---
license: other
language:
- en
- es
- eu
size_categories:
- 1K<n<10K
task_categories:
- translation
- text-generation
pretty_name: CRITICS Science MT Benchmark (critics-SciMTBench)
tags:
- scientific-machine-translation
- machine-translation
- scientific-text
- document-translation
- multilingual
- english
- spanish
- basque
- euskara
- low-resource-languages
- scientific-terminology
- machine-translation-evaluation
- benchmark
- evaluation-only
- arxiv
- chemrxiv
- critics
- chist-era
- science-in-your-own-language
gated: true
extra_gated_heading: Protecting the integrity of critics-SciMTBench for evaluation
extra_gated_description: Please accept the evaluation-only conditions before accessing benchmark content. Article-level redistribution rights are under review. Contact the maintainers for permitted uses.
extra_gated_fields:
  I agree to use critics-SciMTBench solely for evaluating MT systems and not for training, fine-tuning, adaptation, or synthetic training-data generation: checkbox
  I agree not to publish, mirror, or transfer the dataset or derivatives in locations accessible to automated crawling or model training: checkbox
  I agree to retain these conditions with any permitted distribution and to prevent evaluation data from entering training corpora: checkbox
configs:
- config_name: en-es
  data_files:
  - split: test
    path: data/pairs/en-es/test.jsonl
- config_name: en-eu
  data_files:
  - split: test
    path: data/pairs/en-eu/test.jsonl
---

<p align="center">
  <img src="assets/critics-logo.jpg" alt="CRITICS — Critical Science Without Borders project logo" width="260" />
</p>

# CRITICS Science MT Benchmark (critics-SciMTBench)

**A multilingual benchmark for document-level evaluation of machine translation of scientific text into Spanish and Basque.** Developed in the context of **CRITICS - Critical Science Without Borders**, a project on LLM-based access to scientific knowledge across linguistic boundaries.

> **Evaluation-only release.** The benchmark must not be used for model training, fine-tuning, adaptation, or synthetic training-data generation. Access controls and the rights status of underlying source material require independent attention; see [Access, permitted use, and rights](#access-permitted-use-and-rights).

| Attribute | Current benchmark release |
|---|---|
| Source language | English (`en`) |
| Target languages | Spanish (`es`); Basque (`eu`) |
| Main task | Scientific machine translation evaluation at document level|
| Source domains | Scientific excerpts originating from arXiv and ChemRxiv |
| Available configurations | `en-es`, `en-eu` |
| Available splits | `test` only for each configuration |
| Included evaluation material | Source excerpts and corresponding MT outputs |
| Human-post-edited references | Not currently provided; reference fields are `null` |
| Intended uses | Evaluation and comparative analysis of MT systems, subject to access and rights conditions |

## 1. Dataset overview

**critics-SciMTBench** is a working evaluation resource for **scientific machine translation at document level**. Currently it contains English scientific excerpts and corresponding machine-translated texts in Spanish and Basque, with provision for human-post-edited references when independently supplied and approved. At the present stage, human references are `null`; this is **not** a reference-based benchmark release.

The resource is intended to support research on scientific translation in multilingual contexts, particularly translation quality, terminology, and reliable assessment of scientific content. The broader CRITICS project also investigates document-level translation and multidimensional evaluation methods. These are *project research objectives*, not measured capabilities or validated results of the present dataset.

### Scope and limitations

- The published dataset configurations are **English–Spanish** and **English–Basque**; neither additional language pairs nor reference translations should be inferred from the project-level language agenda.
- Both configurations expose a `test` split; there are no `train` or `validation` partitions available.
- No automatic quality scores, human agreement statistics, linguistic-domain breakdowns, or benchmark results are coming soon.

## 2. Loading the dataset

```bash
pip install -U datasets huggingface_hub
hf auth login
```

```python
from datasets import load_dataset

repo = "HiTZ/critics_science_mt_benchmark"

# English -> Basque evaluation data
en_eu = load_dataset(repo, "en-eu", split="test", token=True)

# English -> Spanish evaluation data
en_es = load_dataset(repo, "en-es", split="test", token=True)

print(en_eu.column_names)
print(len(en_eu), len(en_es))
```

To inspect the English source for the first record (using the field name described in the original card):

```python
print(en_eu[0]["source_text"])
```
**Do not combine the two language-pair configurations as if they were independent source documents.** Related records may originate from the same English excerpts; preserve sample and batch IDs when carrying out cross-language comparisons or computing uncertainty estimates.

### Supplementary files

The repository's configured `data_files` load only the language-pair test JSONL files. Additional available files can be retrieved separately, subject to repository access and redistribution restrictions:

```python
from huggingface_hub import snapshot_download

local_repo = snapshot_download(
    repo_id="HiTZ/critics_science_mt_benchmark",
    repo_type="dataset",
    token=True,
)
print(local_repo)
```

The directories below describe the **working repository organization**:

```text
critics_science_mt_benchmark/
├── README.md
├── assets/
│   └── critics-logo.jpg
├── data/
│   ├── pairs/
│   │   ├── en-es/test.jsonl       # Config: en-es
│   │   └── en-eu/test.jsonl       # Config: en-eu
│   ├── canonical/                # canonical files generated from the original input (batches)
│   └── reports/                  # Structural ingestion reports
├── original_input/              # Original files from 10 batches
└── post_edits/                  # human validated reference translations
```

## 3. Data organization and provenance

The dataset is organized by **translation direction**, with two language-pair JSONL evaluation files:

| Hugging Face configuration | Source | Target | Split | File |
|---|---|---|---|---|
| `en-es` | English | Spanish | `test` | `data/pairs/en-es/test.jsonl` |
| `en-eu` | English | Basque | `test` | `data/pairs/en-eu/test.jsonl` |

The source material is extracted from **arXiv and ChemRxiv**. The working pipeline distinguishes original input batches, a canonical multilingual representation, and pairwise evaluation exports for training and inference with language models. The original files are intended to be preserved unchanged; the canonical representation is an authoritative *derived* internal format.

### References and scoring

Human-post-edited reference translations are currently **work in progress** (`null` in the supplied description). Therefore:

- Available MT outputs should be used as gold-standard or human-validated reference translations.
- Reference-based quality metrics should not be reported as if verified human references were present.

## 4. CRITICS project

**CRITICS - Critical Science Without Borders** (*LLMs for Translation of Scientific Knowledge in Multilingual Contexts*) is a research project funded under **CHIST-ERA IV Cofund 2025**, within the **Science in your Own Language (SOL)** topic. Its aim is to improve multilingual access to scientific knowledge while supporting scientific literacy, evidence-based argumentation, and critical thinking.

| Project detail | Information |
|---|---|
| Project name | Critical Science Without Borders (CRITICS) |
| Programme | CHIST-ERA IV Cofund 2025 - Science in your Own Language (SOL) |
| Project period | 15 December 2025 - 14 December 2028 (36 months) |
| **Project coordinator**| **Rodrigo Agerri** HiTZ Center - Ixa, University of the Basque Country EHU|
| Project website | [hitz.eus/critics](https://hitz.eus/critics) |

### CRITICS Consortium

| Partner | Country | Role in CRITICS |
|---|---|---|
| **University of the Basque Country (EHU), HiTZ Center** | Spain | Project coordination, data, and multilingual NLP (WP1–WP2) |
| **Vytautas Magnus University (VMU)** | Lithuania | Scientific-document machine translation (WP3) |
| **CNRS and UCA** | France | Argument mining and assessment (WP4) |
| **Zurich University of Applied Sciences (ZHAW)** | Switzerland | Evaluation and applications (WP5–WP6) |

The project states four connected objectives:

1. **LLM-based translation of scientific documents**, emphasizing document-level fidelity and specialized terminology in low- and medium-resource language settings.
2. **Argument mining and misconception detection**, focusing on evidence-based reasoning and scientific discourse.
3. **Automatic assessment and critical-thinking support**, including critical questions, feedback, and consistent evaluation across languages.
4. **Evaluation and benchmarking**, investigating methods beyond traditional reference-overlap metrics, including multidimensional LLM-as-a-judge methods and human comparison.

**How this dataset contributes:** critics-SciMTBench provides scientific-text evaluation material for the translation and benchmarking of MT at document level. It is complementary to, but distinct from, the [CRITICS Science Assessments dataset](https://huggingface.co/datasets/HiTZ/critics_science_assessments), which focuses on student responses with respect to data-induced assessment rubrics.

For more information: [CRITICS project website](https://hitz.eus/critics) · [HiTZ Center](https://www.hitz.eus/)

## 5. Intended use

**Permitted goal (subject to repository access conditions and underlying rights):** evaluate and compare scientific machine translation systems at document level, including careful analysis of terminology, semantic preservation, and cross-language behavior.

**Not permitted under the stated benchmark-use conditions:** use for training, fine-tuning, model adaptation, or synthetic training-data generation; republish, mirror, or transfer benchmark texts or derivatives to publicly crawlable repositories or training corpora.

Evaluation users should report the language-pair configuration, dataset revision, batch/sample identifiers where relevant, model and decoding configuration, evaluation prompts or scoring protocol, and any human review procedures.

## 6. Access, permitted use, and rights

The benchmark is intended to be distributed as **gated, evaluation-only material**. The YAML `gated: true` field and acceptance checkboxes express the intended conditions, but **repository maintainers must enable and configure access requests in Hugging Face Settings**; metadata alone does not establish access restrictions. Manual approval may be appropriate during initial evaluation.

**Source and derivative rights remain under review.** arXiv and ChemRxiv host materials under potentially different article-level terms. Online availability does not itself authorize redistribution, translation, or relicensing. Before releasing any scientific excerpt or derivative, check the specific work's license, permissions, and any third-party content.

**License metadata note:** Remains pending to select and document appropriate licensing or access terms once the source-rights review (arXiv and ChemRxiv) is complete.

## 7. Citation

If you use this benchmark, cite the **CRITICS project**, a published paper coming soon.

- **Dataset:** [HiTZ/critics_science_mt_benchmark](https://huggingface.co/datasets/HiTZ/critics_science_mt_benchmark)
- **Project:** [CRITICS - Critical Science Without Borders](https://hitz.eus/critics)

## Contact

rodrigo.agerri@ehu.eus and mariana.flores@ehu.eus