Dataset Workflows

Translation

Translate generated datasets with the offline local-pseudo path or explicit provider-backed OpenAI and Anthropic adapters.

Status

Translation is experimental. The public workflow is stable enough to run, but you should treat provider strategy choices, prompts, and locale coverage as something to evaluate on your own datasets before production use.

Offline default

The default strategy is local-pseudo, which keeps the workflow offline:

node dist/cli/index.js translate-dataset outputs/receptionist-sample.jsonl \
  --target-locale es-ES \
  --out outputs/receptionist-sample.es-ES.jsonl

This path transforms translatable message text locally and does not require any provider credentials.

Provider-backed translation

OpenAI and Anthropic translation strategies are implemented behind the shared provider adapter boundary. They require:

an explicit strategy: openai or anthropic
an explicit model via --translation-model
an API key available through the default env var or --translation-api-key-env

OpenAI example:

OPENAI_API_KEY=... node dist/cli/index.js translate-dataset outputs/receptionist-sample.jsonl \
  --strategy openai \
  --translation-model <model> \
  --target-locale es-ES \
  --out outputs/receptionist-sample.es-ES.jsonl

Anthropic example:

ANTHROPIC_API_KEY=... node dist/cli/index.js translate-dataset outputs/receptionist-sample.jsonl \
  --strategy anthropic \
  --translation-model <model> \
  --target-locale fr-CA \
  --out outputs/receptionist-sample.fr-CA.jsonl

Use --translation-api-key-env <ENV_NAME> when the key is stored in a non-default environment variable.

Preservation rules

Translation preserves the schema-bearing parts of each row:

system, user, and assistant text content are translated
assistant tool_calls are preserved exactly
tool result messages are preserved exactly
tool definitions are preserved exactly
metadata is preserved and extended with sourceLocale when known, targetLocale, translationStatus, translationProvider, translationRequestPath, and translationModel when relevant

Target locales must be valid BCP 47 identifiers such as es-ES, fr-CA, or hi-IN.

Failure behavior

Provider-backed translation is intentionally strict:

missing model selection fails before provider calls
missing API key env vars fail before provider calls
empty output for a non-empty source field is rejected
translated rows are validated before being written to disk

That makes it easier to keep translation mistakes from silently leaking into training data.