| import os | |
| import numpy as np | |
| import random | |
| from transformers import GPT2Tokenizer | |
| # --- Settings --- | |
| BIN_PATH = "data_19b.bin" | |
| NUM_SAMPLES = 10 | |
| SAMPLE_LEN = 200 # Number of tokens to decode per sample | |
| def main(): | |
| if not os.path.exists(BIN_PATH): | |
| print(f"Error: {BIN_PATH} not found.") | |
| return | |
| # 1. Load the tokenizer | |
| print("Loading GPT-2 Tokenizer...") | |
| tokenizer = GPT2Tokenizer.from_pretrained("EleutherAI/gpt-neox-20b") | |
| # 2. Map the data (instantly points to the 40GB file) | |
| data = np.memmap(BIN_PATH, dtype=np.uint16, mode='r') | |
| total_tokens = len(data) | |
| print(f"File loaded. Total tokens: {total_tokens:,}") | |
| # 3. Pick random spots and decode | |
| print(f"\n--- Decoding {NUM_SAMPLES} Random Samples ---\n") | |
| for i in range(NUM_SAMPLES): | |
| # Pick a random starting index | |
| start_idx = random.randint(0, total_tokens - SAMPLE_LEN) | |
| end_idx = start_idx + SAMPLE_LEN | |
| # Pull the uint16 tokens and convert to standard Python list | |
| token_ids = data[start_idx:end_idx].tolist() | |
| # Decode to text | |
| decoded_text = tokenizer.decode(token_ids, skip_special_tokens=True) | |
| print(f"Sample {i+1} (Index {start_idx:,}):") | |
| print("-" * 50) | |
| print(decoded_text) | |
| print("-" * 50 + "\n") | |
| if __name__ == "__main__": | |
| main() |