diff --git a/scripts/dump-records.py b/scripts/dump-records.py
new file mode 100644
index 00000000..1ed7f7c9
--- /dev/null
+++ b/scripts/dump-records.py
@@ -0,0 +1,196 @@
+"""Dump records training data for the spam classifier."""
+
+import csv
+import gzip
+import io
+import re
+import string
+from datetime import datetime
+from urllib.parse import urlparse
+
+from bs4 import BeautifulSoup
+from flask import current_app
+from invenio_app.factory import create_api
+from invenio_search.proxies import current_search_client as client
+from opensearch_dsl import Search
+from xrootdpyfs import XRootDPyFS
+
+emoji_pattern = re.compile(
+    "["
+    "\U0001F600-\U0001F64F"  # Emoticons
+    "\U0001F300-\U0001F5FF"  # Symbols & Pictographs
+    "\U0001F680-\U0001F6FF"  # Transport & Map Symbols
+    "\U0001F1E0-\U0001F1FF"  # Flags (iOS)
+    "\U00002700-\U000027BF"  # Dingbats
+    "\U000024C2-\U0001F251"  # Enclosed characters
+    "]+", flags=re.UNICODE)
+
+
+def extract_emojis(text):
+    """Extract emojis from the text."""
+    # Emoji regex pattern
+    return emoji_pattern.findall(text)
+
+
+def links_stats(soup):
+    """Extra number of links and their hosts."""
+    # Parse description and look for links
+    count = 0
+    hosts = set()
+    for link in soup.find_all('a'):
+        count += 1
+        hosts.add(urlparse(link.get('href')).hostname)
+    return count, hosts
+
+
+def make_fs(klass):
+    baseurl = f"root://eosuser.cern.ch//eos/project/z/zenodo/Dumps/{klass}/"
+    return XRootDPyFS(baseurl)
+
+
+def truncate(text, max_length):
+    # strip punctuation as well
+    parts = text.strip().split(" ")
+    if len(parts) >= max_length:
+        parts = parts[:max_length]
+    return ' '.join(parts)
+
+
+def prepare_record(r):
+    """Prepare a record"""
+    # Label
+    if r.parent.is_verified and r.is_published:
+        label = 1  # ham
+    elif r.is_deleted and r.tombstone.removal_reason.id == 'spam':
+        label = 0  # spam
+    else:
+        return None
+
+    # Record ids
+    id = r.id
+
+    # User info
+    uid = r.parent.access.owned_by.user
+    u = users_db.get(uid, None)
+    if u is None:
+        return None
+    ustatus, udomain, udomainstatus = u
+
+    # Text
+    title = getattr(r.metadata, "title", "")
+    soup = BeautifulSoup(getattr(r.metadata, "description", ""), 'html.parser')
+    description = soup.get_text()
+    text = truncate(title + description, 4000)
+
+    # Exclude records with empty text
+    if text.strip() == "":
+        return None
+
+    # Computed properties
+    emojis = len(extract_emojis(title)) + len(extract_emojis(description))
+    title_len = len(title)
+    description_len = len(description)
+    links, links_domains = links_stats(soup)
+
+    # Files
+    files = getattr(r.files, "count", 0)
+    bytes = getattr(r.files, "totalbytes", 0)
+    types = getattr(r.files, "types", [])
+
+    # Todo detect language
+
+    return [label, id, r.created, udomain, title_len, description_len, links, emojis, files, bytes, text, links_domains]
+
+
+########################################################
+
+app = create_api()
+with app.app_context():
+    idxprefix = current_app.config['SEARCH_INDEX_PREFIX']
+
+    # Preload the users database
+    t0 = datetime.now()
+    print("Loading users database ...")
+    users_db = {}
+    s = Search(using=client, index=f'{idxprefix}users-user-v3.0.0').extra(track_total_hits=True)
+    i = 0
+    for u in s.scan():
+        users_db[u.id] = (u.status, u.domain, u.domaininfo.status)
+        i += 1
+    print(f"... loaded {i} users")
+
+    # Dump the records
+    t1 = datetime.now()
+    print("Writing records ...")
+    query = f"(parent.is_verified:true AND is_published:true AND versions.is_latest:true) OR (is_deleted:true AND tombstone.removal_reason.id:spam AND versions.is_latest:true)"
+    s = Search(using=client, index=f'{idxprefix}rdmrecords-records').extra(
+        track_total_hits=True,
+    ).query(
+        "query_string",
+        query=query,
+    )
+
+    # Writing the CSV and compressing at the same time
+    fs = make_fs("records")
+    stamp = datetime.now().date().isoformat()
+    header = (
+        "label",
+        "id",
+        "created",
+        "domain",
+        "title_len",
+        "description_len",
+        "links",
+        "emojis",
+        "files",
+        "bytes",
+        "text",
+    )
+
+    l_header = (
+        "label",
+        "id",
+        "domain",
+    )
+
+    # TODO: write creators, links_domains in separate files.
+
+    try:
+        fp = fs.open(f"records-{stamp}.csv.gz", 'wb')
+        gzfile = gzip.GzipFile(fileobj=fp, mode='wb')
+        wrapper = io.TextIOWrapper(gzfile, encoding='utf-8', newline='')
+        writer = csv.writer(wrapper)
+
+        l_fp = fs.open(f"links-{stamp}.csv.gz", 'wb')
+        l_gzfile = gzip.GzipFile(fileobj=l_fp, mode='wb')
+        l_wrapper = io.TextIOWrapper(l_gzfile, encoding='utf-8', newline='')
+        l_writer = csv.writer(l_wrapper)
+
+        writer.writerow(header)
+        l_writer.writerow(l_header)
+        i = 0
+        for r in s.scan():
+            row = prepare_record(r)
+            if row is not None:
+                (label, id, created, udomain, title_len, description_len, links, emojis, files, bytes, text, links_domains) = row
+                writer.writerow((label, id, created, udomain, title_len, description_len, links, emojis, files, bytes, text))
+                i += 1
+                if i % 100000 == 0:
+                    print("...", i)
+                if links_domains:
+                    for l in links_domains:
+                        l_writer.writerow([label,id,l])
+
+        print(f"... wrote {i} records")
+    finally:
+        wrapper.close()
+        gzfile.close()
+        fp.close()
+
+        l_wrapper.close()
+        l_gzfile.close()
+        l_fp.close()
+
+    t2 = datetime.now()
+    print("Loading users", t1-t0)
+    print("Writing records", t2-t1)