Building a Bilingual Operations Adapter with Limited Data

Balance two languages without relying on naive duplication. The useful result is a method that preserves evidence, exposes constraints, and makes the next decision reviewable. This note describes an engineering workflow rather than a claim about a particular organization or production event.

Establish the boundary

Define supported tasks and language-pair boundaries. Keep the relevant source version, configuration, and stable identifier together while this work is performed. That gives a reviewer a direct path from an observed outcome to the input and rule that produced it. Treat automatic checks and human judgment as separate controls, then record both in the build note.

Use a fixed holdout or review sample for this step. Include ordinary records, rare boundary cases, and at least one input designed to fail safely. A useful change improves the intended condition without removing evidence about a known limitation.

Measure topic and intent coverage separately by language

Measure topic and intent coverage separately by language. Keep the relevant source version, configuration, and stable identifier together while this work is performed. That gives a reviewer a direct path from an observed outcome to the input and rule that produced it. Treat automatic checks and human judgment as separate controls, then record both in the build note.

Parallel examples only where translation preserves operational meaning

Create parallel examples only where translation preserves operational meaning. Keep the relevant source version, configuration, and stable identifier together while this work is performed. That gives a reviewer a direct path from an observed outcome to the input and rule that produced it. Treat automatic checks and human judgment as separate controls, then record both in the build note.

Retain native-language examples and terminology lists

Retain native-language examples and terminology lists. Keep the relevant source version, configuration, and stable identifier together while this work is performed. That gives a reviewer a direct path from an observed outcome to the input and rule that produced it. Treat automatic checks and human judgment as separate controls, then record both in the build note.

Evaluate language consistency, factual retention, and accidental switching

Evaluate language consistency, factual retention, and accidental switching. Keep the relevant source version, configuration, and stable identifier together while this work is performed. That gives a reviewer a direct path from an observed outcome to the input and rule that produced it. Treat automatic checks and human judgment as separate controls, then record both in the build note.

When retrieval or terminology injection is preferable to more fine-tuning

Explain when retrieval or terminology injection is preferable to more fine-tuning. Keep the relevant source version, configuration, and stable identifier together while this work is performed. That gives a reviewer a direct path from an observed outcome to the input and rule that produced it. Treat automatic checks and human judgment as separate controls, then record both in the build note.

Practical check

JSONL examples in English plus a second language and a coverage report script. The example below is intentionally small enough to run before a larger build or publication workflow.

import csv
import json
from pathlib import Path

with Path("labels.csv").open(encoding="utf-8", newline="") as handle:
    rows = list(csv.DictReader(handle))

with Path("train.jsonl").open("w", encoding="utf-8") as handle:
    for row in rows:
        handle.write(json.dumps(row, ensure_ascii=False) + "\n")

print(f"wrote {len(rows)} records")

Evidence for the next decision

Keep the input digest, outcome, and reviewer note with the resulting artifact. Claiming bilingual quality from token-count balance alone. When evidence is incomplete, preserve the failing example, define the missing validation, and defer publication until results can be compared fairly.

Make the process idempotent. Repeating the same step on unchanged input should not append metadata, discard extra detail, or silently change ownership. Idempotence makes retries safe and exposes hidden state when outputs differ.

Review results by failure category rather than a single blended number. A small number of high-risk failures can outweigh an improvement on common cases, particularly when output feeds a release workflow.

Keep accepted and rejected examples together. A concise rejected sample with its expected outcome is a durable regression test and prevents the same edge condition from returning as an unexplained surprise.

Use a separate holdout set for release decisions. Training records can guide implementation, but they cannot demonstrate correct behaviour on new wording, new sources, or changed input order.