Psychias/alee_dataset

ALEE Dataset

This repository contains three multilingual machine-translation benchmarks — FLORES-200, BOUQuET and WMT24++ — each extended with English "foil" (negative) columns: minimally edited, meaning-altered versions of the English text, generated through an Abstract Meaning Representation (AMR) round-trip and filtered with a Natural Language Inference (NLI) model so that the foil is not entailment-equivalent to the original. The wmt24pp config is a merge of two datasets — google/wmt24pp (English source + 55 target languages) and ZurichNLP/wmt24pp-rm (six Romansh varieties) — joined on segment_id.

Each source is a separate config (BOUQuET is further split by level, mirroring facebook/bouquet). Original columns are unchanged; only rows with at least one successful foil are kept.

from datasets import load_dataset
ds = load_dataset("Psychias/alee_dataset", "flores200",               split="test")   # 829 rows
ds = load_dataset("Psychias/alee_dataset", "bouquet_sentence_level",  split="test")   # 671 rows
ds = load_dataset("Psychias/alee_dataset", "bouquet_paragraph_level", split="test")   # 193 rows
ds = load_dataset("Psychias/alee_dataset", "wmt24pp",                 split="test")   # 818 rows

Configs

Config	Rows	Level	Derived from
flores200	829	sentence	Muennighoff/flores200 (FLORES-200 / NLLB)
bouquet_sentence_level	671	sentence	facebook/bouquet (sentence_level)
bouquet_paragraph_level	193	paragraph	facebook/bouquet (paragraph_level)
wmt24pp	818	paragraph	google/wmt24pp + ZurichNLP/wmt24pp-rm

Schema

Every config keeps the original columns of its source dataset (unchanged) and adds the foil columns below. A foil cell holds the generated foil text only when that foil passed validation, otherwise it is null.

Added English foil columns (contrastive against the English column of each source — sentence_eng_Latn for flores200/bouquet_*, en_EN for wmt24pp):

• <eng>_RoleSwap_negative: two leaf arguments of the AMR graph are swapped.
• <eng>_PolarityNegation_negative: a :polarity - edge is added to a node.
• <eng>_AntonymRepl_negative: a concept is replaced by a WordNet antonym.
• <eng>_HypernymSub_negative: a concept is replaced by a WordNet hypernym.

The English code <eng> mirrors each source: eng_ for flores200/bouquet_* (e.g. eng_RoleSwap_negative), en_ for wmt24pp (e.g. en_RoleSwap_negative).

wmt24pp additionally contains six Romansh reference columns (ISO 639-3 roh, Latin script), merged from ZurichNLP/wmt24pp-rm on segment_id:

• roh_puter (Puter), roh_rumgr (Rumantsch Grischun), roh_surmiran (Surmiran), roh_sutsilv (Sutsilvan), roh_sursilv (Sursilvan), roh_vallader (Vallader).

Foil generation

For each English text: parse to an AMR graph (amrlib, spaCy en_core_web_sm) → apply one transform on the AMR triples → regenerate text (amrlib AMR-to-text, T5) → validate with juliussteen/DeBERTa-v3-FaithAug. A foil is a success only if it is not bidirectionally entailing and neither direction exceeds 0.8 entailment probability. Sentence-level rows use best-of-all scoring per transform; paragraph-level rows edit one sentence, splice it back, and apply NLI to the full paragraph (greedy first-valid). Deterministic: random.seed(42).

Reproducing

The full pipeline is shipped in this repo as amr_foil_pipeline.py (dependencies in requirements.txt) — a standalone port of the original AMR_generate_datasets.ipynb:

• --stage prep: loads/transforms the original sources into the parallel-CSV format the pipeline consumes. For wmt24pp it merges the six Romansh varieties from ZurichNLP/wmt24pp-rm and remaps de_DE-rm-<variety> → roh_<variety> (extra carry-along columns; the foil step only reads en_EN, so nothing breaks).
• --stage generate: runs the AMR + NLI foil pipeline and writes the same UTF-16 CSV outputs as the original (identical column names and per-transform splits).

python amr_foil_pipeline.py --stage prep     --dataset all
python amr_foil_pipeline.py --stage generate --dataset wmt24
python amr_foil_pipeline.py --stage all      --dataset all --limit 50   # smoke test

Downstream use (embedding pre-computation)

These datasets feed an embedding pre-computation notebook (2--ALEE_PRE-CALCULATE-Embeddings.ipynb) that reads UTF-16 CSVs and expects the original foil-metadata columns, so the exported inputs use that schema (not the publication *_negative schema):

• WMT24++: foil_<t>_text, foil_<t>_status, boolean is_bad_source, and 5-character language codes. Romansh is exported as 5-char rm_* codes (rm_RG, rm_SV, rm_ST, rm_SM, rm_PU, rm_VA) rather than roh_*, because the notebook discovers languages by a 5-character rule — roh_puter would be skipped, rm_* is embedded. (This is the "remap Romansh back to the original name" step; it does not affect the published dataset.)
• FLORES-200 / BOUQuET: foil_<t>_text is foil_<t>_eng_Latn, plus foil_<t>_status, sentence_<lang> columns, and BOUQuET keeps its level column.

where <t> in {polarity_negation, role_swap, antonym_replacement, hypernym_substitution}.

Licensing

This repository aggregates subsets under different licenses. Use each config under its own license and attribute the original creators.

Config	License
flores200	CC-BY-SA-4.0
bouquet_sentence_level	CC-BY-4.0
bouquet_paragraph_level	CC-BY-4.0
wmt24pp	Apache-2.0

The flores200 subset is CC-BY-SA: any redistribution of it must remain CC-BY-SA-4.0.

Citation

If you use this data, please cite the original datasets it derives from:

@article{nllb2022,
      title={No Language Left Behind: Scaling Human-Centered Machine Translation},
      author={{NLLB Team} and Costa-jussà, Marta R. and Cross, James and others},
      journal={arXiv preprint arXiv:2207.04672},
      year={2022}
}
@inproceedings{andrews-etal-2025-bouquet,
      title={{BOUQuET}: dataset, Benchmark and Open initiative for Universal Quality Evaluation in Translation},
      author={Andrews, Pierre and Artetxe, Mikel and others},
      booktitle={Proceedings of EMNLP 2025},
      year={2025},
      url={https://aclanthology.org/2025.emnlp-main.1400/}
}
@misc{deutsch2025wmt24expandinglanguagecoverage,
      title={{WMT24++: Expanding the Language Coverage of WMT24 to 55 Languages & Dialects}},
      author={Daniel Deutsch and Eleftheria Briakou and Isaac Caswell and others},
      year={2025},
      eprint={2502.12404},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.12404}
}
@inproceedings{vamvas-et-al-2025-expanding,
      title={Expanding the WMT24++ Benchmark with Rumantsch Grischun, Sursilvan, Sutsilvan, Surmiran, Puter, and Vallader},
      author={Vamvas, Jannis and Pérez Prat, Ignacio and Soliva, Not and others},
      booktitle={Proceedings of the Tenth Conference on Machine Translation},
      year={2025},
      url={https://aclanthology.org/2025.wmt-1.79/}
}

Foils were generated with amrlib (models by Brad Jascob), the NLI model juliussteen/DeBERTa-v3-FaithAug, and WordNet (Princeton) via NLTK.

Source datasets

• https://huggingface.co/datasets/Muennighoff/flores200
• https://huggingface.co/datasets/facebook/bouquet
• https://huggingface.co/datasets/google/wmt24pp
• https://huggingface.co/datasets/ZurichNLP/wmt24pp-rm (German → 6 varieties of Romansh)