تخطي إلى المحتوى الرئيسي

i18n-rosetta for Enterprise

Your team translates content regularly. You have a stack of locale files, a CI pipeline, and a process that probably involves someone manually running Google Translate, copying results into JSON, and hoping for the best. Or you're paying for a TMS platform where you're locked into one vendor's translation engine.

There's a better way.

The Pitch​

  1. Pick the best method for each language — not whatever your vendor defaults to
  2. Deploy with one command — npx i18n-rosetta sync translates every locale, every format, every time
  3. Swap methods without changing code — a config change, not a migration
  4. Own your pipeline — no vendor lock-in, no monthly dashboards, no accounts
i18n-rosetta.config.json
{
  "version": 3,
  "pairs": {
    "en:fr": { "method": "deepl" },
    "en:ja": { "method": "llm", "model": "google/gemini-2.5-pro" },
    "en:de": { "method": "google-translate" },
    "en:ko": { "method": "llm", "register": "polite-haeyo" },
    "en:crk": { "methodPlugin": "crk-coached-v3" }
  }
}

French gets DeepL (your team prefers its European fluency). Japanese gets a frontier LLM. German gets Google Translate (fast, cheap, good enough). Korean gets an LLM with a formal register. Plains Cree gets a community-built coached plugin that scored highest on the leaderboard.

Same command. Same CI pipeline. Different methods per pair. One config file.

The Leaderboard → Deploy Workflow​

:::tip Coming soon: rosetta leaderboard CLI The workflow described below is the planned integration between the MT Eval Arena leaderboard and the i18n-rosetta CLI. The infrastructure exists on both sides — the bridge is in development. :::

The MT Eval Arena is where translation methods are benchmarked with reproducible, fingerprinted scoring. Every method gets a composite score across multiple metrics (chrF++, exact match, FST acceptance, semantic scoring). The leaderboard tracks every submission.

The planned workflow:

# Browse the leaderboard from your terminal
npx i18n-rosetta leaderboard --pair en:crk

# Output:
# ┌──────┬───────────────────────┬────────────┬──────────┬───────────┐
# │ Rank │ Method                │ Model      │ chrF++   │ Composite │
# ├──────┼───────────────────────┼────────────┼──────────┼───────────┤
# │  1   │ crk-coached-v3        │ gemini-2.5 │ 43.2     │ 0.67      │
# │  2   │ fst-gated-pipeline    │ gpt-4o     │ 41.8     │ 0.63      │
# │  3   │ prompt-baseline       │ claude-4   │ 38.1     │ 0.55      │
# └──────┴───────────────────────┴────────────┴──────────┴───────────┘

# Install the top-scoring method as a plugin
npx i18n-rosetta plugin install crk-coached-v3

# Use it
npx i18n-rosetta sync

You don't build the method. You don't train the model. You pick the winner and deploy it. If a better method appears on the leaderboard next month, you swap it with one command.

What's Available Today​

The leaderboard-to-CLI bridge is in development. Here's what works right now:

Built-in methods (no plugins needed)​

MethodBest ForCost
llm (default)Quality-focused, any languagePer-token via OpenRouter
geminiQuality + free tierFree (limited), then per-token
google-translateSpeed + volume$20/M characters
deeplEuropean languages$25/M characters
llm-coachedLanguages with coaching dataPer-token via OpenRouter
apiCustom/community-hosted methodsSelf-hosted

Plugin methods (install separately)​

Custom plugins can wrap any translation logic — a fine-tuned model, an FST-gated pipeline, a community API, or anything else that produces JSON. See Build a Plugin.

Enterprise Workflow​

1. Evaluate your current quality​

# See what you're getting today
npx i18n-rosetta status

# Output shows: method per pair, cache hit rate, quality gate stats

2. Run the eval harness on candidates​

The eval harness lets you benchmark multiple methods against the same dataset. Run a sweep, compare scores, pick winners:

# In the eval harness repo
python -m mt_eval_harness.run \
  --methods coached-v3 baseline prompt-tuned \
  --dataset data/your-corpus.json

3. Configure per-pair winners​

Update your config to use the best method per language pair. Different languages have different best methods — that's the point.

4. Integrate into CI/CD​

# In your CI pipeline
npx i18n-rosetta lint        # Catch hardcoded strings
npx i18n-rosetta sync        # Translate what changed
npx i18n-rosetta audit       # Fail if any locale is incomplete
npx i18n-rosetta integrity   # Validate placeholder consistency

Three commands. Zero manual translation. The pipeline catches hardcoded strings, translates them with your chosen methods, and fails the build if anything is missing or corrupted.

5. Professional review (optional)​

For high-stakes content, export to XLIFF for human review:

npx i18n-rosetta xliff export --locale ja --output translations.xliff
# → Send to your translation agency
# → Import corrections back:
npx i18n-rosetta xliff import translations.xliff

Machine-translate the bulk. Human-review the critical paths. Pay for human time only where it matters.

Cost Model​

rosetta has no license fee, no monthly subscription, no per-seat pricing. It's an open-source CLI tool. You pay only for the translation API calls:

VolumeGoogle TranslateLLM (Gemini Flash)LLM (GPT-4o)
1,000 keys × 5 locales~$0.50~$0.30 (free tier)~$2.00
10,000 keys × 15 locales~$15~$8~$60
50,000 keys × 30 locales~$75~$40~$300

Translation Memory means you only pay for changed keys on subsequent syncs. If you update 10 strings out of 10,000, you pay for 10 translations, not 10,000.

vs. TMS Platforms​

rosettaCrowdin / Phrase / Locize
PricingFree (open source) + API costs$50–$500/month + per-seat
Vendor lock-inNone — switch providers in configHigh — data in their cloud
Method choiceAny provider, any model, per pairWhatever they offer
CI/CDFirst-class (lint → sync → audit)Plugin/webhook
Custom methodsPlugin system, community pluginsNot supported
Quality gateBuilt-in (wrong-script, echo, length)Varies
Self-hostedYes (LibreTranslate, custom API)No

See the full comparison for details.

Further Reading​