---
pretty_name: "ORBIT-MT-Eval: Expert Annotations and Prompts for Machine Translation Meta-Evaluation"
language:
- en
- ru
license: apache-2.0
task_categories:
- translation
tags:
- machine-translation
- evaluation
- error-analysis
- rate
size_categories:
- n<10K
configs:
- config_name: rate_annotations
  data_files:
  - split: test
    path: data/rate_annotations/test.jsonl
---

# ORBIT-MT-Eval: Expert Annotations and Prompts for Machine Translation Meta-Evaluation

ORBIT-MT-Eval is an English-to-Russian reference dataset for machine translation meta-evaluation: comparing evaluator outputs against professional expert annotations. Each of the 600 source–translation segments has three independent expert annotations following the [RATE protocol](https://aclanthology.org/2025.findings-emnlp.1203/). In *Beyond Prompts: A Systematic Study of LLM-Based Machine Translation Evaluators*, two annotations serve as gold references and the third is scored against each reference and averaged to produce the Human_E agreement baseline. The release also includes the prompts used by ORBIT and the locally maintained prompting baselines. GEMBA and `unified-mqm-boosted-v5` are linked to their official upstream implementations.

## Dataset at a glance

| Property | Value |
|---|---|
| Language direction | English → Russian |
| Source segments | 300 |
| Source–translation segments | 600 (two translations for each source) |
| Professional expert annotators | 35 |
| Expert annotations | 3 independent annotations per segment from the expert pool, for 1,800 records total |
| Meta-evaluation protocol | 2 reference annotations + 1 Human_E agreement baseline |
| Error spans | 9,771 |
| Annotation scheme | RATE severity 1–5 with separate Accuracy and Fluency scores |
| Domains | Literary, news, social media, and speech |

`expert_1`, `expert_2`, and `expert_3` identify the three annotation slots within each segment. They are not person-level IDs and do not track individual annotators across the dataset.

## Files

- [`data/rate_annotations/test.jsonl`](data/rate_annotations/test.jsonl): the annotation records.
- [`prompts/README.md`](prompts/README.md): readable prompt pages with copy-ready JSON `messages` objects.
- [`prompts/prompts.jsonl`](prompts/prompts.jsonl): the machine-readable prompt collection.

## Load

```python
from datasets import load_dataset

dataset = load_dataset("foksly/orbit-mt-eval", "rate_annotations", split="test")
print(dataset[0])
```

## Paper results

| Evaluator | R | P | F1 | R4+ | P4+ | MQM |
|---|---:|---:|---:|---:|---:|---:|
| Human_E | 46.8 | 44.3 | 45.5 | 65.3 | 65.3 | 75.3 |
| Human_A-SbS | 39.3 | 47.1 | 42.8 | 55.7 | 63.2 | 68.0 |
| Human_A-PW | 31.8 | 46.4 | 37.7 | 49.8 | 57.1 | 65.0 |
| Human_WMT | 12.3 | 61.9 | 20.6 | 26.6 | 62.8 | 42.7 |
| GEMBA-MQM | 24.6 | 43.0 | 31.3 | 45.0 | 51.0 | 58.7 |
| GEMBA-ESA | 24.9 | 41.3 | 31.0 | 44.6 | 45.7 | 57.3 |
| MQM-AE | 24.0 | 51.8 | 32.8 | 45.4 | 54.5 | 60.0 |
| ESA-AE | 27.0 | 54.8 | 36.2 | 49.5 | 58.6 | 53.7 |
| UMB-v5 | 27.8 | 44.5 | 34.2 | 48.7 | 56.4 | 57.3 |
| xCOMET-XXL | 18.3 | 36.4 | 24.3 | 33.5 | 36.8 | 58.3 |
| ORBIT-SC | 51.5 | 46.2 | 48.7 | 74.7 | 63.5 | 67.0 |
| ORBIT-MC | **60.3** | 45.8 | **52.1** | **82.0** | **70.6** | **69.3** |

Values are on a 0–100 scale for the English→Russian evaluation set. R/P/F1 are micro-averaged span metrics excluding RATE severity 1, R4+/P4+ use severity ≥4, and MQM is pairwise ranking accuracy.

Prompting baselines and ORBIT-SC use GPT-5.4. ORBIT-MC fuses three evaluator backbones with an LLM fuser.

## Citation

```bibtex
@inproceedings{popov-etal-2026-beyond-prompts,
  title     = {Beyond Prompts: A Systematic Study of LLM-Based Machine Translation Evaluators},
  author    = {Popov, Dmitry and Bokhyan, Roman and Enikeeva, Ekaterina and
               Mekhraliev, Artem and Vysotsky, Stepan and Karpachev, Nikolay},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}
```

## License

The original RATE annotations and prompt material are released under the [Apache License 2.0](LICENSE). Source and translation material retain their applicable upstream terms.