merge.py

import os
import sys
import json
import tqdm
import argparse
import transformers

def parse_args():
    parser = argparse.ArgumentParser(description="Merge the boards into a single train set.")
    parser.add_argument("-t", "--tokenizer-name-or-path", default="meta-llama/Llama-3.1-8B", help="The name or path for the tokenizer")
    parser.add_argument("-l", "--limit", type=int, default=4096, help="Length limit in tokens for each post")
    parser.add_argument("-m", "--min", type=int, default=5, help="Minimum amount of message in each post")
    parser.add_argument("-i", "--id", default="<|start_header_id|>", help="Prefix token for message IDs")
    parser.add_argument("-c", "--content", default="<|end_header_id|>", help="Prefix token for message contents")
    return parser.parse_args()

def main():
    args = parse_args()
    tokenizer = transformers.AutoTokenizer.from_pretrained(args.tokenizer_name_or_path, legacy=True)
    if args.id not in tokenizer.vocab:
        print(f"The message ID prefix token \"{args.id}\" is not a token in \"{args.tokenizer_name_or_path}\", it will work but it's better to be a token in the tokenizer.")
    if args.content not in tokenizer.vocab:
        print(f"The message content prefix token \"{args.content}\" is not a token in \"{args.tokenizer_name_or_path}\", it will work but it's better to be a token in the tokenizer.")
    boards_dir = "boards"
    total_token_count = 0
    with open("merged_strings_train.jsonl", "w", encoding="utf8") as output:
        for board_path in tqdm.tqdm(os.listdir(boards_dir), desc="Boards"):
            board_name, ext = os.path.splitext(board_path)
            if ext != ".json":
                continue
            board_path = os.path.join(boards_dir, board_path)
            if not os.path.isfile(board_path):
                continue
            with open(board_path, "r", encoding="utf8") as f:
                board = json.load(f)
            for post in tqdm.tqdm(board, desc="Posts"):
                if len(post) < args.min:
                    continue
                post_content = board_name
                post_token_count = len(tokenizer.encode(post_content, add_special_tokens=False)) + 2 # Add 2 for the start of string and end of string tokens.
                for message in post:
                    formatted = f"{args.id}{message["id"]}{args.content}{message["content"]}"
                    formatted_token_count = len(tokenizer.encode(formatted, add_special_tokens=False))
                    added_token_count = post_token_count + formatted_token_count
                    if added_token_count > args.limit:
                        break
                    post_content += formatted
                    post_token_count = added_token_count
                json.dump({"input": "", "output": post_content}, output, ensure_ascii=False)
                output.write("\n")
                total_token_count += post_token_count
    print("Merge finished, total token count:", total_token_count)

if __name__ == "__main__":
    try:
        main()
    except KeyboardInterrupt:
        print("\nScript interrupted by user, exiting...")
        sys.exit(1)