diff --git a/.gitignore b/.gitignore
index 2b51c6e..b87dd4a 100644
--- a/.gitignore
+++ b/.gitignore
@@ -109,4 +109,6 @@ climateguard/experiments/claim_extraction/climateguard_claim_extraction
 .aws/
 germany_s3_bucket/
 france_s3_bucket/
-.gcloud
\ No newline at end of file
+.gcloud
+.claude
+CLAUDE.md
\ No newline at end of file
diff --git a/data/update_huggingface_dataset.py b/data/update_huggingface_dataset.py
new file mode 100644
index 0000000..a899322
--- /dev/null
+++ b/data/update_huggingface_dataset.py
@@ -0,0 +1,137 @@
+import os
+from datetime import date, datetime, timedelta
+from typing import List
+
+import numpy as np
+import pandas as pd
+import psycopg
+import requests
+from datasets import Dataset, DatasetDict, load_dataset
+from dotenv import load_dotenv
+
+load_dotenv(os.path.join(os.path.dirname(__file__), ".env"))
+
+
+def fetch_articles(columns: List[str] = None) -> pd.DataFrame:
+    conninfo = (
+        f"host={os.getenv('PG_HOST', 'localhost')} port={os.getenv('PG_PORT', 5432)} "
+        f"dbname={os.getenv('PG_DATABASE', 'postgres')} user={os.getenv('PG_USER', 'user')} "
+        f"password={os.getenv('PG_PASSWORD', 'supersecret')}"
+    )
+    columns_str = ", ".join(columns) if columns else "*"
+    query = f"""
+        SELECT {columns_str}
+        FROM analytics.task_global_completion
+        where task_completion_aggregate_id is not null
+    """
+    with psycopg.connect(conninfo) as conn:
+        with conn.cursor() as cur:
+            cur.execute(query)
+            rows = cur.fetchall()
+            columns = [desc.name for desc in cur.description]
+
+    return pd.DataFrame(rows, columns=columns)
+
+
+def get_week_number(record):
+    format_str = "%Y-%m-%dT%H:%M:%S"
+    if len(record["data_item_start"]) > 19:
+        format_str = format_str + "%z"
+    week_number = (
+        datetime.strptime(record["data_item_start"], format_str).isocalendar().week
+    )
+    return week_number
+
+
+def split_df_by_week(df):
+    df["week_number"] = df.data_item_start.dt.isocalendar().week
+    df_train = df.loc[df.week_number % 5 > 0]
+    df_test = df.loc[df.week_number % 5 == 0]
+    return df_train, df_test
+
+def cast_to_int_with_nan(col: pd.Series, int_type: str='uint8'):
+    col = col.fillna(-1)
+    col = col.astype(int_type)
+    col[col==-1] = np.nan
+    return col
+
+
+def format_dtypes(df: pd.DataFrame):
+    df.data_item_day = df.data_item_day.astype('f2')
+    df.data_item_month = df.data_item_month.astype('f2')
+    df.data_item_year = df.data_item_year.astype('f2')
+    df.data_item_model_result = df.data_item_model_result.astype('f2')
+    df.mesinfo_correct = df.mesinfo_correct.astype('f2')
+    df.mesinfo_incorrect = df.mesinfo_incorrect.astype('f2')
+    df.speaker_journalist = df.speaker_journalist.astype('f2')
+    df.speaker_commentator = df.speaker_commentator.astype('f2')
+    df.speaker_guest = df.speaker_guest.astype('f2')
+    df.speaker_politician = df.speaker_politician.astype('f2')
+    df.speaker_audience = df.speaker_audience.astype('f2')
+    df.speaker_unknown = df.speaker_unknown.astype('f2')
+    df.mesinfo_corrected_bool = df.mesinfo_corrected_bool.astype('f2')
+    df.mesinfo_corrected = df.mesinfo_corrected.fillna("")
+    return df
+
+def generate_hf_dataset(df_train, df_test):
+    df_train.data_item_start = df_train.data_item_start.dt.strftime(
+        "%Y-%m-%dT%H:%M:%S%z"
+    )
+    df_test.data_item_start = df_test.data_item_start.dt.strftime("%Y-%m-%dT%H:%M:%S%z")
+    dataset = DatasetDict(
+        {
+            "train": Dataset.from_pandas(df_train),
+            "test": Dataset.from_pandas(df_test),
+        }
+    )
+    return dataset
+
+
+DATASET_COLUMNS = [
+    "task_completion_aggregate_id",
+    "task_aggregate_id",
+    "created_at",
+    "updated_at",
+    "is_labeled",
+    "project_id",
+    "country",
+    "data_item_id",
+    "data_item_channel",
+    "data_item_channel_name",
+    "data_item_channel_title",
+    "data_item_channel_program",
+    "data_item_channel_program_type",
+    "data_item_day",
+    "data_item_month",
+    "data_item_year",
+    "data_item_start",
+    "data_item_model_name",
+    "data_item_model_reason",
+    "data_item_model_result",
+    "data_item_plaintext",
+    "data_item_plaintext_whisper",
+    "data_item_url_mediatree",
+    "mesinfo_choice",
+    "locuteur_choice",
+    "mesinfo_correct",
+    "mesinfo_incorrect",
+    "speaker_journalist",
+    "speaker_commentator",
+    "speaker_guest",
+    "speaker_politician",
+    "speaker_audience",
+    "speaker_unknown",
+    "mesinfo_corrected",
+    "mesinfo_corrected_bool",
+    "debunk_references",
+    "claims",
+    "explanations",
+    "other_comments",
+]
+
+if __name__ == "__main__":
+    df = fetch_articles(DATASET_COLUMNS)
+    df = format_dtypes(df)
+    df_train, df_test = split_df_by_week(df)
+    dataset = generate_hf_dataset(df_train, df_test)
+    dataset.push_to_hub("DataForGood/climateguard-training", private=True, token=True)