Merge branch 'efficient-transfer' of https://github.com/bminixhofer/w…

…tpsplit into efficient-transfer

configs/xlmr_stratify_0.1_3layers_p_v3.json

@@ -0,0 +1,45 @@
+{
+    "model_name_or_path": "xlm-roberta-base",
+    "output_dir": "xlmr-3l-v3",
+    "train_text_path": "data/sentence/train.parquet",
+    "valid_text_path": "data/sentence/valid.parquet",
+    "block_size": 256,
+    "eval_stride": 128,
+    "use_bert": true,
+    "do_train": true,
+    "do_eval": true,
+    "evaluation_strategy": "steps",
+    "per_device_train_batch_size": 128,
+    "per_device_eval_batch_size": 64,
+    "gradient_accumulation_steps": 1,
+    "eval_accumulation_steps": 8,
+    "dataloader_num_workers": 4,
+    "preprocessing_num_workers": 32,
+    "learning_rate": 1e-4,
+    "save_strategy": "steps",
+    "fp16": false,
+    "max_steps": 200000,
+    "save_steps": 50000,
+    "eval_steps": 5000,
+    "logging_steps": 50,
+    "report_to": "wandb",
+    "is_decoder": false,
+    "remove_unused_columns": false,
+    "lookahead": null,
+    "one_sample_per_line": false,
+    "do_sentence_training": true,
+    "do_auxiliary_training": true,
+    "auxiliary_remove_prob": 0.2,
+    "warmup_steps": 5000,
+    "adapter_warmup_steps": 0,
+    "adapter_lr_multiplier": 1,
+    "ngram_order": 1,
+    "non_punctuation_sample_ratio": 0.1,
+    "prediction_loss_only": true,
+    "use_auxiliary": true,
+    "ddp_timeout": 3600,
+    "use_subwords": true,
+    "num_hidden_layers": 3,
+    "custom_punctuation_file": "punctuation_xlmr_unk.txt",
+    "log_level": "warning"
+}

configs/xlmr_stratify_0.1_3layers_p_v3_bs512.json

@@ -0,0 +1,45 @@
+{
+    "model_name_or_path": "xlm-roberta-base",
+    "output_dir": "xlmr-3l-v3",
+    "train_text_path": "data/sentence/train.parquet",
+    "valid_text_path": "data/sentence/valid.parquet",
+    "block_size": 512,
+    "eval_stride": 256,
+    "use_bert": true,
+    "do_train": true,
+    "do_eval": true,
+    "evaluation_strategy": "steps",
+    "per_device_train_batch_size": 64,
+    "per_device_eval_batch_size": 32,
+    "gradient_accumulation_steps": 1,
+    "eval_accumulation_steps": 8,
+    "dataloader_num_workers": 4,
+    "preprocessing_num_workers": 32,
+    "learning_rate": 1e-4,
+    "save_strategy": "steps",
+    "fp16": false,
+    "max_steps": 200000,
+    "save_steps": 50000,
+    "eval_steps": 5000,
+    "logging_steps": 50,
+    "report_to": "wandb",
+    "is_decoder": false,
+    "remove_unused_columns": false,
+    "lookahead": null,
+    "one_sample_per_line": false,
+    "do_sentence_training": true,
+    "do_auxiliary_training": true,
+    "auxiliary_remove_prob": 0.2,
+    "warmup_steps": 5000,
+    "adapter_warmup_steps": 0,
+    "adapter_lr_multiplier": 1,
+    "ngram_order": 1,
+    "non_punctuation_sample_ratio": 0.1,
+    "prediction_loss_only": true,
+    "use_auxiliary": true,
+    "ddp_timeout": 3600,
+    "use_subwords": true,
+    "num_hidden_layers": 3,
+    "custom_punctuation_file": "punctuation_xlmr_unk.txt",
+    "log_level": "info"
+}