wikipediaを間引く

Author: yuiseki

src/train.py

@@ -113,9 +113,9 @@ def prepare_train_data(dataset_id):
     if "dataset_load_config" in train_config:
         dataset_load_config = train_config["dataset_load_config"]
         data = load_dataset(dataset_id, dataset_load_config, split="train", num_proc=32)
-        if dataset_load_config == "20231101.ja" or dataset_load_config == "20231101.vi" or dataset_load_config == "20231101.es":
+        if dataset_load_config == "20231101.ja" or dataset_load_config == "20231101.vi" or dataset_load_config == "20231101.es" or dataset_load_config == "20231101.it":
             data = data.filter(lambda item, idx: idx % 3 == 0, with_indices=True)
-        if dataset_load_config == "20231101.de":
+        if dataset_load_config == "20231101.de" or dataset_load_config == "20231101.fr":
             data = data.filter(lambda item, idx: idx % 5 == 0, with_indices=True)
     else:
         data = load_dataset(dataset_id, split="train", num_proc=32)
@@ -162,7 +162,8 @@ def prepare_train_data(dataset_id):
                 lambda x: simple_template_for_train(x[input_field_name], x[output_field_name]),
                 axis=1,
             )
-
+    # keep only text field
+    data = data_df[["text"]]
     data = Dataset.from_pandas(data_df)
     data = data.train_test_split(seed=42, test_size=0.2)
     print(len(data["train"]))
@@ -281,7 +282,7 @@ def load_model_and_tokenizer(model_id):
     args=training_arguments,
     tokenizer=tokenizer,
     packing=False,
-    max_seq_length=512,
+    max_seq_length=1024,
 )
 
 #