Using the dynaword package rather than custom functions.

data/danske-taler/create.py

@@ -4,7 +4,10 @@
 #     "beautifulsoup4==4.13.3",
 #     "datasets>=3.0.0",
 #     "transformers",
+#     "dynaword"
 # ]
+# [tool.uv.sources]
+# dynaword = { git = "https://huggingface.co/datasets/danish-foundation-models/danish-dynaword", rev = "00e7f2aee7f7ad2da423419f77ecbb9c0536de0d" }
 # ///
 """
 Danske Taler API Downloader
@@ -34,15 +37,20 @@ import time
 from datetime import date
 from pathlib import Path
 from typing import Any
-from functools import partial
 
 from datasets import Dataset
-from transformers import AutoTokenizer
 import pandas as pd
 import requests
 from bs4 import BeautifulSoup, NavigableString
 from tqdm import tqdm
 
+from dynaword.process_dataset import (
+    add_token_count,
+    ensure_column_order,
+    remove_duplicate_text,
+    remove_empty_texts,
+)
+
 logger = logging.getLogger(__name__)
 
 # Configuration
@@ -101,32 +109,6 @@ def contains_html_tags(text):
     return any(tag.name in KNOWN_HTML_TAGS for tag in soup.find_all())
 
 
-def _tokenize_function(
-    examples: dict[str, Any], tokenizer: AutoTokenizer
-) -> dict[str, Any]:
-    token_count = [
-        len(tokens)
-        for tokens in tokenizer(examples["text"], padding=False)[  # type: ignore
-            "input_ids"
-        ]
-    ]
-    examples["token_count"] = token_count
-    return examples
-
-
-def add_token_count(
-    ds: Dataset,
-    tokenizer_name: str = "AI-Sweden-Models/Llama-3-8B-instruct",
-    num_proc: int = 4,
-) -> Dataset:
-    tokenizer = AutoTokenizer.from_pretrained(tokenizer_name, use_fast=True)
-
-    tokenize = partial(_tokenize_function, tokenizer=tokenizer)  # type: ignore
-
-    ds = ds.map(tokenize, batched=True, num_proc=num_proc)
-    return ds
-
-
 def get_all_speeches() -> list[dict[str, Any]]:
     # fetch first page, notably the total number of pages
     url = f"{API_BASE_URL}/speeches?per_page=50"
@@ -294,24 +276,13 @@ def main():
     df = df[df["license"] == "cc0"]
     logger.info(f"Removed {len_df - len(df)} documents without a cc0 license")
 
-    # remove duplicate ids
-    len_df = len(df)
-    df = df.drop_duplicates(subset=["id"])
-    logger.info(f"Removed {len_df - len(df)} duplicate ids")
+    dataset = Dataset.from_pandas(df, preserve_index=False)
 
-    # remove rows with empty text
-    len_df = len(df)
-    df = df[df["text"].str.strip() != ""]
-    logger.info(f"Removed {len_df - len(df)} rows with empty text")
-
-    # remove rows with duplicate text
-    len_df = len(df)
-    df = df.drop_duplicates(subset=["text"])
-    logger.info(f"Removed {len_df - len(df)} rows with duplicate text")
+    dataset = remove_empty_texts(dataset)  # remove rows with empty text
+    dataset = remove_duplicate_text(dataset)  # remove rows with duplicate text
+    dataset = add_token_count(dataset)
+    dataset = ensure_column_order(dataset)
 
-    dataset = Dataset.from_pandas(
-        df[["id", "text", "source", "added", "created"]], preserve_index=False
-    )
     assert len(set(dataset["id"])) == len(dataset), "IDs are not unique"
     assert len(set(dataset["text"])) == len(dataset), "Texts are not unique"
     assert len(set(df["license"])) == 1, "Multiple licenses found"
@@ -319,8 +290,6 @@ def main():
     # check for html tags in text
     assert not df["text"].apply(contains_html_tags).any(), "HTML tags found in text"
 
-    dataset = add_token_count(dataset)
-
     dataset.to_parquet(save_path)