initial commit

README.md

@@ -0,0 +1,29 @@
+---
+language: dv
+---
+
+# dv-labse
+
+This is an experiment in cross-lingual transfer learning, to insert Dhivehi word and
+word-piece tokens into Google's LaBSE model.
+
+- Original model weights: https://huggingface.co/setu4993/LaBSE
+- Original model announcement: https://ai.googleblog.com/2020/08/language-agnostic-bert-sentence.html
+
+This currently outperforms dv-wave and dv-MuRIL (a similar transfer learning model) on 
+the Maldivian News Classification task https://github.com/Sofwath/DhivehiDatasets
+
+- mBERT: 52%
+- dv-wave (ELECTRA): 89%
+- dv-muril: 90.7%
+- dv-labse: 91.5% (may continue training)
+
+## Training
+
+- Start with LaBSE (similar to mBERT) with no Thaana vocabulary
+- Based on PanLex dictionaries, attach 1,100 Dhivehi words to Sinhalese or English embeddings
+- Add remaining words and word-pieces from dv-wave's vocabulary to vocab.txt
+- Continue BERT pretraining on Dhivehi text
+
+CoLab notebook: 
+https://colab.research.google.com/drive/1CUn44M2fb4Qbat2pAvjYqsPvWLt1Novi

config.json

@@ -0,0 +1,23 @@
+{
+  "architectures": [
+    "BertForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "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": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "position_embedding_type": "absolute",
+  "transformers_version": "4.3.0",
+  "type_vocab_size": 2,
+  "use_cache": true,
+  "vocab_size": 528954
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e83b53c49ec3cc36f160bd0cef7dcb5cd96f153286358b7e69430341105dfa6
+size 1971321682

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:569bc9c6c7fcdaf89d336c1f63940b2c559b12d971b7055c8f4ce06ed71bce40
+size 2159