[Dataset] CLI-1M: 975K NL→shell pairs — 13 languages, 6 shells, Apache-2.0

Hi HF community! I just published carosh/cli-1m — the first large-scale multilingual dataset for natural-language → shell command generation.

Numbers:

• 975,933 training pairs

• 6 shells: bash, zsh, fish, PowerShell, nushell, oils-osh

• 13 languages

• 18 industry buckets

• 108× NL2Bash (the previous reference dataset)

Load it:

from datasets import load_dataset

# SFT training

ds = load_dataset(“carosh/cli-1m”, split=“train”)

# 50k browse-friendly subset

ds = load_dataset(“carosh/cli-1m”, name=“sample”, split=“train”)

# Domain-specific (security only)

ds = load_dataset(“carosh/cli-1m”, name=“domains”, split=“security”)

Interactive explorer: CLI-1M Explorer - a Hugging Face Space by carosh

Help wanted: Looking for native speakers of Hebrew, Arabic, Hindi, Korean, or Russian to spot-check 50 translations each (~30 min, full credit in dataset card). DM @CaroDaShellShib on X or reply here to join the contributor waitlist.

Apache-2.0. Feedback welcome.