Implement review comments from OA

README.md

@@ -37,8 +37,11 @@ pretty_name: Open-Assistant StackExchange Instruction
 
 This dataset is taken from https://archive.org/details/stackexchange.
 
-There's a single parquet file combining all stackexchange sites.
-The threads have been filtered as follows: only threads with an accepted answer, for which both the question and response is less than 1000 characters have been choosen. Other answers, or questions without accepted answers, or long entries have been droppped.
+There's a single parquet file combining all stackexchange sites. The threads
+have been filtered as follows: only threads with an accepted answer, for which
+both the question and response is less than 1000 characters have been choosen.
+Other answers, or questions without accepted answers, or long entries have been
+droppped.
 
 Each row consists of
 
@@ -49,6 +52,31 @@ Each row consists of
 
 Original extraction code by https://github.com/b-mc2
 
+
+## How to Reproduce this Dataset
+
+1. Download all XML files from the stackexchange archive into the xml/ folder
+```
+./download.py
+```
+2. Process the XML, filter conversations and convert to OA format into parquet/ folder
+
+```
+./process.py
+```
+3. Run stats on all files in the parquet/ folder
+```
+./stats.py
+```
+4. Combine all parquet files into one large stackexchange.parquet file
+```
+./combine.py
+```
+5. Upload to huggingface hub, you'll first need use huggingface-cli login
+```
+./upload.py
+```
+
 ## Statistics
 
 - 3dprinting: 1,006

combine.py

@@ -1,7 +1,8 @@
 #!/usr/bin/env python3
 # Combine (and shorten) parquet files into a single file
 
-import os, glob, sys, re
+import glob
+
 import pandas as pd
 from merge_parquets import merge_parquet_dir
 
@@ -15,18 +16,12 @@ for file in glob.glob("full/*.parquet"):
     df["METADATA"] = df["METADATA"].apply(
         lambda meta: {
             "tags": meta["tags"],
-            "answer_score": int(meta["answer_score"])
-            if "answer_score" in meta and meta["answer_score"]
-            else 0,
-            "question_score": int(meta["question_score"])
-            if "question_score" in meta and meta["question_score"]
-            else 0,
+            "answer_score": int(meta["answer_score"]) if "answer_score" in meta and meta["answer_score"] else 0,
+            "question_score": int(meta["question_score"]) if "question_score" in meta and meta["question_score"] else 0,
         }
     )
     df.to_parquet(file)
     after = len(df)
-    print(
-        f"Shortened {file} from {before} to {after} rows ({100 * after / before:.2f})"
-    )
+    print(f"Shortened {file} from {before} to {after} rows ({100 * after / before:.2f})")
 
 merge_parquet_dir("full", "stackexchange.parquet")

download.py

@@ -3,19 +3,19 @@
 # Simple script to download StackExchange archive XML files with posts (threaded version)
 #
 # Note: you probably want to download stackoverflow.com-Posts.7z manually, as it is 18GB
-# and takes a long time to download. You can try using torrent:
+# and takes a days to download otherwise. You can try using the torrent:
 #
 # webtorrent https://archive.org/download/stackexchange/stackexchange_archive.torrent --select 658
 #
 
-import requests
 import concurrent.futures
 import os
-from bs4 import BeautifulSoup as bs
-import pandas as pd
 import re
 
-base_url = "https://ia600107.us.archive.org/view_archive.php?archive=/27/items/stackexchange/{0}&file=Posts.xml"
+import requests
+from bs4 import BeautifulSoup as bs
+
+BASE_URL = "https://ia600107.us.archive.org/view_archive.php?archive=/27/items/stackexchange/{0}&file=Posts.xml"
 DOWNLOAD_DIR = "xml/"
 NUM_PARALLEL = 20
 RE_IGNORE = r"_meta|stackoverflow\.com\-"
@@ -24,31 +24,26 @@ RE_IGNORE = r"_meta|stackoverflow\.com\-"
 def get_all_filenames():
     """
     Retrieve all urls from stackexchange archive.
-    This needs quite some mangling because of special cases.
+    This needs quite some mangling because of special cases (i.e. stackoverflow is not in one 7z archive).
+    Ignore meta files.
     """
     response = requests.get("https://archive.org/download/stackexchange")
     if response.ok:
         soup = bs(response.content, "html.parser")
         table = soup.find("table")
         link_tags = table.find_all("a")
-        urls = {
-            "stackoverflow": "https://archive.org/download/stackexchange/stackoverflow.com-Posts.7z"
-        }
+        urls = {"stackoverflow": "https://archive.org/download/stackexchange/stackoverflow.com-Posts.7z"}
         for link in link_tags:
             url = link["href"]
             name = url.split(".stackexchange")[0].replace(".", "_").replace("-", "_")
             name = name.replace("_com_7z", "")
             if url.endswith("7z") and not re.search(RE_IGNORE, url):
-                urls[name] = base_url.format(url)
+                urls[name] = BASE_URL.format(url)
         return urls
 
 
-urls = get_all_filenames()
-
-
 def download_url(dataset_name: str, url: str):
-    if not os.path.exists(DOWNLOAD_DIR):
-        os.mkdir(DOWNLOAD_DIR)
+    os.makedirs(DOWNLOAD_DIR, exist_ok=True)
     cache_path = os.path.join(DOWNLOAD_DIR, dataset_name + ".xml")
     if os.path.exists(cache_path):
         print("Using cached: ", cache_path)
@@ -62,12 +57,18 @@ def download_url(dataset_name: str, url: str):
         return cache_path
 
 
-with concurrent.futures.ThreadPoolExecutor(max_workers=NUM_PARALLEL) as executor:
-    futures = [
-        executor.submit(download_url, dataset, url) for dataset, url in urls.items()
-    ]
+def download_all():
+    urls = get_all_filenames()
+    with concurrent.futures.ThreadPoolExecutor(max_workers=NUM_PARALLEL) as executor:
+        futures = [executor.submit(download_url, dataset, url) for dataset, url in urls.items()]
+
+    # Wait for all downloads to complete
+    concurrent.futures.wait(futures)
+    print("All downloads complete, except for the large stackoverflow XML file")
+    print("Use torrent to download this one much quicker, then uncompress the 7z file")
+    print("and move the extracted stackoverflow.com-Posts.xml to xml/stackoverflow.xml")
+    print("webtorrent https://archive.org/download/stackexchange/stackexchange_archive.torrent --select 658")
 
-# Wait for all downloads to complete
-concurrent.futures.wait(futures)
 
-print("All downloads complete")
+if __name__ == "__main__":
+    download_all()

merge_parquets.py

@@ -34,17 +34,7 @@ def stream_from_parquets(paths: Iterable[Path]) -> Iterable[pa.Table]:
 T = TypeVar("T")
 
 
-def coalesce(
-    items: Iterable[T], max_size: int, sizer: Callable[[T], int] = len
-) -> Iterable[list[T]]:
-    """Coalesce items into chunks. Tries to maximize chunk size and not exceed max_size.
-    If an item is larger than max_size, we will always exceed max_size, so make a
-    best effort and place it in its own chunk.
-    You can supply a custom sizer function to determine the size of an item.
-    Default is len.
-    >>> list(coalesce([1, 2, 11, 4, 4, 1, 2], 10, lambda x: x))
-    [[1, 2], [11], [4, 4, 1], [2]]
-    """
+def coalesce(items: Iterable[T], max_size: int, sizer: Callable[[T], int] = len) -> Iterable[list[T]]:
     batch = []
     current_size = 0
     for item in items:
@@ -59,9 +49,7 @@ def coalesce(
         yield batch
 
 
-def coalesce_parquets(
-    paths: Iterable[Path], outpath: Path, max_size: int = 2**20
-) -> None:
+def coalesce_parquets(paths: Iterable[Path], outpath: Path, max_size: int = 2**20) -> None:
     tables = stream_from_parquets(paths)
     # Instead of coalescing using number of rows as your metric, you could
     # use pa.Table.nbytes or something.

process.py

@@ -2,23 +2,24 @@
 # Simple script to convert StackExchange XML to Open Assistant format
 # Original code by https://github.com/b-mc2
 
-import os, gc, glob, sys, re
-from bs4 import BeautifulSoup as bs
+import gc
+import glob
+import os
+import re
+import subprocess
+import sys
+
 import pandas as pd
 from html2text import html2text
-from datasets import load_dataset
 from lxml import etree
 from tqdm import tqdm
-import subprocess
-from merge_parquets import merge_parquet_dir
-
 
 XML_DIR = "./xml"
 SOURCE = "stackexchange-{0}"
 MAX_ANSWERS = 10
 QUESTION_SCORE_TRESHOLD = 0
 ANSWER_SCORE_TRESHOLD = 0
-PARQUET_FILE = "{0}.parquet"
+PARQUET_FILE = "parquet/{0}.parquet"
 MAX_LENGTH = 1000  # max length of question or answer
 
 
@@ -115,9 +116,7 @@ def convert_to_oa(all, source):
 
     del all
 
-    merged["INSTRUCTION"] = (
-        merged["Title_q"] + "\n" + merged["Body_q"].apply(to_markdown)
-    )
+    merged["INSTRUCTION"] = merged["Title_q"] + "\n" + merged["Body_q"].apply(to_markdown)
     merged["RESPONSE"] = merged["Body_a"].apply(to_markdown)
     merged["SOURCE"] = source
     merged["METADATA"] = merged.apply(create_metadata, axis=1)
@@ -142,6 +141,7 @@ def save_parquet(df, dataset):
     Save Dataframe to Parquet. See here for specs:
     https://projects.laion.ai/Open-Assistant/docs/data/datasets#creating-a-dataset-on-hugging-face
     """
+    os.makedirs("parquet", exist_ok=True)
     parquet_file = PARQUET_FILE.format(dataset)
     df.to_parquet(parquet_file, row_group_size=100, engine="pyarrow", index=False)
     print(f"Converted {len(df)} instructions into {parquet_file}")

stats.md

@@ -1,186 +0,0 @@
-- 3dprinting: 1,006
-- academia: 6,956
-- ai: 1,169
-- android: 11,591
-- anime: 3,688
-- apple: 32,603
-- arduino: 3,725
-- askubuntu: 78,472
-- astronomy: 2,425
-- aviation: 4,945
-- avp: 1,949
-- beer: 387
-- bicycles: 4,835
-- bioacoustics: 70
-- bioinformatics: 903
-- biology: 5,344
-- bitcoin: 7,456
-- blender: 25,527
-- boardgames: 4,538
-- bricks: 1,457
-- buddhism: 911
-- cardano: 670
-- chemistry: 7,430
-- chess: 2,185
-- chinese: 4,897
-- christianity: 1,248
-- civicrm: 3,221
-- codegolf: 943
-- codereview: 2,171
-- coffee: 350
-- cogsci: 645
-- computergraphics: 540
-- conlang: 101
-- cooking: 7,951
-- craftcms: 4,533
-- crafts: 438
-- crypto: 4,425
-- cs: 9,478
-- cseducators: 71
-- cstheory: 2,196
-- datascience: 5,045
-- dba: 16,850
-- devops: 961
-- diy: 14,400
-- drones: 190
-- drupal: 24,090
-- dsp: 4,470
-- earthscience: 922
-- ebooks: 323
-- economics: 2,120
-- electronics: 41,717
-- elementaryos: 1,769
-- ell: 30,428
-- emacs: 7,140
-- engineering: 2,314
-- english: 42,415
-- eosio: 626
-- es_stackoverflow: 21,475
-- esperanto: 617
-- ethereum: 9,603
-- expatriates: 973
-- expressionengine: 3,638
-- fitness: 1,833
-- freelancing: 338
-- french: 5,193
-- gamedev: 9,678
-- gaming: 44,899
-- gardening: 4,492
-- genealogy: 487
-- german: 6,715
-- gis: 30,249
-- graphicdesign: 10,563
-- ham: 790
-- hardwarerecs: 647
-- health: 804
-- hermeneutics: 782
-- hinduism: 1,036
-- history: 1,776
-- homebrew: 2,357
-- hsm: 484
-- interpersonal: 199
-- iot: 331
-- iota: 292
-- islam: 1,496
-- italian: 1,356
-- ja_stackoverflow: 9,734
-- japanese: 13,862
-- joomla: 1,875
-- judaism: 6,156
-- korean: 754
-- languagelearning: 135
-- latin: 1,387
-- law: 3,475
-- lifehacks: 934
-- linguistics: 1,507
-- literature: 582
-- magento: 20,537
-- martialarts: 364
-- materials: 338
-- math: 501,019
-- matheducators: 316
-- mathematica: 19,529
-- mathoverflow_net_7z: 23,803
-- mechanics: 4,735
-- meta: 34,161
-- meta_askubuntu: 2,076
-- meta_mathoverflow_net_7z: 333
-- meta_serverfault: 823
-- meta_stackoverflow: 12,641
-- meta_superuser: 1,748
-- moderators: 39
-- monero: 1,443
-- money: 7,996
-- movies: 6,789
-- music: 5,740
-- musicfans: 781
-- mythology: 271
-- networkengineering: 4,637
-- opendata: 1,117
-- opensource: 805
-- or: 586
-- outdoors: 1,503
-- parenting: 815
-- patents: 582
-- pets: 1,081
-- philosophy: 1,505
-- photo: 6,386
-- physics: 35,386
-- pm: 982
-- poker: 431
-- politics: 1,903
-- portuguese: 658
-- proofassistants: 87
-- pt_stackoverflow: 27,650
-- puzzling: 11,959
-- quant: 3,303
-- quantumcomputing: 1,604
-- raspberrypi: 6,794
-- retrocomputing: 1,016
-- reverseengineering: 1,606
-- robotics: 1,020
-- rpg: 9,517
-- ru_stackoverflow: 106,714
-- rus: 8,210
-- russian: 1,960
-- salesforce: 27,962
-- scicomp: 1,403
-- scifi: 15,174
-- security: 11,733
-- serverfault: 81,229
-- sharepoint: 24,934
-- sitecore: 2,691
-- skeptics: 1,043
-- softwareengineering: 10,526
-- softwarerecs: 3,032
-- solana: 602
-- sound: 2,031
-- space: 3,145
-- spanish: 3,049
-- sports: 1,715
-- sqa: 1,944
-- stackapps: 702
-- stackoverflow: 4,269,779
-- stats: 23,102
-- stellar: 373
-- substrate: 812
-- superuser: 128,488
-- sustainability: 240
-- tex: 42,808
-- tezos: 635
-- tor: 887
-- travel: 9,957
-- tridion: 1,769
-- ukrainian: 577
-- unix: 54,338
-- ux: 7,403
-- vegetarianism: 151
-- vi: 4,360
-- webapps: 10,159
-- webmasters: 9,413
-- windowsphone: 1,110
-- woodworking: 677
-- wordpress: 24,270
-- workplace: 4,104
-- worldbuilding: 2,766
-- writers: 1,957

stats.py

@@ -1,4 +1,6 @@
-import re, glob
+import glob
+import re
+
 from pyarrow.parquet import ParquetDataset
 
 

upload.py

@@ -10,7 +10,7 @@ HF_DATASET = "donfu/oa-stackexchange"
 
 def upload_hf():
     """
-    Upload to Hugging Face
+    Upload to Hugging Face. Make sure you are logged in beforehand with `huggingface-cli login`
     """
     parquet_file = PARQUET_FILE
     dataset = load_dataset("parquet", data_files=parquet_file, name="oa-stackexchange")