initial commit

README.md

@@ -0,0 +1,34 @@
+# SciBERT Longformer finetuned to SDG classification
+
+This is a Lonformer version of the [SciBERT uncased](https://huggingface.co/allenai/scibert_scivocab_uncased) model by Allen AI, finetuned to Sustainable Development Goals classification. The model is slower than SciBERT (~2.5x in my benchmarks) but can allow for 8x wider `max_seq_length` (4096 vs. 512) which is handy in case of working with long texts, e.g. scientific full texts. 
+
+The conversion to Longformer was performed with a [tutorial](https://github.com/allenai/longformer/blob/master/scripts/convert_model_to_long.ipynb) by Allen AI: see a [Google Colab Notebook](https://colab.research.google.com/drive/1NPTnMkeAYOF2MWH3_uJYesuxxdOzxrFn?usp=sharing) by [Yury](https://yorko.github.io/) which closely follows the tutorial. 
+
+Note: 
+
+- no additional MLM pretraining of the Longformer was performed, the [collab notebook](https://colab.research.google.com/drive/1NPTnMkeAYOF2MWH3_uJYesuxxdOzxrFn?usp=sharing) stops at step 3, and step 4 is not done. The model can be improved with this additional MLM pretraining, better to do so with scientific texts, e.g. [S@ORC](https://github.com/allenai/s2orc), again by Allen AI. 
+- no extensive benchmarks SciBERT Longformer vs. SciBERT were performed in terms of downstream task performance
+
+Links:
+ - the original [SciBERT repo](https://github.com/allenai/scibert)
+ - the original [Longformer repo](https://github.com/allenai/longformer)
+
+
+If using these models, please consider citing the following papers:
+```
+@inproceedings{beltagy-etal-2019-scibert,
+    title = "SciBERT: A Pretrained Language Model for Scientific Text",
+    author = "Beltagy, Iz  and Lo, Kyle  and Cohan, Arman",
+    booktitle = "EMNLP",
+    year = "2019",
+    publisher = "Association for Computational Linguistics",
+    url = "https://www.aclweb.org/anthology/D19-1371"
+}
+
+@article{Beltagy2020Longformer,
+  title={Longformer: The Long-Document Transformer},
+  author={Iz Beltagy and Matthew E. Peters and Arman Cohan},
+  journal={arXiv:2004.05150},
+  year={2020},
+}
+```

config.json

@@ -0,0 +1,34 @@
+{
+  "architectures": [
+    "BertForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "attention_window": [
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512
+  ],
+  "gradient_checkpointing": false,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 4096,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "type_vocab_size": 2,
+  "vocab_size": 31090
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48db774448d6c458effaee86de8b6d656b6571f4de8df2f148542a6c7db8b7c7
+size 450822016

special_tokens_map.json

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

tokenizer_config.json

@@ -0,0 +1 @@
+{"model_max_length": 4096, "special_tokens_map_file": null, "full_tokenizer_file": null}