README.md

---
license: cc-by-4.0
language:
  - en
  - fr
  - id
  - uk
  - km
  - jv
task_categories:
  - translation
tags:
  - machine-translation
  - instruction-following
  - llm-as-judge
extra_gated_heading: Protecting the integrity of these evaluation benchmarks
extra_gated_description: >-
  This dataset derives from BOUQuET and WMT24++, both translation evaluation
  benchmarks. It is gated so that its contents are not picked up by web crawlers
  and absorbed into language-model training data. The terms below are BOUQuET's
  own, retained here as its licence requires. One scope note on the first term:
  it does not extend to datasets/smol/, which derives from SMOL rather than from
  a test set and is the training data for the Appendix I experiment.
extra_gated_fields:
  This data is for evaluation purposes only; You may not use any of this data or its derivatives for training machine learning / AI models: checkbox
  You may only distribute, embed, or otherwise transfer this data or its derivatives via a mechanism that is either private or that implements protections against automated crawling (such as using a password-protected archive or a gating mechanism that requires users to accept these terms before accessing the dataset): checkbox
  Your distributions must retain these terms: checkbox
---

# Tailoring MT to Audience and Intent: data release

Data accompanying *"Beyond 'To whom it may concern': Tailoring Machine Translation to
Audience and Intent"* (EMNLP 2026).

- **Paper:** https://arxiv.org/abs/2606.03259
- **Code:** https://github.com/raphaelmerx/purpose-mt

## Why this is gated

The user instructions here are generated from BOUQuET and WMT24++ metadata and embed
source text from both benchmarks. BOUQuET is distributed under terms that forbid
re-hosting it where web crawlers can reach it, and ask users to keep its contents out of
training data. WMT24++ is Apache-2.0 and could legally be posted openly, but it is also a
test set, so we apply the same protection to it.

This is why the data lives here rather than in the GitHub repository, and why the three
gate terms above are BOUQuET's own, carried over verbatim.

One scope note: `datasets/smol/` derives from [SMOL](https://huggingface.co/datasets/google/smol)
(CC-BY-4.0), not from a test set. It is the training data for the Appendix I distillation
experiment, so the no-training term does not extend to it.

## Contents

Paths mirror the layout the code expects, so `python -m data.download` in the GitHub
repository places every file where the scripts look for it.

### `datasets/`: generated user instructions

| File | Rows | Paper |
|---|---|---|
| `bouquet_instructions_dev.jsonl` | 504 | §3.2, main results |
| `bouquet_instructions_test.jsonl` | 854 | held-out split, few-shot retrieval pool |
| `bouquet_instructions_dev_context.jsonl` | 504 | Appendix G, context-only ablation |
| `bouquet_instructions_dev_purpose.jsonl` | 504 | Appendix G, purpose-only ablation |
| `bouquet_instructions_dev_self-para-gemma-3-27b-it.jsonl` | 504 | §5, self-instruction |
| `bouquet_instructions_dev_self-para-gemma-4-31b-it.jsonl` | 504 | §5, self-instruction |
| `bouquet_instructions_test_self-para-gemma-3-27b-it.jsonl` | 854 | §5, self-instruction |
| `wmt24pp_instructions.jsonl` | 997 | Appendix C |
| `wmt24pp_instructions_self-para.jsonl` | 997 | Appendix C, self-instruction |
| `smol/en_sources.jsonl` | 7,815 | Appendix I, SMOL sources |
| `smol/en_instructions.jsonl` | 7,815 | Appendix I, instructions for distillation |

BOUQuET instruction schema: `uniq_id`, `tgt_text` (English source), `domain`,
`par_comment`, `tags`, `register`, `user_instruction`. Every field except
`user_instruction` comes from BOUQuET; `user_instruction` is drafted by Gemini-3-Flash
from that metadata and then revised by a human annotator.

### `annotation_exports/`: human and LLM-judge evaluations

`human/` holds 160 rated items per language for French, Indonesian, Ukrainian, Khmer and
Javanese: error spans, a 0-100 ESA translation rating, and a 0-100 adaptedness score.
These are the five annotation projects reported in the paper; earlier pilot projects are
not included.

`llm/` holds the LLM-judge scores over the same items, produced by
`human_eval.match_judge`, which re-judges the exported text itself rather than joining on
`uniq_id`. The `_refbased` files are the reference-based judge run behind the
reference-free vs reference-based comparison in §3.4.

`comet/` holds XCOMET-XL scores computed the same way, on exactly the translations the
annotators rated (`human_eval.match_comet`). This matters: joining human ratings to the
full test-set metric files on `uniq_id` pairs 38% of rated rows with a score computed on
a different translation, because the test-set translations were regenerated after the
annotations were collected. `human_eval.human_metric_correlation --matched` uses these
files and recomputes ChrF++ from the export, so judge and metric are measured on
identical text. **Coverage:** French covers 100 of its 160 rated rows; the other four
languages are complete.

Annotators are identified only by an integer `annotator_id`. Free-text
`annotator_comment` fields are linguistic notes and contain no personal data.

### `scores/`: aggregated results

The per-condition means behind every table and figure, produced by
`python -m analysis.aggregate`. No benchmark text; these are the numbers the paper
reports. See the GitHub README for the column definitions.

### `translation_results/controlled_mt/`: controlled-MT scores

Aggregate CoCoA-MT and MT-GenEval scores (Appendix K): M-Acc, coverage, commit rate,
gender accuracy. No source text.

## Not included

Item-level model outputs and judge evaluations (~770 MB) are not released: they embed
large amounts of benchmark source text, and the aggregates in `scores/` carry every
number the paper reports. They are reproducible from the pipeline in the GitHub
repository.

## Licence and attribution

Our contributions, meaning the generated instructions, the human annotations and the
aggregated scores, are released under **CC-BY-4.0**.

Derived from, and subject to the terms of, the following:

- **BOUQuET** ([facebook/bouquet](https://huggingface.co/datasets/facebook/bouquet)), CC-BY-4.0, gated
- **WMT24++** ([google/wmt24pp](https://huggingface.co/datasets/google/wmt24pp)), Apache-2.0
- **SMOL** ([google/smol](https://huggingface.co/datasets/google/smol)), CC-BY-4.0
- **CoCoA-MT**, CDLA-Sharing-1.0
- **MT-GenEval**, CC-BY-SA-3.0

## Citation

```bibtex
@inproceedings{merx2026beyond,
  title     = {Beyond ``To whom it may concern'': Tailoring Machine Translation to Audience and Intent},
  author    = {Merx, Raphael and Vylomova, Ekaterina and Cohn, Trevor},
  booktitle = {Proceedings of EMNLP 2026},
  year      = {2026}
}
```