Translation and Audit Pipeline

English | 简体中文

The scripts in this directory preserve the Phase 1 Chinese-to-English data translation, validation, repair, and human adjudication workflow. API keys are read only from the environment; do not add them to configuration files.

Configuration

Copy translation_api_config.example.json to translation_api_config.json, then set the environment variable named by api.api_key_env. The configuration supports OpenAI-compatible Chat Completions endpoints, custom base URLs, headers, model names, concurrency, token fields, and additional top-level request body fields through extra_body.

cp translation_api_config.example.json translation_api_config.json
export TRANSLATION_API_KEY="your-key"

--no-api runs the dictionary and record-processing stages without loading an API configuration file or requiring an API key.

The local configuration is ignored by Git. It can independently override API parameters for translation, audit, repair, and thinking-enabled re-audit.

Workflow

1. Build the DSL dictionary and translated query records:

   python scripts/build_translation_assets.py
   

2. Run deterministic checks and the configured LLM audit:

   python scripts/audit_phase1_translations.py
   

3. Repair only records selected by the audit and verify the repairs:

   python scripts/repair_phase1_translations.py
   

4. Re-audit rejected candidates conservatively, or export them for review:

   python scripts/reaudit_phase1_translations.py
   python scripts/export_phase1_translation_review.py
   

5. Apply the reviewed decisions to produce a final data directory:

   python scripts/apply_phase1_human_adjudication.py
   

repair_translated_dsl.py is a separate deterministic utility for rebuilding syntax-invalid translated DSL expressions from the canonical Chinese source. All scripts accept path overrides; run a script with --help for details.

Conservative Change Policy

The pipeline does not rewrite every record flagged by one LLM call. Repair candidates must match the configured statuses and verification policy. The two-stage re-audit keeps uncertain cases as manual_review, and the human adjudication step records the final set of changes. This preserves already correct translations whenever possible.