Aller au contenu principal

MT Evaluation

The MT Eval Harness is a language-pair-agnostic recording surface for machine translation experiments. It doesn't judge how you translate — it measures what comes out.

Run any translation approach against a standardized dataset. The harness produces a run card — a self-contained JSON document recording every input, output, metric, and environment detail. Run cards are the unit of comparison: deterministic, reproducible, and auditable.

How It Works​

  1. Pick a dataset — A curated set of source→target pairs with gold-standard references
  2. Run the harness — It sends each entry through your model, collects outputs, and scores them
  3. Get a run card — chrF++, exact match rate, FST acceptance, latency, cost — all in one JSON file
  4. Submit to the leaderboard — Compare your method against others on the same dataset

:::danger DO NOT TRAIN on evaluation data

These datasets exist to measure translation quality, not to improve it. Methods trained, fine-tuned, or few-shot-prompted with evaluation data will produce artificially inflated scores and will be disqualified from the leaderboard.

If you need training data, use separate corpora. The evaluation sets must remain unseen by your model. :::

What Gets Measured​

The harness computes several complementary metrics per run:

MetricWhat it tells you
chrF++Character-level n-gram overlap with the gold standard. Language-agnostic, no tokenizer needed.
Exact match ratePercentage of outputs identical to the gold standard. A high bar — useful for morphologically precise targets.
FST acceptance ratePercentage of outputs accepted by a finite-state transducer analyzer (when available). Measures morphological validity.
Latency (avg / median / p95)How fast your method responds.
Cost (total / per-entry)Token usage and USD cost for API-based methods.

Next Steps​