upload create_board_sft_dataset.py

uploading script used to automate the dataset creation. Needs to be a board folder with thread text files in the same format as the ones found in the lesserfield/4chan-datasets repo (https://huggingface.co/datasets/lesserfield/4chan-datasets)

create_board_sft_dataset.py

@@ -0,0 +1,139 @@
+import os
+import json
+import sys
+import re
+from pathlib import Path
+import argparse
+import subprocess
+import importlib
+
+def install_tqdm_if_needed():
+    """Checks for tqdm and installs it if not found."""
+    try:
+        importlib.import_module('tqdm')
+    except ImportError:
+        print("`tqdm` library not found. Installing...")
+        try:
+            subprocess.check_call([sys.executable, "-m", "pip", "install", "tqdm"])
+        except subprocess.CalledProcessError as e:
+            print(f"Error: Failed to install tqdm. Please install it manually using 'pip install tqdm'.")
+            sys.exit(1)
+
+def create_sft_dataset(input_folder, output_file):
+    """
+    Scans a directory for .txt thread files, parses them, and creates a
+    JSONL dataset suitable for Supervised Fine-Tuning (SFT).
+    """
+    from tqdm import tqdm
+
+    folder_path = Path(input_folder)
+    output_path = Path(output_file)
+
+    if not folder_path.is_dir():
+        print(f"Error: The input folder '{input_folder}' does not exist.")
+        sys.exit(1)
+
+    try:
+        output_path.parent.mkdir(parents=True, exist_ok=True)
+    except Exception as e:
+        print(f"Error: Could not create output directory '{output_path.parent}'. Error: {e}")
+        sys.exit(1)
+        
+    thread_files = list(folder_path.glob("*.txt"))
+
+    if not thread_files:
+        print(f"Error: No .txt files found in '{input_folder}'.")
+        sys.exit(1)
+        
+    print(f"Found {len(thread_files)} thread files. Processing...")
+    
+    total_conversations = 0
+    system_prompt = "You are an anonymous message board user. You are opinionated and use informal language. You discuss various topics."
+    
+    reply_pattern = re.compile(r'>>(\d{8,9})')
+
+    with open(output_path, 'w', encoding='utf-8') as f_out:
+        for file_path in tqdm(thread_files, desc="Processing Threads"):
+            try:
+                with open(file_path, 'r', encoding='utf-8') as f:
+                    content = f.read()
+                
+                posts_map = {}
+                ordered_post_ids = []
+                raw_posts = [p.strip() for p in content.split('--- ') if p.strip()]
+                
+                for post_text in raw_posts:
+                    lines = post_text.splitlines()
+                    if lines and lines[0].strip().isdigit():
+                        post_id = lines[0].strip()
+                        post_content = "\n".join(lines[1:]).strip()
+                        if post_content:
+                            posts_map[post_id] = post_content
+                            ordered_post_ids.append(post_id)
+                
+                if not ordered_post_ids:
+                    continue
+                    
+                op_id = ordered_post_ids[0]
+                op_content = posts_map[op_id]
+
+                # --- LOGIC REMAINS THE SAME FOR DIRECT REPLIES ---
+                for post_id, post_content in posts_map.items():
+                    reply_ids = reply_pattern.findall(post_content)
+                    if reply_ids:
+                        assistant_text = post_content.strip()
+                        for parent_id in reply_ids:
+                            if parent_id in posts_map:
+                                human_text = posts_map[parent_id]
+                                conversation = {
+                                    "conversations": [
+                                        {"from": "system", "value": system_prompt},
+                                        {"from": "human", "value": human_text},
+                                        {"from": "assistant", "value": assistant_text}
+                                    ]
+                                }
+                                f_out.write(json.dumps(conversation) + '\n')
+                                total_conversations += 1
+
+                # --- NEW LOGIC FOR STANDALONE POSTS ---
+                for post_id, post_content in posts_map.items():
+                    # A post is standalone if it's not the OP and has no reply links
+                    if post_id != op_id and not reply_pattern.search(post_content):
+                        assistant_text = post_content.strip()
+                        conversation = {
+                            "conversations": [
+                                {"from": "system", "value": system_prompt},
+                                {"from": "human", "value": op_content},
+                                {"from": "assistant", "value": assistant_text}
+                            ]
+                        }
+                        f_out.write(json.dumps(conversation) + '\n')
+                        total_conversations += 1
+
+            except Exception as e:
+                tqdm.write(f"Warning: Could not read or process file {file_path}. Error: {e}")
+
+    print(f"\nProcessing complete.")
+    print(f"Successfully created SFT dataset with {total_conversations} conversations.")
+    print(f"Output file: {output_path}")
+
+if __name__ == "__main__":
+    install_tqdm_if_needed()
+
+    parser = argparse.ArgumentParser(
+        description="Create an SFT dataset in JSONL format from a folder of message board thread files.",
+        formatter_class=argparse.RawTextHelpFormatter
+    )
+    parser.add_argument(
+        "input_folder",
+        help="The path to the folder containing your .txt thread files."
+    )
+    parser.add_argument(
+        "-o", "--output_file",
+        default="sft_dataset.jsonl",
+        help="The name of the output .jsonl file. Defaults to 'sft_dataset.jsonl'."
+    )
+    
+    args = parser.parse_args()
+    
+    create_sft_dataset(args.input_folder, args.output_file)