Offline translation across ~419 languages via MADLAD-400-3B-MT (CTranslate2 + SentencePiece). No cloud. No account. The model downloads once (~3 GB) and runs locally on CPU or Apple Silicon thereafter. Use whenever the user wants to translate text TO or FROM a non-English language — short phrases, paragraphs, whole documents, idioms, low-resource languages, indigenous languages, dead languages MADLAD has seen, polite forms, casual registers. Also trigger on "translate", "say this in <lang>", "what does <foreign text> mean", "render this in <language>", "<lang> for <phrase>", "how do you say X in Y", "localize", or any request that involves moving text between human languages. DO NOT use for code translation between programming languages. DO NOT use for transliteration (script conversion) without a target meaning — MADLAD translates meaning, not glyphs.
82%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Passed
No findings from the security scan
Offline machine translation in ~419 languages, running entirely on the user's machine. Backed by Google's MADLAD-400-3B-MT (Apache 2.0), served through CTranslate2 with int8 quantization for fast CPU/Metal inference, and SentencePiece for tokenization.
The model downloads once on first use (~3 GB) into
~/.local/share/translator/madlad400-3b-ct2/ and never phones home
again. After install, translation works with the network unplugged.
Use it for:
--list-langs or the
model card for the full set)Don't use it for:
For higher quality on low-resource languages specifically (under-served
African, indigenous American, and Pacific languages), the user can opt
into the NLLB-200 paradigm — see rness/paradigms/translation.md.
NLLB-200 is CC-BY-NC, so it's gated behind that paradigm flag for
personal-use-only contexts. Don't reach for it without checking the
paradigm.
Python deps (once):
pip install -r requirements.txtThen download the default model (~3 GB, one-time):
python scripts/bootstrap.py --installscripts/bootstrap.py --check is a fast yes/no gate you can call
before any translation to confirm the model is present.
Before translating, confirm the model is present:
python scripts/bootstrap.py --checkIf the exit code is non-zero, install it:
python scripts/bootstrap.py --installThen translate. Three invocation patterns, in order of preference:
1. Single short phrase, text on the command line:
python scripts/translate.py --target es "Hello, world."The output is the translated text on stdout. The <2es> prefix MADLAD
expects is added for you — you pass plain text, you get plain text.
2. Paragraph or document via stdin:
cat draft.md | python scripts/translate.py --target jaParagraph breaks (blank lines) are preserved; each paragraph is translated independently so the model doesn't lose track over very long inputs.
3. Many short strings via stdin, one per output line:
python scripts/translate.py --target fr --batch < strings.txtUse this for UI string lists, glossaries, or any case where each input line is its own translation unit.
scripts/translate.pypython scripts/translate.py --target <lang> [text]
python scripts/translate.py --target <lang> < file.txt
python scripts/translate.py --target <lang> --batch < lines.txt
python scripts/translate.py --list-langs<lang> is a short BCP-47-style code: es, fr, ja, zh, ar,
sw, qu, haw, etc. MADLAD covers ~419 languages — --list-langs
prints a curated reference of common ones, but the canonical full list
lives at https://huggingface.co/google/madlad400-3b-mt#languages.
The model auto-detects the source language; you only specify the target. Mixed-language inputs work but the model will translate the whole thing into the target.
scripts/bootstrap.pypython scripts/bootstrap.py --check # is the 3B model installed?
python scripts/bootstrap.py --install # download MADLAD-400-3B-MT (~3 GB)
python scripts/bootstrap.py --install-7b # opt-in: also fetch the 7B variant (~7 GB)
python scripts/bootstrap.py --status # what's installed and how big
python scripts/bootstrap.py --where # print install root
python scripts/bootstrap.py --uninstall # remove the 3B install (asks first)Models go to ~/.local/share/translator/. Override with the
TRANSLATOR_HOME env var if the user wants them somewhere else (e.g.
on an external SSD).
bootstrap.py --install-7b) — same family,
meaningfully betterrness/knowledge/rosetta-primers/ — bundled Swadesh lists,
grammar sketches, and parallel Genesis translations from the
Long Now Rosetta Project, useful as a sanity check or as
few-shot grounding when MADLAD is shakyMADLAD is a sentence-to-paragraph model, not a long-document model. Very
long single inputs (>2000 tokens) may degrade. translate.py handles this
by splitting on paragraph breaks (\n\n) and translating each chunk
independently — so a 50-paragraph document works fine. If the user has
one paragraph that's 3000 words long, that's where you'd want to split
manually first.
This skill is invoked directly by you via the shell tool. It does NOT call other skills, and is NOT called by other
skills as a library. The contract is:
translate.py with the right --target.The skill should NEVER be invoked through the LLM — translate.py is a direct ct2 call. Sending translation work to llama-server is a category error: the LLM is the router, MADLAD is the translator.
This skill is a deliberate implementation of enough's "no cloud dependency, no SDK lock-in" principle for a high-stakes capability:
defaults/skills/translator/ into
any RNESS install and it works after pip install -r requirements.txt && python bootstrap.py --install.~/.local/share/translator/ (the model
cache) and the project's own session logs (your choice).If the user wants higher-quality translation for low-resource languages
and is willing to accept a non-commercial license for personal use, the
NLLB-200 paradigm (gated behind a flag in
rness/paradigms/translation.md) provides that path — explicitly,
with the license tradeoff visible.
enough-tooltip-text: "use the translator skill and the translation paradigm to translate any text into another language."
0gsd/enough
3cf8a60
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.