Upload preprocessing/create_dataset.py with huggingface_hub

preprocessing/create_dataset.py

@@ -0,0 +1,76 @@
+import argparse
+import os
+
+from datasets import DatasetDict, concatenate_datasets, load_from_disk
+
+
+SOURCE_SPECS = {
+    "qqp": {"sentence1": "question1", "sentence2": "question2", "keep": lambda label: label == 1},
+    "mrpc": {"sentence1": "sentence1", "sentence2": "sentence2", "keep": lambda label: label == 1},
+    "stsb": {"sentence1": "sentence1", "sentence2": "sentence2", "keep": lambda label: label >= 3},
+    "rte": {"sentence1": "sentence1", "sentence2": "sentence2", "keep": lambda label: label == 1},
+    "anli": {"sentence1": "sentence1", "sentence2": "sentence2", "keep": lambda label: label == 0},
+    "smf": {"sentence1": "sentence1", "sentence2": "sentence2", "keep": lambda label: label == 0},
+}
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--data_dir", default="./data")
+    parser.add_argument("--sources", default="qqp,mrpc,stsb")
+    parser.add_argument("--output", default="ling_conversion")
+    return parser.parse_args()
+
+
+def rename_dev_split(data):
+    if "dev" not in data and "validation" in data:
+        data["dev"] = data["validation"]
+        del data["validation"]
+    return data
+
+
+def normalize_columns(data, source_name, sentence1_name, sentence2_name):
+    for split in data.keys():
+        split_data = data[split]
+        rename_map = {}
+        if sentence1_name in split_data.column_names and sentence1_name != "sentence1":
+            rename_map[sentence1_name] = "sentence1"
+        if sentence2_name in split_data.column_names and sentence2_name != "sentence2":
+            rename_map[sentence2_name] = "sentence2"
+        if rename_map:
+            split_data = split_data.rename_columns(rename_map)
+        if "source" not in split_data.column_names:
+            split_data = split_data.add_column("source", [source_name] * len(split_data))
+        keep_columns = [column for column in split_data.column_names if column.startswith("sentence") or column == "source"]
+        data[split] = split_data.remove_columns(sorted(set(split_data.column_names) - set(keep_columns)))
+    return data
+
+
+def main():
+    args = parse_args()
+    sources = [source.strip() for source in args.sources.split(",") if source.strip()]
+
+    datasets = {}
+    for source_name in sources:
+        if source_name not in SOURCE_SPECS:
+            raise ValueError(f"Unsupported source dataset: {source_name}")
+
+        spec = SOURCE_SPECS[source_name]
+        dataset = load_from_disk(os.path.join(args.data_dir, source_name))
+        dataset = rename_dev_split(dataset)
+        dataset = dataset.filter(lambda row: spec["keep"](row["label"]))
+        dataset = normalize_columns(dataset, source_name, spec["sentence1"], spec["sentence2"])
+        datasets[source_name] = dataset
+
+    merged = DatasetDict(
+        {
+            "train": concatenate_datasets([dataset["train"] for dataset in datasets.values() if "train" in dataset]),
+            "dev": concatenate_datasets([dataset["dev"] for dataset in datasets.values() if "dev" in dataset]),
+            "test": concatenate_datasets([dataset["test"] for dataset in datasets.values() if "test" in dataset]),
+        }
+    )
+    merged.save_to_disk(os.path.join(args.data_dir, args.output))
+
+
+if __name__ == "__main__":
+    main()