch.Tensor | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+        **kwargs,
+    ) -> tuple | TokenClassifierOutput:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.deberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions
+        )
+
+
+@auto_docstring
+class DebertaV2ForQuestionAnswering(DebertaV2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.deberta = DebertaV2Model(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @auto_docstring
+    # Copied from transformers.models.deberta.modeling_deberta.DebertaForQuestionAnswering.forward with Deberta->DebertaV2
+    def forward(
+        self,
+        input_ids: torch.Tensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        token_type_ids: torch.Tensor | None = None,
+        position_ids: torch.Tensor | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+        start_positions: torch.Tensor | None = None,
+        end_positions: torch.Tensor | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+        **kwargs,
+    ) -> tuple | QuestionAnsweringModelOutput:
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.deberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1).contiguous()
+        end_logits = end_logits.squeeze(-1).contiguous()
+
+        total_loss = None
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[1:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@auto_docstring
+class DebertaV2ForMultipleChoice(DebertaV2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        num_labels = getattr(config, "num_labels", 2)
+        self.num_labels = num_labels
+
+        self.deberta = DebertaV2Model(config)
+        self.pooler = ContextPooler(config)
+        output_dim = self.pooler.output_dim
+
+        self.classifier = nn.Linear(output_dim, 1)
+        drop_out = getattr(config, "cls_dropout", None)
+        drop_out = self.config.hidden_dropout_prob if drop_out is None else drop_out
+        self.dropout = nn.Dropout(drop_out)
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.deberta.get_input_embeddings()
+
+    def set_input_embeddings(self, new_embeddings):
+        self.deberta.set_input_embeddings(new_embeddings)
+
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.Tensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        token_type_ids: torch.Tensor | None = None,
+        position_ids: torch.Tensor | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+        labels: torch.Tensor | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+        **kwargs,
+    ) -> tuple | MultipleChoiceModelOutput:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
+            num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
+            `input_ids` above)
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
+
+        flat_input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
+        flat_position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
+        flat_token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
+        flat_attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        flat_inputs_embeds = (
+            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
+            if inputs_embeds is not None
+            else None
+        )
+
+        outputs = self.deberta(
+            flat_input_ids,
+            position_ids=flat_position_ids,
+            token_type_ids=flat_token_type_ids,
+            attention_mask=flat_attention_mask,
+            inputs_embeds=flat_inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        encoder_layer = outputs[0]
+        pooled_output = self.pooler(encoder_layer)
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+        reshaped_logits = logits.view(-1, num_choices)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(reshaped_logits, labels)
+
+        if not return_dict:
+            output = (reshaped_logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return MultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+__all__ = [
+    "DebertaV2ForMaskedLM",
+    "DebertaV2ForMultipleChoice",
+    "DebertaV2ForQuestionAnswering",
+    "DebertaV2ForSequenceClassification",
+    "DebertaV2ForTokenClassification",
+    "DebertaV2Model",
+    "DebertaV2PreTrainedModel",
+]

tokenizer.json

@@ -1046,33 +1046,6 @@
       "rstrip": false,
       "normalized": true,
       "special": false
-    },
-    {
-      "id": 50368,
-      "content": "[START]",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": false,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 50369,
-      "content": "[END]",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": false,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 50370,
-      "content": "[GLOSS]",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": false,
-      "normalized": false,
-      "special": true
     }
   ],
   "normalizer": {

tokenizer_config.json

@@ -2,11 +2,6 @@
   "backend": "tokenizers",
   "clean_up_tokenization_spaces": true,
   "cls_token": "[CLS]",
-  "extra_special_tokens": [
-    "[START]",
-    "[END]",
-    "[GLOSS]"
-  ],
   "is_local": false,
   "mask_token": "[MASK]",
   "model_input_names": [

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7d495bea733a9a1d58c23c5a0a1a14cb54e0ab61bd6b81fc5d20360188143cd1
+oid sha256:76ebe973e3dccc2be47ff0763275e5783a298a4f33403e7a15d8ea9e90eeb842
 size 5265