prepare_finetune.py

import os
import numpy as np
import tiktoken
from datasets import load_dataset
from tqdm import tqdm

OUTPUT_DIR = "data/alpaca_cleaned_mixed"
TOKENIZER_NAME = "gpt2" 
SEED = 1337

FINEWEB_SAMPLES = 2500 

enc = tiktoken.get_encoding(TOKENIZER_NAME)
EOS_TOKEN = "<|endoftext|>"

def format_prompt_with_mask(instruction, input_text, output):
    """
    Formatiert den Prompt und erstellt die Loss-Maske.
    Format:
    Instruction: ...
    Input: ... (optional)
    Response: ... <|endoftext|>
    """
    if input_text and input_text.strip():
        prompt_text = f"Instruction:\n{instruction}\n\nInput:\n{input_text}\n\nResponse:\n"
    else:
        prompt_text = f"Instruction:\n{instruction}\n\nResponse:\n"
    
    completion_text = f"{output}{EOS_TOKEN}"
    
    prompt_ids = enc.encode(prompt_text, allowed_special={'<|endoftext|>'})
    completion_ids = enc.encode(completion_text, allowed_special={'<|endoftext|>'})
    
    full_ids = prompt_ids + completion_ids
    
    mask = [0] * len(prompt_ids) + [1] * len(completion_ids)
    
    return full_ids, mask

def main():
    np.random.seed(SEED)
    print(f"🚀 Starting Prepare-Script for SmaLLMPro (350M Instruct)...")
    print(f"📚 Tokenizer: {TOKENIZER_NAME}")
    
    os.makedirs(OUTPUT_DIR, exist_ok=True)

    print("📥 Loading 'yahma/alpaca-cleaned' (Chat-Instructions)...")
    alpaca = load_dataset("yahma/alpaca-cleaned", split='train')
    
    print(f"📥 Loading 'HuggingFaceFW/fineweb-edu' (Sample-10BT) for {FINEWEB_SAMPLES} Samples...")
    fineweb = load_dataset("HuggingFaceFW/fineweb-edu", name="sample-10BT", split='train', streaming=True)
    
    all_tokens = []
    all_masks = []
    
    print("⚙️  Processing Alpaca...")
    for ex in tqdm(alpaca, desc="Alpaca"):
        ids, mask = format_prompt_with_mask(ex['instruction'], ex['input'], ex['output'])
        all_tokens.extend(ids)
        all_masks.extend(mask)
        
    alpaca_len = len(all_tokens)
    print(f"   -> Alpaca Tokens: {alpaca_len:,}")

    print("⚙️  Processing FineWeb (Anti-Forgetting)...")
    fw_iter = iter(fineweb)
    fw_count = 0
    fw_tokens_count = 0
    
    for _ in tqdm(range(FINEWEB_SAMPLES), desc="FineWeb"):
        try:
            ex = next(fw_iter)
            text = ex['text'] + EOS_TOKEN
            ids = enc.encode(text, allowed_special={EOS_TOKEN})
            
            all_tokens.extend(ids)
            all_masks.extend([1] * len(ids)) 
            
            fw_tokens_count += len(ids)
            fw_count += 1
        except StopIteration:
            break
            
    print(f"   -> FineWeb Tokens: {fw_tokens_count:,} (from {fw_count} documents)")

    total_tokens = len(all_tokens)
    print(f"\n💾 Saving {total_tokens:,} Tokens in '{OUTPUT_DIR}'...")
    
    token_arr = np.array(all_tokens, dtype=np.uint16)
    token_arr.tofile(os.path.join(OUTPUT_DIR, "train.bin"))
    
    mask_arr = np.array(all_masks, dtype=np.uint8)
    mask_arr.tofile(os.path.join(OUTPUT_DIR, "train_mask.bin"))
    
    print("\n🔍 --- SANITY CHECK ---")
    print("I decode the first 50 tokens of the first sample, to check, if everything is okay.")
    print("Green (TRAIN) = The things the model learns. Grey (IGNORE) = The things the model only reads.")
    
    check_len = 100
    sample_ids = all_tokens[:check_len]
    sample_mask = all_masks[:check_len]
    
    decoded_parts = []
    for t_id, m_val in zip(sample_ids, sample_mask):
        token_str = enc.decode([t_id])
        if m_val == 1:
            decoded_parts.append(f"\033[92m{token_str}\033[0m")
        else:
            decoded_parts.append(f"\033[90m{token_str}\033[0m")
            
    print("".join(decoded_parts))
    print("\n(Legend: \033[90mGrey=Prompt/Ignored\033[0m, \033[Green=Response/Learned\033[0m)")
    
    if len(token_arr) != len(mask_arr):
        print("\n❌ Warning: Token and Mask Array have different lengths! Something has gone wrong!")
    else:
        print("\n✅ Everything seems to be fine. The arrays are synchronized. You can now start the training.")

if __name__ == "__main__":
    main()