scripts/readme

README.md

@@ -36,7 +36,7 @@ pretty_name: ArXiv-CC0
 ### Dataset Description
 
 *ArXiv CC0* is a cleaned dataset of a raw scrape of arXiv using the latest metadata from January 2024.
-Filtering to a total amount of tokens of **~-B** (llama-2-7b-chat-tokenizer) / **~-B** (RWKV Tokenizer) from primarily English language.
+Filtering to a total amount of tokens of **~2.77B** (llama-2-7b-chat-tokenizer) / **~2.43B** (RWKV Tokenizer) from primarily English language.
 
 - **Curated by:** M8than
 - **Funded by [optional]:** Recursal.ai
@@ -53,6 +53,11 @@ Filtering to a total amount of tokens of **~-B** (llama-2-7b-chat-tokenizer) / *
 
 The entries are filtered down to CC0 only, so these are all public domain. The PDF's have their text pulled from them (not via OCR, the text is embedded). The documents are then run through a filtering and text transformation process within our pipeline to produce readable outputs, filtering out excessive newlines or spaces and also converting to markdown.
 
+#### How to run the scripts:
+- Download the metadata json file from kaggle
+- Run process_metadata_PDFs.py (after adjusting the variables at the top)
+- Run filter.py (after adjusting the variables at the top)
+
 ### Data Splits
 
 - final

filter.py

@@ -0,0 +1,129 @@
+import json
+import os
+import re
+
+dir_in = "raw_jsonl"
+dir_out = "jsonl"
+
+def remove_single_newlines_with_space(text):
+    new_text = []
+    i = 0
+
+    def is_valid_char(c):
+        # Check if the character is a letter, space, quote or closing parenthesis
+        return c.isalpha() or c in {' ', '"', "'", ')', ']', '(' , '[', '-', ',', '.', '!', '?', ';', ':'} or c.isdigit()
+
+    while i < len(text):
+        if i > 0 and i < len(text) - 1 and text[i] == '\n':
+            # Check if the characters before and after the newline are valid
+            if is_valid_char(text[i-1]) and is_valid_char(text[i+1]):
+                new_text.append(' ')  # Add a space instead of the newline
+                i += 1
+                continue
+        new_text.append(text[i])
+        i += 1
+
+    return ''.join(new_text)
+
+char_count_per_chunk = 10737418240
+cur_dir = os.path.dirname(os.path.realpath(__file__))
+
+allowed_licenses = [
+    "https://creativecommons.org/share-your-work/public-domain/cc0/",
+    "http://creativecommons.org/share-your-work/public-domain/cc0/",
+    "http://creativecommons.org/publicdomain/zero/1.0/",
+    "http://creativecommons.org/licenses/publicdomain/",
+]
+
+disallowed_licenses = [
+    "https://creativecommons.org/licenses/by/4.0/",
+    "http://creativecommons.org/licenses/by/4.0/",
+    "http://creativecommons.org/licenses/by/3.0/",
+    
+    
+    "https://creativecommons.org/licenses/by-sa/4.0/",
+    "http://creativecommons.org/licenses/by-sa/4.0/",
+    
+    "https://creativecommons.org/licenses/by-nc-sa/4.0/",
+    "http://creativecommons.org/licenses/by-nc-sa/4.0/",
+    "http://creativecommons.org/licenses/by-nc-sa/3.0/",
+    
+    "https://creativecommons.org/licenses/by-nc/4.0/",
+    "http://creativecommons.org/licenses/by-nc/4.0/",
+    
+    "https://creativecommons.org/licenses/by-nc-nd/4.0/",
+    "http://creativecommons.org/licenses/by-nc-nd/4.0/",
+    
+    "https://arxiv.org/licenses/nonexclusive-distrib/1.0/",
+    "http://arxiv.org/licenses/nonexclusive-distrib/1.0/",
+]
+
+# generate a list of file paths from raw_jsonl
+file_paths = os.listdir(cur_dir + '/' + dir_in)
+
+# sort file paths in alphanumeric order
+def alphanumeric_sort(file_paths):
+    # Function to convert text to integer if possible
+    convert = lambda text: int(text) if text.isdigit() else text.lower()
+    # Function to split the file path into parts and convert numbers
+    alphanum_key = lambda key: [convert(c) for c in re.split('([0-9]+)', key)]
+    # Sort the list with the custom key
+    return sorted(file_paths, key=alphanum_key)
+
+file_paths = alphanumeric_sort(file_paths)
+
+file_paths = [cur_dir + f'/{dir_in}/' + file for file in file_paths]
+
+
+buffer = []
+current_char_count = 0
+chunk_id = 0
+# loop through each file path building up the buffer till its over the char count
+print("Processing files...")
+line_count = 0
+for file in file_paths:
+    print("Processing file " + file + "...")
+    with open(file) as f:
+        for line in f.readlines():
+            line_json = json.loads(line)
+            if line_json['license'] == None:
+                continue
+            elif line_json['license'] not in allowed_licenses:
+                if line_json['license'] not in disallowed_licenses:
+                    # log the license
+                    print("License not in allowed licenses or disallowed licenses: " + line_json['license'])
+                continue
+            
+            
+            current_char_count += len(line)
+            
+            # filter line text here
+            
+            line_json['text'] = line_json['text'].replace('\\"', '')
+            line_json['text'] = line_json['text'].replace('\"', '')
+            line_json['text'] = line_json['text'].replace('\\', '')
+            line_json['text'] = line_json['text'].replace('\f', '')
+            line_json['text'] = remove_single_newlines_with_space(line_json['text'])
+    
+            buffer.append(line_json)
+            line_count += 1
+            
+            if current_char_count > char_count_per_chunk:
+                # write buffer to file
+                with open(cur_dir + f'/{dir_out}/' + "arxiv_" + str(chunk_id) + ".jsonl", 'w') as f:
+                    for line in buffer:
+                        f.write(json.dumps(line) + '\n')
+                        
+                # reset buffer and char count
+                buffer = []
+                current_char_count = 0
+                chunk_id += 1
+                
+            if line_count % 1000 == 0:
+                print("Found " + str(line_count) + " copyright-free lines...")
+                
+                
+
+with open(cur_dir + f'/{dir_out}/' + "arxiv_" + str(chunk_id) + ".jsonl", 'w') as f:
+    for line in buffer:
+        f.write(json.dumps(line) + '\n')

process_metadata_PDFs.py

@@ -0,0 +1,321 @@
+dir = "20-Jan-2024"
+chunk_doc_size = 100000
+remove_every_word_over_x_characters = 16
+ignore_first_page = True
+ignore_last_page = True
+max_processes = 12
+attempt_resume = True
+remove_false_newlines = True
+
+import re
+import requests
+import threading
+import PyPDF2
+from io import BytesIO
+import time
+
+def remove_single_newlines_with_space(text):
+    new_text = []
+    i = 0
+
+    def is_valid_char(c):
+        # Check if the character is a letter, space, quote or closing parenthesis
+        return c.isalpha() or c in {' ', '"', "'", ')', ']', '(' , '[', '-', ',', '.', '!', '?', ';', ':'} or c.isdigit()
+
+    while i < len(text):
+        if i > 0 and i < len(text) - 1 and text[i] == '\n':
+            # Check if the characters before and after the newline are valid
+            if is_valid_char(text[i-1]) and is_valid_char(text[i+1]):
+                new_text.append(' ')  # Add a space instead of the newline
+                i += 1
+                continue
+        new_text.append(text[i])
+        i += 1
+
+    return ''.join(new_text)
+
+def extract_text_from_pdf_url(pdf_url):
+    # Send a GET request to the URL
+    response = requests.get(pdf_url)
+    
+    # Check if the request was successful
+    if response.status_code == 200:
+        # Read the PDF file from the response
+        file = BytesIO(response.content)
+        
+        # Create a PDF reader object
+        reader = PyPDF2.PdfReader(file)
+        
+        # Initialize a variable to store extracted text
+        text = ""
+        
+        # Loop through each page in the PDF
+        for page in range(len(reader.pages)):
+            if page == 0 and ignore_first_page:
+                continue
+            
+            if page == len(reader.pages) - 1 and ignore_last_page:
+                continue
+
+            # Get a specific page
+            pdf_page = reader.pages[page]
+            
+            # Extract text from the page
+            text += pdf_page.extract_text()
+        
+        return text
+    else:
+        return ""
+
+
+# read json file
+import json
+
+print("Loading arxiv metadata")
+
+file = open(f"{dir}/arxiv-metadata-oai-snapshot.json", "r")
+# read it line by line because its a jsonl
+data = []
+for line in file.readlines():
+    # load the json
+    data.append(json.loads(line))
+file.close()
+
+print(f"Loaded {len(data)} documents from metadata")
+
+
+to_skip = 0
+chunk = 0
+count = 0
+
+print("Reading previous data to resume progress")
+
+if attempt_resume:
+    # count how many lines in jsonl folder
+    import os
+    for filename in os.listdir('jsonl'):
+        if filename.endswith(".jsonl"):
+            chunk += 1
+            count += sum(1 for line in open(f'jsonl/{filename}', 'r'))
+            count -= 1 # remove the last line because it is empty
+            
+    # count how many lines in failed.txt
+    count += sum(1 for line in open('failed.txt', 'r'))
+    
+    # count how many lines in failed.jsonl
+    count += sum(1 for line in open('failed.jsonl', 'r'))
+    
+    
+    count -= 2 # remove the last lines because it is empty
+    chunk -= 1
+    to_skip = count
+    
+print(f"Skipping {to_skip} documents (to resume progress)")
+
+# # id: ArXiv ID (can be used to access the paper, see below)
+# # submitter: Who submitted the paper
+# # authors: Authors of the paper
+# # title: Title of the paper
+# # comments: Additional info, such as number of pages and figures
+# # journal-ref: Information about the journal the paper was published in
+# # doi: [https://www.doi.org](Digital Object Identifier)
+# # abstract: The abstract of the paper
+# # categories: Categories / tags in the ArXiv system
+# # versions: A version history
+# with open(f'jsonl/arxiv_{chunk}.jsonl', 'w') as f:
+#     for i in range(len(data)):
+#         print(f"Processing {i} of {len(data)}")
+#         row = data[i]
+
+        
+        
+#         # extract text from pdf file
+#         text = extract_text_from_pdf_url("https://arxiv.org/pdf/" + row['id'])
+
+#         if text == "":
+#             with open(f'failed.txt') as w:
+#                 w.write(row['id'] + '\n')
+
+#         f.write(json.dumps({"text": text.encode('utf-8').decode('unicode_escape')} + row) + '\n')
+#         if count % 100000 == 0:
+#             print(text)
+#         count+=1
+#         if count % chunk_doc_size == 0:
+#             f.close()
+#             chunk += 1
+#             f = open(f'jsonl/arxiv_{chunk}.jsonl', 'w')
+
+import threading
+import json
+from unidecode import unidecode
+import multiprocessing
+
+def filter_text(text):
+    final_text = text.encode('utf-8', errors='replace').decode('unicode_escape', errors='ignore')
+    final_text = re.sub(r'\\U[\da-fA-F]{8}', '', final_text)
+
+    # remove_every_word_over_x_characters
+    words = final_text.split(' ')
+    for word in words:
+        if len(word) > remove_every_word_over_x_characters:
+            final_text = final_text.replace(word, '')
+        if word.count('\\') > len(word) * 0.25:
+            final_text = final_text.replace(word, '')
+            
+            
+    # remove_false_newlines
+    if remove_false_newlines:
+        final_text = remove_single_newlines_with_space(final_text)
+        
+        
+    final_text = final_text.replace('\\"', '')
+    final_text = final_text.replace('\"', '')
+    final_text = final_text.replace('\\', '')
+    final_text = final_text.replace('\f', '')
+    return final_text
+    
+
+def extract_text_process(row):
+    text = extract_text_from_pdf_url("https://arxiv.org/pdf/" + row['id'])
+
+            
+    return filter_text(text)
+    
+def extract_text_process_safe(row):
+    try:
+        return extract_text_process(row)
+    except Exception as e:
+        print(e)
+        return "" 
+
+
+from queue import Empty
+def process_queue(queue, count, chunk, chunk_doc_size, stop_event):
+    output_file = open(f'jsonl/arxiv_{chunk}.jsonl', 'a')
+    failed_file = open('failed.jsonl', 'a')
+    
+    print("Starting queue process")
+    print(stop_event.is_set())
+    
+    while not stop_event.is_set():
+        while not queue.empty():
+            try:
+                row, text = queue.get(timeout=0.5)  # Timeout to check for stop event
+            except Empty as e:
+                print(e)
+                continue
+            print("Looping queue process " + str(chunk) + "  " + str(queue.qsize()))
+
+            row['title'] = filter_text(row['title'])
+            row['abstract'] = filter_text(row['abstract'])
+            
+            if len(text) < 1000:
+                failed_file.write(json.dumps(row) + '\n')
+            else:
+                output_file.write(json.dumps({**{"text": text}, **row}) + '\n')
+            
+            count += 1
+            if count % chunk_doc_size == 0:
+                output_file.close()
+                chunk += 1
+                output_file = open(f'jsonl/arxiv_{chunk}.jsonl', 'w')
+            time.sleep(0.005)  # Sleep 
+        time.sleep(0.005)  # Sleep 
+
+    output_file.close()
+    failed_file.close()
+
+def process_worker(queue, count, chunk, chunk_doc_size, stop_event):
+    try:
+        process_queue(queue, count, chunk, chunk_doc_size, stop_event)
+    except Exception as e:
+        print(f"Error in worker process: {e}")
+        # exit whole app
+        os._exit(1)
+        
+def worker(i, row, queue):
+    try:
+        text = extract_text_process_safe(row) 
+        queue.put((row, text))
+        print(f"Processed {i} of {len(data)}") # Adjust this line as needed
+    except Exception as e:
+        print(f"Error in worker process: {e}")
+    time.sleep(0.5)  # Sleep to avoid spamming requests
+
+    
+from concurrent.futures import ThreadPoolExecutor
+
+def process_data(data, chunk_doc_size):
+    global to_skip, count, chunk
+    queue = multiprocessing.Manager().Queue()
+    stop_event = multiprocessing.Event()
+    
+    stop_event.clear()
+
+    # Start the thread for processing the queue
+    queue_process = multiprocessing.Process(target=process_worker, args=(queue, count, chunk, chunk_doc_size, stop_event))
+    queue_process.start()
+
+    # Create a pool of worker threads
+    # with ThreadPoolExecutor(max_workers=max_processes) as executor:
+    #     for i, row in enumerate(data):
+    #         if i < to_skip:
+    #             continue
+
+    #         executor.submit(worker, i, row, queue)
+    #         time.sleep(0.0001)
+
+    # Create a pool of worker processes
+    with multiprocessing.Pool(max_processes) as pool:
+        for i, row in enumerate(data):
+            if i < to_skip:
+                continue
+
+            pool.apply_async(worker, args=(i, row, queue))
+            time.sleep(0.0001)
+
+        # Close the pool and wait for the work to finish
+        pool.close()
+        pool.join()
+
+    # Signal the queue processing thread to stop and wait for it to finish
+    stop_event.set()
+    queue_process.join()
+
+# output_file = open(f'jsonl/arxiv_{chunk}.jsonl', 'a')
+# failed_file = open('failed.txt', 'a')
+
+# def process_results(result):
+#     global count, chunk, failed_file, output_file, chunk_doc_size
+#     row, text = result
+#     if text == "":
+#         failed_file.write(row['id'] + '\n')
+#     else:
+#         output_file.write(json.dumps({**{"text": text}, **row}) + '\n')
+        
+#     if count % 100 == 0:
+#         print(f"Processed {count} of {len(data)}")
+
+#     count += 1
+        
+#     if count % chunk_doc_size == 0:
+#         output_file.close()
+#         chunk += 1
+#         output_file = open(f'jsonl/arxiv_{chunk}.jsonl', 'w')
+        
+    
+
+# def process_data(data, chunk_doc_size, pool_size=20):
+#     global output_file, failed_file
+
+#     with multiprocessing.Pool(pool_size) as pool:
+#         results = [pool.apply_async(extract_text_process, args=(data[i],), callback=lambda res: process_results(res)) for i in range(to_skip, len(data))]
+
+#         for result in results:
+#             result.wait()
+
+#     failed_file.close()
+#     output_file.close()
+
+# Example usage
+process_data(data, chunk_doc_size)