init

ELC_ParserBERT_10M_textonly_predictions.json.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:62805b3ad647effbdd913914d89158d728eec3c044ba4daf136b4f245d85b8f4
+size 1184499

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   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.

README.md

@@ -1,3 +1,95 @@
 ---
 license: apache-2.0
 ---
+
+# ELC-ParserBERT
+
+This model is an adaptation of the [Every Layer Counts BERT model](<https://aclanthology.org/2023.conll-babylm.20/>), but it incorporates the `Parser Network` from the [StructFormer](<https://arxiv.org/abs/2012.00857>). It was trained for the [BabyLM 2024 challenge](https://babylm.github.io/index.html)'s Strict-Small track.
+
+## Dataset
+
+The training data for the challenge can be accessed through OSF [here](https://osf.io/ad7qg/). This model was trained on the 10M token training dataset.
+
+### Order in Pretraining
+
+After the segmentation of the data, the segements are ordered in increasing difficulty according to the flesch_reading_ease metric. This ordering can either be maintained by not including the shuffle flag when training or rejected (and allowing shuffling of the data to happen); this model did shuffle the data.
+
+## Hyperparameters
+
+### Base Model
+
+| Hyperparameter | Value |
+| -------------- | ----- |
+| Initial learning rate | 5e-3 |
+| Batch size | 256 |
+| Steps | 13495 |
+| shuffled      | True |
+|attention_probs_dropout_prob | 0.1 |
+| classifier_dropout | 0.2 |
+| hidden_dropout_prob | 0.1 |
+| hidden_size | 384 |
+| intermediate_size | 1024 |
+| layer_norm_eps | 1e-07 |
+| max_position_embeddings | 512 |
+| num_attention_heads | 6 |
+| num_hidden_layers | 12 |
+| vocab_size | 16384 |
+| n_parser_layers | 4 |
+| parser_conv_size |9 |
+
+### Fine-tuning
+
+The fine-tuning parameters were unchanged from the organizer outside of following the ELC-BERT model's patience approach for last year, in particular:
+
+| Hyperparameter | Value |
+| -------------- | ----- |
+| Initial learning rate | 5e-5 |
+| Batch size | 64 |
+| Maximum epochs | 10 |
+| Evaluate every (epochs) | 1 |
+| Patience | 10 (for CoLA, MRPC, RTE, BoolQ, MultiRC, and WSC), 100 (for MNLI, MNLI-MM, QQP, QNLI, and SST-2) |
+| Seed | 12 |
+
+## Credit
+
+As mentioned above, this model is an adapatation of Every Layer Counts (ELC) BERT and StructFormer, the citations and code repositories for which can be found here
+
+* StructFormer
+  * [StructFormer Github](<https://github.com/google-research/google-research/tree/master/structformer>)
+
+  * ```bibtex 
+    @misc{shen2020structformer,
+      title={StructFormer: Joint Unsupervised Induction of Dependency and Constituency Structure from Masked Language Modeling}, 
+      author={Yikang Shen and Yi Tay and Che Zheng and Dara Bahri and Donald Metzler and Aaron Courville},
+      year={2020},
+      eprint={2012.00857},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}}```
+* ELC-BERT:
+  * [ELC-BERT Github](<https://github.com/ltgoslo/elc-bert>)
+  * [ELC-BERT 10M Hugging Face](https://huggingface.co/lgcharpe/ELC_BERT_small_baby_10M)
+  * ```bibtex
+    @inproceedings{georges-gabriel-charpentier-samuel-2023-layers,
+    title = "Not all layers are equally as important: Every Layer Counts {BERT}",
+    author = "Georges Gabriel Charpentier, Lucas  and
+      Samuel, David",
+    editor = "Warstadt, Alex  and
+      Mueller, Aaron  and
+      Choshen, Leshem  and
+      Wilcox, Ethan  and
+      Zhuang, Chengxu  and
+      Ciro, Juan  and
+      Mosquera, Rafael  and
+      Paranjabe, Bhargavi  and
+      Williams, Adina  and
+      Linzen, Tal  and
+      Cotterell, Ryan",
+    booktitle = "Proceedings of the BabyLM Challenge at the 27th Conference on Computational Natural Language Learning",
+    month = dec,
+    year = "2023",
+    address = "Singapore",
+    publisher = "Association for Computational Linguistics",
+    url = "https://aclanthology.org/2023.conll-babylm.20",
+    doi = "10.18653/v1/2023.conll-babylm.20",
+    pages = "238--252",
+    }```

config.json

@@ -0,0 +1,26 @@
+{
+  "architectures": [
+    "LtgBertForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "auto_map": {
+    "AutoConfig": "configuration_ltgbert.LtgBertConfig",
+    "AutoModelForMaskedLM": "modeling_ltgbert.LtgBertForMaskedLM",
+    "AutoModelForSequenceClassification": "modeling_ltgbert.LtgBertForSequenceClassification"
+  },
+  "classifier_dropout": 0.2,
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 384,
+  "intermediate_size": 1024,
+  "layer_norm_eps": 1e-07,
+  "max_position_embeddings": 512,
+  "model_type": "ltgbert",
+  "num_attention_heads": 6,
+  "num_hidden_layers": 12,
+  "output_all_encoded_layers": true,
+  "pad_token_id": 3,
+  "position_bucket_size": 32,
+  "torch_dtype": "float32",
+  "transformers_version": "4.26.0",
+  "vocab_size": 16384
+}

configuration_ltgbert.py

@@ -0,0 +1,106 @@
+# coding=utf-8
+# Copyright 2023 Language Technology Group from University of Oslo and The HuggingFace 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.
+
+""" LTG-BERT configutation """
+
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+LTG_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "bnc-bert-span": "https://huggingface.co/ltg/bnc-bert-span",
+    "bnc-bert-span-2x": "https://huggingface.co/ltg/bnc-bert-span-2x",
+    "bnc-bert-span-0.5x": "https://huggingface.co/ltg/bnc-bert-span-0.5x",
+    "bnc-bert-span-0.25x": "https://huggingface.co/ltg/bnc-bert-span-0.25x",
+    "bnc-bert-span-order": "https://huggingface.co/ltg/bnc-bert-span-order",
+    "bnc-bert-span-document": "https://huggingface.co/ltg/bnc-bert-span-document",
+    "bnc-bert-span-word": "https://huggingface.co/ltg/bnc-bert-span-word",
+    "bnc-bert-span-subword": "https://huggingface.co/ltg/bnc-bert-span-subword",
+    "norbert3-xs": "https://huggingface.co/ltg/norbert3-xs/config.json",
+    "norbert3-small": "https://huggingface.co/ltg/norbert3-small/config.json",
+    "norbert3-base": "https://huggingface.co/ltg/norbert3-base/config.json",
+    "norbert3-large": "https://huggingface.co/ltg/norbert3-large/config.json",
+    "norbert3-oversampled-base": "https://huggingface.co/ltg/norbert3-oversampled-base/config.json",
+    "norbert3-ncc-base": "https://huggingface.co/ltg/norbert3-ncc-base/config.json",
+    "norbert3-nak-base": "https://huggingface.co/ltg/norbert3-nak-base/config.json",
+    "norbert3-nb-base": "https://huggingface.co/ltg/norbert3-nb-base/config.json",
+    "norbert3-wiki-base": "https://huggingface.co/ltg/norbert3-wiki-base/config.json",
+    "norbert3-c4-base": "https://huggingface.co/ltg/norbert3-c4-base/config.json",
+}
+
+
+class LtgBertConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`LtgBertModel`]. It is used to
+    instantiate an LTG-BERT model according to the specified arguments, defining the model architecture.
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+    Args:
+        vocab_size (`int`, *optional*, defaults to 16384):
+            Vocabulary size of the LTG-BERT model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`LtgBertModel`].
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 2048):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        classifier_dropout (`float`, *optional*):
+            The dropout ratio for the classification head.
+    """
+    model_type = "ltgbert"
+
+    def __init__(
+        self,
+        vocab_size=16384,
+        attention_probs_dropout_prob=0.1,
+        hidden_dropout_prob=0.1,
+        hidden_size=768,
+        intermediate_size=2048,
+        max_position_embeddings=512,
+        position_bucket_size=32,
+        num_attention_heads=12,
+        num_hidden_layers=12,
+        layer_norm_eps=1.0e-7,
+        pad_token_id=4,
+        output_all_encoded_layers=True,
+        classifier_dropout=None,
+        **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.output_all_encoded_layers = output_all_encoded_layers
+        self.position_bucket_size = position_bucket_size
+        self.layer_norm_eps = layer_norm_eps
+        self.classifier_dropout = classifier_dropout

modeling_ltgbert.py

@@ -0,0 +1,1294 @@
+# coding=utf-8
+# Copyright 2023 Language Technology Group from University of Oslo and The HuggingFace Inc. team.
+# And Copyright 2024 The Google Research Authors.
+#
+# 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.
+
+# Base implementation of the LTG-BERT/ELC-BERT Model is from Language Technology Group from University of Oslo and The HuggingFace Inc., Team
+# The StructFormer components is from The Google Research Authors - the authors were Yikang Shen and Yi Tay and Che Zheng and Dara Bahri and Donald Metzler and Aaron Courville
+# (and the code can be from here: https://github.com/google-research/google-research/tree/master/structformer), both were using Apache license, Version 2.0
+
+""" PyTorch LTG-(ELC)-ParserBERT model."""
+
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.utils import checkpoint
+
+from .configuration_ltgbert import LtgBertConfig
+from transformers.modeling_utils import PreTrainedModel
+from transformers.activations import gelu_new
+from transformers.modeling_outputs import (
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+    BaseModelOutput,
+)
+from transformers.pytorch_utils import softmax_backward_data
+from transformers.utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+)
+
+
+_CHECKPOINT_FOR_DOC = "ltg/bnc-bert-span"
+_CONFIG_FOR_DOC = "LtgBertConfig"
+
+
+LTG_BERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "bnc-bert-span",
+    "bnc-bert-span-2x",
+    "bnc-bert-span-0.5x",
+    "bnc-bert-span-0.25x",
+    "bnc-bert-span-order",
+    "bnc-bert-span-document",
+    "bnc-bert-span-word",
+    "bnc-bert-span-subword",
+    "norbert3-xs",
+    "norbert3-small",
+    "norbert3-base",
+    "norbert3-large",
+    "norbert3-oversampled-base",
+    "norbert3-ncc-base",
+    "norbert3-nak-base",
+    "norbert3-nb-base",
+    "norbert3-wiki-base",
+    "norbert3-c4-base",
+]
+
+
+class Conv1d(nn.Module):
+    """1D convolution layer."""
+
+    def __init__(self, hidden_size, kernel_size, dilation=1):
+        """Initialization.
+
+        Args:
+          hidden_size: dimension of input embeddings
+          kernel_size: convolution kernel size
+          dilation: the spacing between the kernel points
+        """
+        super(Conv1d, self).__init__()
+
+        if kernel_size % 2 == 0:
+            padding = (kernel_size // 2) * dilation
+            self.shift = True
+        else:
+            padding = ((kernel_size - 1) // 2) * dilation
+            self.shift = False
+        self.conv = nn.Conv1d(
+            hidden_size, hidden_size, kernel_size, padding=padding, dilation=dilation
+        )
+
+    def forward(self, x):
+        """Compute convolution.
+
+        Args:
+          x: input embeddings
+        Returns:
+          conv_output: convolution results
+        """
+
+        if self.shift:
+            return self.conv(x.transpose(1, 2)).transpose(1, 2)[:, 1:]
+        else:
+            return self.conv(x.transpose(1, 2)).transpose(1, 2)
+
+
+def cumprod(x, reverse=False, exclusive=False):
+    """cumulative product."""
+    if reverse:
+        x = x.flip([-1])
+
+    if exclusive:
+        x = F.pad(x[:, :, :-1], (1, 0), value=1)
+
+    cx = x.cumprod(-1)
+
+    if reverse:
+        cx = cx.flip([-1])
+    return cx
+
+
+def cumsum(x, reverse=False, exclusive=False):
+    """cumulative sum."""
+    bsz, _, length = x.size()
+    device = x.device
+    if reverse:
+        if exclusive:
+            w = torch.ones([bsz, length, length], device=device).tril(-1)
+        else:
+            w = torch.ones([bsz, length, length], device=device).tril(0)
+        cx = torch.bmm(x, w)
+    else:
+        if exclusive:
+            w = torch.ones([bsz, length, length], device=device).triu(1)
+        else:
+            w = torch.ones([bsz, length, length], device=device).triu(0)
+        cx = torch.bmm(x, w)
+    return cx
+
+
+def cummin(x, reverse=False, exclusive=False, max_value=1e4):
+    """cumulative min."""
+    if reverse:
+        if exclusive:
+            x = F.pad(x[:, :, 1:], (0, 1), value=max_value)
+        x = x.flip([-1]).cummin(-1)[0].flip([-1])
+    else:
+        if exclusive:
+            x = F.pad(x[:, :, :-1], (1, 0), value=max_value)
+        x = x.cummin(-1)[0]
+    return x
+
+
+class ParserNetwork(nn.Module):
+    def __init__(
+        self,
+        config,
+        pad=0,
+        n_parser_layers=4,
+        conv_size=9,
+        relations=("head", "child"),
+        weight_act="softmax",
+    ):
+        """
+        hidden_size: dimension of input embeddings
+        nlayers: number of layers
+        ntokens: number of output categories
+        nhead: number of self-attention heads
+        dropout: dropout rate
+        pad: pad token index
+        n_parser_layers: number of parsing layers
+        conv_size: convolution kernel size for parser
+        relations: relations that are used to compute self attention
+        weight_act: relations distribution activation function
+        """
+        super(ParserNetwork, self).__init__()
+        self.hidden_size = config.hidden_size
+        self.num_hidden_layers = config.num_hidden_layers
+        self.num_attention_heads = config.num_attention_heads
+
+        self.parser_layers = nn.ModuleList(
+            [
+                nn.Sequential(
+                    Conv1d(self.hidden_size, conv_size),
+                    nn.LayerNorm(self.hidden_size, elementwise_affine=False),
+                    nn.Tanh(),
+                )
+                for _ in range(n_parser_layers)
+            ]
+        )
+
+        self.distance_ff = nn.Sequential(
+            Conv1d(self.hidden_size, 2),
+            nn.LayerNorm(self.hidden_size, elementwise_affine=False),
+            nn.Tanh(),
+            nn.Linear(self.hidden_size, 1),
+        )
+
+        self.height_ff = nn.Sequential(
+            nn.Linear(self.hidden_size, self.hidden_size),
+            nn.LayerNorm(self.hidden_size, elementwise_affine=False),
+            nn.Tanh(),
+            nn.Linear(self.hidden_size, 1),
+        )
+
+        n_rel = len(relations)
+        self._rel_weight = nn.Parameter(
+            torch.zeros((self.num_hidden_layers, self.num_attention_heads, n_rel))
+        )
+        self._rel_weight.data.normal_(0, 0.1)
+
+        self._scaler = nn.Parameter(torch.zeros(2))
+
+        self.n_parse_layers = n_parser_layers
+        self.weight_act = weight_act
+        self.relations = relations
+        self.pad = pad
+
+    @property
+    def scaler(self):
+        return self._scaler.exp()
+
+    @property
+    def rel_weight(self):
+        if self.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight)
+        elif self.weight_act == "softmax":
+            return torch.softmax(self._rel_weight, dim=-1)
+
+    def parse(self, x, h):
+        """
+        Parse input sentence.
+        Args:
+            x: input tokens (required).
+            h: static embeddings
+        Returns:
+            distance: syntactic distance
+            height: syntactic height
+        """
+
+        mask = x != self.pad
+        mask_shifted = F.pad(mask[:, 1:], (0, 1), value=0)
+
+        for i in range(self.n_parse_layers):
+            h = h.masked_fill(~mask[:, :, None], 0)
+            h = self.parser_layers[i](h)
+
+        height = self.height_ff(h).squeeze(-1)
+        height.masked_fill_(~mask, -1e4)
+
+        distance = self.distance_ff(h).squeeze(-1)
+        distance.masked_fill_(~mask_shifted, 1e4)
+
+        # Calbrating the distance and height to the same level
+        length = distance.size(1)
+        height_max = height[:, None, :].expand(-1, length, -1)
+        height_max = torch.cummax(
+            height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e4, dim=-1
+        )[0].triu(0)
+
+        margin_left = torch.relu(
+            F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e4) - height_max
+        )
+        margin_right = torch.relu(distance[:, None, :] - height_max)
+        margin = torch.where(
+            margin_left > margin_right, margin_right, margin_left
+        ).triu(0)
+
+        margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+        margin.masked_fill_(~margin_mask, 0)
+        margin = margin.max()
+
+        distance = distance - margin
+
+        return distance, height
+
+    def compute_block(self, distance, height):
+        """Compute constituents from distance and height."""
+
+        beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler[0]
+
+        gamma = torch.sigmoid(-beta_logits)
+        ones = torch.ones_like(gamma)
+
+        block_mask_left = cummin(
+            gamma.tril(-1) + ones.triu(0), reverse=True, max_value=1
+        )
+        block_mask_left = block_mask_left - F.pad(
+            block_mask_left[:, :, :-1], (1, 0), value=0
+        )
+        block_mask_left.tril_(0)
+
+        block_mask_right = cummin(
+            gamma.triu(0) + ones.tril(-1), exclusive=True, max_value=1
+        )
+        block_mask_right = block_mask_right - F.pad(
+            block_mask_right[:, :, 1:], (0, 1), value=0
+        )
+        block_mask_right.triu_(0)
+
+        block_p = block_mask_left[:, :, :, None] * block_mask_right[:, :, None, :]
+        block = cumsum(block_mask_left).tril(0) + cumsum(
+            block_mask_right, reverse=True
+        ).triu(1)
+
+        return block_p, block
+
+    def compute_head(self, height):
+        """Estimate head for each constituent."""
+
+        _, length = height.size()
+        head_logits = height * self.scaler[1]
+        index = torch.arange(length, device=height.device)
+
+        mask = (index[:, None, None] <= index[None, None, :]) * (
+            index[None, None, :] <= index[None, :, None]
+        )
+        head_logits = head_logits[:, None, None, :].repeat(1, length, length, 1)
+        head_logits.masked_fill_(~mask[None, :, :, :], -1e4)
+
+        head_p = torch.softmax(head_logits, dim=-1)
+
+        return head_p
+
+    def generate_mask(self, x, distance, height):
+        """Compute head and cibling distribution for each token."""
+
+        batch_size, length = x.size()
+
+        eye = torch.eye(length, device=x.device, dtype=torch.bool)
+        eye = eye[None, :, :].expand((batch_size, -1, -1))
+
+        block_p, block = self.compute_block(distance, height)
+        head_p = self.compute_head(height)
+        head = torch.einsum("blij,bijh->blh", block_p, head_p)
+        head = head.masked_fill(eye, 0)
+        child = head.transpose(1, 2)
+        cibling = torch.bmm(head, child).masked_fill(eye, 0)
+
+        rel_list = []
+        if "head" in self.relations:
+            rel_list.append(head)
+        if "child" in self.relations:
+            rel_list.append(child)
+        if "cibling" in self.relations:
+            rel_list.append(cibling)
+
+        rel = torch.stack(rel_list, dim=1)
+
+        rel_weight = self.rel_weight
+
+        dep = torch.einsum("lhr,brij->lbhij", rel_weight, rel)
+        att_mask = dep.reshape(
+            self.num_hidden_layers, batch_size, self.num_attention_heads, length, length
+        )
+
+        return att_mask, cibling, head, block
+
+    def forward(self, x, embeddings):
+        """
+        Pass the x tokens through the parse network, get the syntactic height and distances
+        and compute the distribution for each token
+        """
+        
+        x = torch.transpose(x, 0, 1)
+        embeddings = torch.transpose(embeddings, 0, 1)
+
+        distance, height = self.parse(x, embeddings)
+        att_mask, cibling, head, block = self.generate_mask(x, distance, height)
+        return att_mask, cibling, head, block
+
+
+class Encoder(nn.Module):
+    def __init__(self, config, activation_checkpointing=False):
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [EncoderLayer(config, i) for i in range(config.num_hidden_layers)]
+        )
+
+        for i, layer in enumerate(self.layers):
+            layer.mlp.mlp[1].weight.data *= math.sqrt(1.0 / (2.0 * (1 + i)))
+            layer.mlp.mlp[-2].weight.data *= math.sqrt(1.0 / (2.0 * (1 + i)))
+
+        self.activation_checkpointing = activation_checkpointing
+
+    def forward(self, hidden_states, attention_mask, relative_embedding):
+        hidden_states, attention_probs = [hidden_states], []
+
+        for i in range(len(self.layers)):
+            if self.activation_checkpointing:
+                hidden_state, attention_p = checkpoint.checkpoint(
+                    self.layers[i], hidden_states, attention_mask, relative_embedding
+                )
+            else:
+                hidden_state, attention_p = self.layers[i](
+                    hidden_states, attention_mask[i], relative_embedding
+                )
+
+            hidden_states.append(hidden_state)
+            attention_probs.append(attention_p)
+
+        return hidden_states, attention_probs
+
+
+class MaskClassifier(nn.Module):
+    def __init__(self, config, subword_embedding):
+        super().__init__()
+        self.nonlinearity = nn.Sequential(
+            nn.LayerNorm(
+                config.hidden_size, config.layer_norm_eps, elementwise_affine=False
+            ),
+            nn.Linear(config.hidden_size, config.hidden_size),
+            nn.GELU(),
+            nn.LayerNorm(
+                config.hidden_size, config.layer_norm_eps, elementwise_affine=False
+            ),
+            nn.Dropout(config.hidden_dropout_prob),
+            nn.Linear(subword_embedding.size(1), subword_embedding.size(0)),
+        )
+        self.initialize(config.hidden_size, subword_embedding)
+
+    def initialize(self, hidden_size, embedding):
+        std = math.sqrt(2.0 / (5.0 * hidden_size))
+        nn.init.trunc_normal_(
+            self.nonlinearity[1].weight, mean=0.0, std=std, a=-2 * std, b=2 * std
+        )
+        self.nonlinearity[-1].weight = embedding
+        self.nonlinearity[1].bias.data.zero_()
+        self.nonlinearity[-1].bias.data.zero_()
+
+    def forward(self, x, masked_lm_labels=None):
+        if masked_lm_labels is not None:
+            x = torch.index_select(
+                x.flatten(0, 1),
+                0,
+                torch.nonzero(masked_lm_labels.flatten() != -100).squeeze(),
+            )
+        x = self.nonlinearity(x)
+        return x
+
+
+class EncoderLayer(nn.Module):
+    def __init__(self, config, layer_num):
+        super().__init__()
+        self.attention = Attention(config)
+        self.mlp = FeedForward(config)
+        temp = torch.zeros(layer_num + 1)
+        temp[-1] = 1
+        self.prev_layer_weights = nn.Parameter(temp)
+
+    def forward(self, hidden_states, padding_mask, relative_embedding):
+        prev_layer_weights = F.softmax(self.prev_layer_weights, dim=-1)
+        x = prev_layer_weights[0] * hidden_states[0]
+        for i, hidden_state in enumerate(hidden_states[1:]):
+            x = x + prev_layer_weights[i + 1] * hidden_state
+        attention_output, attention_probs = self.attention(
+            x, padding_mask, relative_embedding
+        )
+        x = attention_output
+        x = x + self.mlp(x)
+        return x, attention_probs
+
+
+class GeGLU(nn.Module):
+    def forward(self, x):
+        x, gate = x.chunk(2, dim=-1)
+        x = x * gelu_new(gate)
+        return x
+
+
+class FeedForward(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            nn.LayerNorm(
+                config.hidden_size, eps=config.layer_norm_eps, elementwise_affine=False
+            ),
+            nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False),
+            GeGLU(),
+            nn.LayerNorm(
+                config.intermediate_size,
+                eps=config.layer_norm_eps,
+                elementwise_affine=False,
+            ),
+            nn.Linear(config.intermediate_size, config.hidden_size, bias=False),
+            nn.Dropout(config.hidden_dropout_prob),
+        )
+        self.initialize(config.hidden_size)
+
+    def initialize(self, hidden_size):
+        std = math.sqrt(2.0 / (5.0 * hidden_size))
+        nn.init.trunc_normal_(
+            self.mlp[1].weight, mean=0.0, std=std, a=-2 * std, b=2 * std
+        )
+        nn.init.trunc_normal_(
+            self.mlp[-2].weight, mean=0.0, std=std, a=-2 * std, b=2 * std
+        )
+
+    def forward(self, x):
+        return self.mlp(x)
+
+
+class MaskedSoftmax(torch.autograd.Function):
+    @staticmethod
+    def forward(self, x, mask, dim):
+        self.dim = dim
+        x.masked_fill_(mask, float("-inf"))
+        x = torch.softmax(x, self.dim)
+        x.masked_fill_(mask, 0.0)
+        self.save_for_backward(x)
+        return x
+
+    @staticmethod
+    def backward(self, grad_output):
+        (output,) = self.saved_tensors
+        input_grad = softmax_backward_data(self, grad_output, output, self.dim, output)
+        return input_grad, None, None
+
+
+class Attention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+
+        self.config = config
+
+        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 heads {config.num_attention_heads}"
+            )
+
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_size = config.hidden_size // config.num_attention_heads
+
+        self.in_proj_qk = nn.Linear(
+            config.hidden_size, 2 * config.hidden_size, bias=True
+        )
+        self.in_proj_v = nn.Linear(config.hidden_size, config.hidden_size, bias=True)
+        self.out_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=True)
+
+        self.pre_layer_norm = nn.LayerNorm(
+            config.hidden_size, config.layer_norm_eps, elementwise_affine=False
+        )
+        self.post_layer_norm = nn.LayerNorm(
+            config.hidden_size, config.layer_norm_eps, elementwise_affine=True
+        )
+
+        position_indices = torch.arange(
+            config.max_position_embeddings, dtype=torch.long
+        ).unsqueeze(1) - torch.arange(
+            config.max_position_embeddings, dtype=torch.long
+        ).unsqueeze(
+            0
+        )
+        position_indices = self.make_log_bucket_position(
+            position_indices,
+            config.position_bucket_size,
+            config.max_position_embeddings,
+        )
+        position_indices = config.position_bucket_size - 1 + position_indices
+        self.register_buffer("position_indices", position_indices, persistent=True)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.scale = 1.0 / math.sqrt(3 * self.head_size)
+        self.initialize()
+
+    def make_log_bucket_position(self, relative_pos, bucket_size, max_position):
+        sign = torch.sign(relative_pos)
+        mid = bucket_size // 2
+        abs_pos = torch.where(
+            (relative_pos < mid) & (relative_pos > -mid),
+            mid - 1,
+            torch.abs(relative_pos).clamp(max=max_position - 1),
+        )
+        log_pos = (
+            torch.ceil(
+                torch.log(abs_pos / mid)
+                / math.log((max_position - 1) / mid)
+                * (mid - 1)
+            ).int()
+            + mid
+        )
+        bucket_pos = torch.where(abs_pos <= mid, relative_pos, log_pos * sign).long()
+        return bucket_pos
+
+    def initialize(self):
+        std = math.sqrt(2.0 / (5.0 * self.hidden_size))
+        nn.init.trunc_normal_(
+            self.in_proj_qk.weight, mean=0.0, std=std, a=-2 * std, b=2 * std
+        )
+        nn.init.trunc_normal_(
+            self.in_proj_v.weight, mean=0.0, std=std, a=-2 * std, b=2 * std
+        )
+        nn.init.trunc_normal_(
+            self.out_proj.weight, mean=0.0, std=std, a=-2 * std, b=2 * std
+        )
+        self.in_proj_qk.bias.data.zero_()
+        self.in_proj_v.bias.data.zero_()
+        self.out_proj.bias.data.zero_()
+
+    def compute_attention_scores(self, hidden_states, relative_embedding):
+        key_len, batch_size, _ = hidden_states.size()
+        query_len = key_len
+
+        if self.position_indices.size(0) < query_len:
+            position_indices = torch.arange(query_len, dtype=torch.long).unsqueeze(
+                1
+            ) - torch.arange(query_len, dtype=torch.long).unsqueeze(0)
+            position_indices = self.make_log_bucket_position(
+                position_indices, self.position_bucket_size, 512
+            )
+            position_indices = self.position_bucket_size - 1 + position_indices
+            self.position_indices = position_indices.to(hidden_states.device)
+
+        hidden_states = self.pre_layer_norm(hidden_states)
+
+        query, key = self.in_proj_qk(hidden_states).chunk(2, dim=2)  # shape: [T, B, D]
+        value = self.in_proj_v(hidden_states)  # shape: [T, B, D]
+
+        query = query.reshape(
+            query_len, batch_size * self.num_heads, self.head_size
+        ).transpose(0, 1)
+        key = key.reshape(
+            key_len, batch_size * self.num_heads, self.head_size
+        ).transpose(0, 1)
+        value = value.view(
+            key_len, batch_size * self.num_heads, self.head_size
+        ).transpose(0, 1)
+
+        attention_scores = torch.bmm(query, key.transpose(1, 2) * self.scale)
+
+        query_pos, key_pos = self.in_proj_qk(self.dropout(relative_embedding)).chunk(
+            2, dim=-1
+        )  # shape: [2T-1, D]
+        query_pos = query_pos.view(
+            -1, self.num_heads, self.head_size
+        )  # shape: [2T-1, H, D]
+        key_pos = key_pos.view(
+            -1, self.num_heads, self.head_size
+        )  # shape: [2T-1, H, D]
+
+        query = query.view(batch_size, self.num_heads, query_len, self.head_size)
+        key = key.view(batch_size, self.num_heads, query_len, self.head_size)
+
+        attention_c_p = torch.einsum(
+            "bhqd,khd->bhqk", query, key_pos.squeeze(1) * self.scale
+        )
+        attention_p_c = torch.einsum(
+            "bhkd,qhd->bhqk", key * self.scale, query_pos.squeeze(1)
+        )
+
+        position_indices = self.position_indices[:query_len, :key_len].expand(
+            batch_size, self.num_heads, -1, -1
+        )
+        attention_c_p = attention_c_p.gather(3, position_indices)
+        attention_p_c = attention_p_c.gather(2, position_indices)
+
+        attention_scores = attention_scores.view(
+            batch_size, self.num_heads, query_len, key_len
+        )
+        attention_scores.add_(attention_c_p)
+        attention_scores.add_(attention_p_c)
+
+        return attention_scores, value
+
+    def compute_output(self, attention_probs, value):
+        attention_probs = self.dropout(attention_probs)
+        context = torch.bmm(attention_probs.flatten(0, 1), value)  # shape: [B*H, Q, D]
+        context = context.transpose(0, 1).reshape(
+            context.size(1), -1, self.hidden_size
+        )  # shape: [Q, B, H*D]
+        context = self.out_proj(context)
+        context = self.post_layer_norm(context)
+        context = self.dropout(context)
+        return context
+
+    def forward(self, hidden_states, attention_mask, relative_embedding):
+        attention_scores, value = self.compute_attention_scores(
+            hidden_states, relative_embedding
+        )
+        attention_probs = torch.sigmoid(attention_scores) * attention_mask
+        return self.compute_output(attention_probs, value), attention_probs.detach()
+
+
+class Embedding(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.word_embedding = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.word_layer_norm = nn.LayerNorm(
+            config.hidden_size, eps=config.layer_norm_eps, elementwise_affine=False
+        )
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+        self.relative_embedding = nn.Parameter(
+            torch.empty(2 * config.position_bucket_size - 1, config.hidden_size)
+        )
+        self.relative_layer_norm = nn.LayerNorm(
+            config.hidden_size, eps=config.layer_norm_eps
+        )
+
+        self.initialize()
+
+    def initialize(self):
+        std = math.sqrt(2.0 / (5.0 * self.hidden_size))
+        nn.init.trunc_normal_(
+            self.relative_embedding, mean=0.0, std=std, a=-2 * std, b=2 * std
+        )
+        nn.init.trunc_normal_(
+            self.word_embedding.weight, mean=0.0, std=std, a=-2 * std, b=2 * std
+        )
+
+    def forward(self, input_ids):
+        word_embedding = self.dropout(
+            self.word_layer_norm(self.word_embedding(input_ids))
+        )
+        relative_embeddings = self.relative_layer_norm(self.relative_embedding)
+        return word_embedding, relative_embeddings
+
+
+#
+# HuggingFace wrappers
+#
+
+
+class LtgBertPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = LtgBertConfig
+    base_model_prefix = "bnc-bert"
+    supports_gradient_checkpointing = True
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, Encoder):
+            module.activation_checkpointing = value
+
+    def _init_weights(self, _):
+        pass  # everything is already initialized
+
+
+LTG_BERT_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+    Parameters:
+        config ([`LtgBertConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+LTG_BERT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+            [What are attention masks?](../glossary#attention-mask)
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare LTG-BERT transformer outputting raw hidden-states without any specific head on top.",
+    LTG_BERT_START_DOCSTRING,
+)
+class LtgBertModel(LtgBertPreTrainedModel):
+    def __init__(self, config, add_mlm_layer=False):
+        super().__init__(config)
+        self.config = config
+
+        self.embedding = Embedding(config)
+        self.parser_network = ParserNetwork(config, pad=config.pad_token_id)
+        self.transformer = Encoder(config, activation_checkpointing=False)
+        self.classifier = (
+            MaskClassifier(config, self.embedding.word_embedding.weight)
+            if add_mlm_layer
+            else None
+        )
+
+    def get_input_embeddings(self):
+        return self.embedding.word_embedding
+
+    def set_input_embeddings(self, value):
+        self.embedding.word_embedding = value
+
+    def get_contextualized_embeddings(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> List[torch.Tensor]:
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            raise ValueError("You have to specify input_ids")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device
+
+        static_embeddings, relative_embedding = self.embedding(input_ids.t())
+        att_mask, cibling, head, block = self.parser_network(
+            input_ids.t(), static_embeddings
+        )
+        contextualized_embeddings, attention_probs = self.transformer(
+            static_embeddings, att_mask, relative_embedding
+        )
+        contextualized_embeddings = [
+            e.transpose(0, 1) for e in contextualized_embeddings
+        ]
+        last_layer = contextualized_embeddings[-1]
+        contextualized_embeddings = [contextualized_embeddings[0]] + [
+            contextualized_embeddings[i] - contextualized_embeddings[i - 1]
+            for i in range(1, len(contextualized_embeddings))
+        ]
+        return last_layer, contextualized_embeddings, attention_probs
+
+    @add_start_docstrings_to_model_forward(
+        LTG_BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], 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
+        )
+
+        (
+            sequence_output,
+            contextualized_embeddings,
+            attention_probs,
+        ) = self.get_contextualized_embeddings(input_ids, attention_mask)
+
+        if not return_dict:
+            return (
+                sequence_output,
+                *([contextualized_embeddings] if output_hidden_states else []),
+                *([attention_probs] if output_attentions else []),
+            )
+
+        return BaseModelOutput(
+            last_hidden_state=sequence_output,
+            hidden_states=contextualized_embeddings if output_hidden_states else None,
+            attentions=attention_probs if output_attentions else None,
+        )
+
+
+@add_start_docstrings(
+    """LTG-BERT model with a `language modeling` head on top.""",
+    LTG_BERT_START_DOCSTRING,
+)
+class LtgBertForMaskedLM(LtgBertModel):
+    _keys_to_ignore_on_load_unexpected = ["head"]
+
+    def __init__(self, config):
+        super().__init__(config, add_mlm_layer=True)
+
+    def get_output_embeddings(self):
+        return self.classifier.nonlinearity[-1].weight
+
+    def set_output_embeddings(self, new_embeddings):
+        self.classifier.nonlinearity[-1].weight = new_embeddings
+
+    @add_start_docstrings_to_model_forward(
+        LTG_BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple[torch.Tensor], 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
+        )
+
+        (
+            sequence_output,
+            contextualized_embeddings,
+            attention_probs,
+        ) = self.get_contextualized_embeddings(input_ids, attention_mask)
+        subword_prediction = self.classifier(sequence_output)
+
+        masked_lm_loss = None
+        if labels is not None:
+            masked_lm_loss = F.cross_entropy(
+                subword_prediction.flatten(0, 1), labels.flatten()
+            )
+
+        if not return_dict:
+            output = (
+                subword_prediction,
+                *([contextualized_embeddings] if output_hidden_states else []),
+                *([attention_probs] if output_attentions else []),
+            )
+            return (
+                ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+            )
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=subword_prediction,
+            hidden_states=contextualized_embeddings if output_hidden_states else None,
+            attentions=attention_probs if output_attentions else None,
+        )
+
+
+class Classifier(nn.Module):
+    def __init__(self, config, num_labels: int):
+        super().__init__()
+
+        drop_out = getattr(config, "classifier_dropout", config.hidden_dropout_prob)
+
+        self.nonlinearity = nn.Sequential(
+            nn.LayerNorm(
+                config.hidden_size, config.layer_norm_eps, elementwise_affine=False
+            ),
+            nn.Linear(config.hidden_size, config.hidden_size),
+            nn.GELU(),
+            nn.LayerNorm(
+                config.hidden_size, config.layer_norm_eps, elementwise_affine=False
+            ),
+            nn.Dropout(drop_out),
+            nn.Linear(config.hidden_size, num_labels),
+        )
+        self.initialize(config.hidden_size)
+
+    def initialize(self, hidden_size):
+        std = math.sqrt(2.0 / (5.0 * hidden_size))
+        nn.init.trunc_normal_(
+            self.nonlinearity[1].weight, mean=0.0, std=std, a=-2 * std, b=2 * std
+        )
+        nn.init.trunc_normal_(
+            self.nonlinearity[-1].weight, mean=0.0, std=std, a=-2 * std, b=2 * std
+        )
+        self.nonlinearity[1].bias.data.zero_()
+        self.nonlinearity[-1].bias.data.zero_()
+
+    def forward(self, x):
+        x = self.nonlinearity(x)
+        return x
+
+
+@add_start_docstrings(
+    """
+    LTG-BERT model with a sequence classification/regression head on top (a linear layer on top of the pooled
+    output) e.g. for GLUE tasks.
+    """,
+    LTG_BERT_START_DOCSTRING,
+)
+class LtgBertForSequenceClassification(LtgBertModel):
+    _keys_to_ignore_on_load_unexpected = ["classifier"]
+    _keys_to_ignore_on_load_missing = ["head"]
+
+    def __init__(self, config):
+        super().__init__(config, add_mlm_layer=False)
+
+        self.num_labels = config.num_labels
+        self.head = Classifier(config, self.num_labels)
+
+    @add_start_docstrings_to_model_forward(
+        LTG_BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple[torch.Tensor], 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
+        )
+
+        (
+            sequence_output,
+            contextualized_embeddings,
+            attention_probs,
+        ) = self.get_contextualized_embeddings(input_ids, attention_mask)
+        logits = self.head(sequence_output[:, 0, :])
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (
+                    labels.dtype == torch.long or labels.dtype == torch.int
+                ):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = nn.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 = nn.CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = nn.BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (
+                logits,
+                *([contextualized_embeddings] if output_hidden_states else []),
+                *([attention_probs] if output_attentions else []),
+            )
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=contextualized_embeddings if output_hidden_states else None,
+            attentions=attention_probs if output_attentions else None,
+        )
+
+
+@add_start_docstrings(
+    """
+    LTG-BERT model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    LTG_BERT_START_DOCSTRING,
+)
+class LtgBertForTokenClassification(LtgBertModel):
+    _keys_to_ignore_on_load_unexpected = ["classifier"]
+    _keys_to_ignore_on_load_missing = ["head"]
+
+    def __init__(self, config):
+        super().__init__(config, add_mlm_layer=False)
+
+        self.num_labels = config.num_labels
+        self.head = Classifier(config, self.num_labels)
+
+    @add_start_docstrings_to_model_forward(
+        LTG_BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        (
+            sequence_output,
+            contextualized_embeddings,
+            attention_probs,
+        ) = self.get_contextualized_embeddings(input_ids, attention_mask)
+        logits = self.head(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = nn.CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (
+                logits,
+                *([contextualized_embeddings] if output_hidden_states else []),
+                *([attention_probs] if output_attentions else []),
+            )
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=contextualized_embeddings if output_hidden_states else None,
+            attentions=attention_probs if output_attentions else None,
+        )
+
+
+@add_start_docstrings(
+    """
+    LTG-BERT model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
+    layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    LTG_BERT_START_DOCSTRING,
+)
+class LtgBertForQuestionAnswering(LtgBertModel):
+    _keys_to_ignore_on_load_unexpected = ["classifier"]
+    _keys_to_ignore_on_load_missing = ["head"]
+
+    def __init__(self, config):
+        super().__init__(config, add_mlm_layer=False)
+
+        self.num_labels = config.num_labels
+        self.head = Classifier(config, self.num_labels)
+
+    @add_start_docstrings_to_model_forward(
+        LTG_BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        start_positions: Optional[torch.Tensor] = None,
+        end_positions: Optional[torch.Tensor] = None,
+    ) -> Union[Tuple[torch.Tensor], QuestionAnsweringModelOutput]:
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        (
+            sequence_output,
+            contextualized_embeddings,
+            attention_probs,
+        ) = self.get_contextualized_embeddings(input_ids, attention_mask)
+        logits = self.head(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 = nn.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,
+                *([contextualized_embeddings] if output_hidden_states else []),
+                *([attention_probs] if output_attentions else []),
+            )
+            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=contextualized_embeddings if output_hidden_states else None,
+            attentions=attention_probs if output_attentions else None,
+        )
+
+
+@add_start_docstrings(
+    """
+    LTG-BERT model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
+    softmax) e.g. for RocStories/SWAG tasks.
+    """,
+    LTG_BERT_START_DOCSTRING,
+)
+class LtgBertForMultipleChoice(LtgBertModel):
+    _keys_to_ignore_on_load_unexpected = ["classifier"]
+    _keys_to_ignore_on_load_missing = ["head"]
+
+    def __init__(self, config):
+        super().__init__(config, add_mlm_layer=False)
+
+        self.num_labels = getattr(config, "num_labels", 2)
+        self.head = Classifier(config, self.num_labels)
+
+    @add_start_docstrings_to_model_forward(
+        LTG_BERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], MultipleChoiceModelOutput]:
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+        num_choices = input_ids.shape[1]
+
+        flat_input_ids = input_ids.view(-1, input_ids.size(-1))
+        flat_attention_mask = (
+            attention_mask.view(-1, attention_mask.size(-1))
+            if attention_mask is not None
+            else None
+        )
+
+        (
+            sequence_output,
+            contextualized_embeddings,
+            attention_probs,
+        ) = self.get_contextualized_embeddings(flat_input_ids, flat_attention_mask)
+        logits = self.head(sequence_output)
+        reshaped_logits = logits.view(-1, num_choices)
+
+        loss = None
+        if labels is not None:
+            loss_fct = nn.CrossEntropyLoss()
+            loss = loss_fct(reshaped_logits, labels)
+
+        if not return_dict:
+            output = (
+                reshaped_logits,
+                *([contextualized_embeddings] if output_hidden_states else []),
+                *([attention_probs] if output_attentions else []),
+            )
+            return ((loss,) + output) if loss is not None else output
+
+        return MultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=contextualized_embeddings if output_hidden_states else None,
+            attentions=attention_probs if output_attentions else None,
+        )

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b7628dc36c1c996847b8ac8e32cd959e09cfab55511e6dd935b904f2374f0b0
+size 159209798

results.md

@@ -0,0 +1,125 @@
+# Results 
+
+The results here are taken from running `score_predictions.py` from the [babylm evaluation pipeline](https://github.com/babylm/evaluation-pipeline-2024) on the `ELC_ParserBERT_10M_textonly_predictions.json.gz` file in this directory, which contains the predictions for the different evaluation tasks.
+
+## Overall Results
+
+Here are the average results per section and the macroscore, compared with the baseline models:
+
+| Model | BLiMP | BLiMP Supplement | EWoK | GLUE | *Macroaverage* |
+| --- | --- | --- | --- | --- | --- |
+| BabyLlama | 69.8 | 59.5 | 50.7 | 63.3 | 60.8 |
+| LTG-BERT | 60.6 | 60.8 | 48.9 | 60.3 | 57.7 |
+| ELC-ParserBERT | 59.6 | 57.7 | 63.1 | 44.5 | 56.2 |
+
+## The Breakdown Per Section
+
+|glue subtask   |    Score | 
+|-------------- |  ------- |
+|cola (MCC)     |    0.042 |
+|sst2           |    0.502 |
+|mrpc (F1)      |    0.82  |
+|qqp (F1)       |    0     |
+|mnli           |    0.357 |
+|mnli-mm        |    0.355 |
+|qnli           |    0.491 |
+|rte            |    0.496 |
+|boolq          |    0.585 |
+|multirc        |    0.63  |
+|wsc            |    0.615 |
+|*Average*      |    0.445 |
+
+| blimp subtask                                       |   Score  |
+| --------------------------------------------------- | -------  |
+| adjunct_island                                      |   0.712  |
+| anaphor_gender_agreement                            |   0.593  |
+| anaphor_number_agreement                            |   0.647  |
+| animate_subject_passive                             |   0.594  |
+| animate_subject_trans                               |   0.47   |
+| causative                                           |   0.726  |
+| complex_NP_island                                   |   0.447  |
+| coordinate_structure_constraint_complex_left_branch |   0.39   |
+| coordinate_structure_constraint_object_extraction   |   0.806  |
+| determiner_noun_agreement_1                         |   0.793  |
+| determiner_noun_agreement_2                         |   0.936  |
+| determiner_noun_agreement_irregular_1               |   0.467  |
+| determiner_noun_agreement_irregular_2               |   0.394  |
+| determiner_noun_agreement_with_adj_2                |   0.889  |
+| determiner_noun_agreement_with_adj_irregular_1      |   0.834  |
+| determiner_noun_agreement_with_adj_irregular_2      |   0.848  |
+| determiner_noun_agreement_with_adjective_1          |   0.758  |
+| distractor_agreement_relational_noun                |   0.212  |
+| distractor_agreement_relative_clause                |   0.282  |
+| drop_argument                                       |   0.485  |
+| ellipsis_n_bar_1                                    |   0.505  |
+| ellipsis_n_bar_2                                    |   0.342  |
+| existential_there_object_raising                    |   0.447  |
+| existential_there_quantifiers_1                     |   0.385  |
+| existential_there_quantifiers_2                     |   0.396  |
+| existential_there_subject_raising                   |   0.476  |
+| expletive_it_object_raising                         |   0.44   |
+| inchoative                                          |   0.527  |
+| intransitive                                        |   0.484  |
+| irregular_past_participle_adjectives                |   0.348  |
+| irregular_past_participle_verbs                     |   0.594  |
+| irregular_plural_subject_verb_agreement_1           |   0.634  |
+| irregular_plural_subject_verb_agreement_2           |   0.687  |
+| left_branch_island_echo_question                    |   0.634  |
+| left_branch_island_simple_question                  |   0.615  |
+| matrix_question_npi_licensor_present                |   0.206  |
+| npi_present_1                                       |   0.362  |
+| npi_present_2                                       |   0.347  |
+| only_npi_licensor_present                           |   0.964  |
+| only_npi_scope                                      |   0.89   |
+| passive_1                                           |   0.514  |
+| passive_2                                           |   0.482  |
+| principle_A_c_command                               |   0.635  |
+| principle_A_case_1                                  |   0.999  |
+| principle_A_case_2                                  |   0.78   |
+| principle_A_domain_1                                |   0.893  |
+| principle_A_domain_2                                |   0.623  |
+| principle_A_domain_3                                |   0.556  |
+| principle_A_reconstruction                          |   0.339  |
+| regular_plural_subject_verb_agreement_1             |   0.628  |
+| regular_plural_subject_verb_agreement_2             |   0.663  |
+| sentential_negation_npi_licensor_present            |   0.93   |
+| sentential_negation_npi_scope                       |   0.722  |
+| sentential_subject_island                           |   0.361  |
+| superlative_quantifiers_1                           |   0.702  |
+| superlative_quantifiers_2                           |   0.498  |
+| tough_vs_raising_1                                  |   0.351  |
+| tough_vs_raising_2                                  |   0.648  |
+| transitive                                          |   0.645  |
+| wh_island                                           |   0.719  |
+| wh_questions_object_gap                             |   0.657  |
+| wh_questions_subject_gap                            |   0.861  |
+| wh_questions_subject_gap_long_distance              |   0.937  |
+| wh_vs_that_no_gap                                   |   0.969  |
+| wh_vs_that_no_gap_long_distance                     |   0.969  |
+| wh_vs_that_with_gap                                 |   0.222  |
+| wh_vs_that_with_gap_long_distance                   |   0.063  |
+| *Average*                                           |   0.596  |
+
+| blimp_supplement subtask   |    Score |
+| -------------------------- |  ------- |
+| hypernym                   |    0.531 |
+| qa_congruence_easy         |    0.641 |
+| qa_congruence_tricky       |    0.521 |
+| subject_aux_inversion      |    0.614 |
+| turn_taking                |    0.579 |
+| *Average*                  |    0.577 |
+
+| ewok subtask            |   Score |
+| ----------------------- | ------- |
+| agent-properties        |   0.738 |
+| material-dynamics       |   0.81  |
+| material-properties     |   0.6   |
+| physical-dynamics       |   0.383 |
+| physical-interactions   |   0.599 |
+| physical-relations      |   0.817 |
+| quantitative-properties |   0.427 |
+| social-interactions     |   0.565 |
+| social-properties       |   0.561 |
+| social-relations        |   0.807 |
+| spatial-relations       |   0.635 |
+| *Average*               |   0.631 |

special_tokens_map.json

@@ -0,0 +1,7 @@
+{
+  "cls_token": "[CLS]",
+  "mask_token": "[MASK]",
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "unk_token": "[UNK]"
+}

tokenizer_config.json

@@ -0,0 +1,4 @@
+{
+  "model_max_length": 1000000000000000019884624838656,
+  "tokenizer_class": "PreTrainedTokenizerFast"
+}