Skip to content
Dextalo
Back to leaderboard
Benchmark · Dextalo

Benchmark Methodology

How we evaluate LLM translation quality across Asian & ASEAN languages — a multi-judge, MQM-weighted panel over a fixed corpus.

01 Approach

Overview

The Dextalo Translation Benchmark systematically evaluates large language models on their ability to translate English text into Asian and ASEAN languages. Rather than relying on reference-based metrics (BLEU, COMET) that penalize valid alternative phrasings, we use a multi-judge LLM-as-Judge panel to score translations on four quality dimensions.

02 What we test

Test Corpus

The benchmark uses a curated corpus of 80+ English source sentences organized into eight domain categories:

Business

Corporate communications, financial reports, formal correspondence

Casual

Everyday conversation, informal tone, colloquialisms

Technical

Software documentation, IT infrastructure, APIs

Legal

Contracts, compliance, regulatory language

Marketing

Promotional copy, brand voice, persuasive content

Medical

Clinical documentation, pharmaceutical, patient care

Idiomatic

Figurative language, proverbs, cultural expressions

E-Commerce

Product descriptions, reviews, shipping policies

Each domain tests different translation skills: terminology precision, register matching, cultural adaptation, and technical accuracy.

03 Direction

Language Pairs

Models are tested on four English-to-target language pairs:

TH English → Thai
JA English → Japanese
ZH English → Chinese (Simplified)
VI English → Vietnamese

These languages were chosen to represent diverse writing systems, grammatical structures, and levels of LLM training-data availability.

04 Pipeline

Translation Process

Each model translates all sentences through the OpenRouter API using domain-specific system prompts. Sentences are batched by domain to provide context, and all models receive identical prompts to ensure a fair comparison.

Performance metrics — time-to-first-token, tokens per second, and cost — are measured via streaming responses and averaged across three runs per model.

05 LLM-as-Judge

Scoring Panel

Each translation is independently scored by three judge models:

Claude Haiku 4.6 Anthropic
Qwen 3.5 Flash Qwen
Gemini 3.1 Flash Lite Google

Using multiple judges from different providers reduces single-model bias. Final scores use median aggregation across the three judges, which is more robust to outliers than averaging.

06 Rubric

MQM Quality Dimensions

Judges score each translation on four dimensions from the Multidimensional Quality Metrics (MQM) framework, each on a 0–100 scale. The composite score is their weighted average, used for overall ranking:

Dimension Weight What it measures
Accuracy 0.40 How faithfully the translation conveys the meaning of the source text. Measures semantic equivalence.
Fluency 0.30 How natural and grammatically correct the translation reads in the target language.
Terminology 0.20 How appropriate and consistent domain-specific terms are in the translation.
Style 0.10 How well the translation matches the expected register, tone, and style for the domain.
composite = 0.40·Accuracy + 0.30·Fluency + 0.20·Terminology + 0.10·Style
07 Speed & cost

Performance Metrics

TTFT Time to First Token — latency from request to first response token, measured via streaming
Tokens/sec Output generation speed
Cost/1K words Normalized cost using OpenRouter pricing

Performance metrics are averaged across three runs to reduce variance from network conditions.

08 Caveats

Limitations

Sources

References

Kocmi & Federmann (2023). "GEMBA-MQM: Detecting Translation Quality Error Spans with GPT-4." Microsoft Research.
MQM Framework — themqm.org
OpenRouter API — openrouter.ai/docs