subset of modded nanogpt data, estimate according to
```
import glob, array
from pathlib import Path
import torch # only used for fast host-side dtype conversion
# ------------------------------------------------------------------
# 1. How many tokens do we need?
# ------------------------------------------------------------------
NUM_GPUS = 8
GRADIENT_ACCUM_STEPS = 1 # “no gradient accumulation”
TRAIN_SEQ_LEN = 48 * 1024 # 49 152
NUM_ITERATIONS = 1_770
TOKENS_PER_STEP = NUM_GPUS * TRAIN_SEQ_LEN // GRADIENT_ACCUM_STEPS
TOKENS_NEEDED = TOKENS_PER_STEP * NUM_ITERATIONS # 695 992 320
print(f"{TOKENS_NEEDED:,} tokens will be copied into the new file")
# ------------------------------------------------------------------
# 2. Helper to load a shard (same logic as _load_data_shard in train_gpt.py)
# ------------------------------------------------------------------
MAGIC = 20240520
VERSION = 1
H…
