H200 FineWeb Tokenization Summary | Generated by AI

Home 2026.07

Here’s the full summary of everything done on the H200 server:


Server: 38.80.152.249:30416 — NVIDIA H200 (140 GB VRAM, CUDA 13.0) Workdir: /workspace/zz/

1. Setup

2. Data Download

3. Tokenization

4. Results

Metric Value
Output shards 358 (1 val + 357 train)
Total tokens ~35.8B (incl. EOT prefixes)
Output location /workspace/zz/datasets/tokenized/
Output files edufineweb_val_000000.npy + edufineweb_train_000001-000357.npy
Disk used ~104 GB on /workspace (187 GB total)
Wall time ~45 min

5. Commands

# Download (already done)
cd /workspace/zz
python3 h200/download_fineweb.py

# Tokenize (already done)
cd /workspace/zz
python3 h200/tokenize_fineweb.py

The tokenized data is ready for GPT-3 ablation training — nanoGPT-compatible format.


Back Donate