Datasets:

dappalumbo91
/

pflt-fsot-sample

license: apache-2.0
pretty_name: PFLT FSOT Sample + Benchmark Snapshot
tags:
  - translation
  - multilingual
  - evaluation
  - fsot
  - sacrebleu
  - offline
task_categories:
  - translation
language:
  - multilingual
size_categories:
  - n<1K

PFLT FSOT — sample densify + benchmark snapshot

Version: 0.2.0 (2026-07-21)
Companion model card: dappalumbo91/pflt-fsot
Code: protofluid-language-translator

File	Role
`sample_densify.tsv`	Small form→gloss densify sample for demos
`metrics_snapshot.json`	Public catalog + MT benchmarks (sacreBLEU / chrF)
`DATASET_README.md`	This card

What this is not

Not the full multi-GB densify/gold lexicon (local product only)
Not FLORES scores (Hub parquet still gated)
Not a claim of commercial DeepL news SOTA

Headline numbers (v0.2.0)

Track	Metric	Value
Catalog	langs / form→gloss product	113 / ~99.99%
Chat neural open-set	mean best sacreBLEU (16 langs)	50.19
Hybrid oracle	mean sacreBLEU	53.58
WMT14 de→en	opus-mt-de-en sacreBLEU	33.88
WMT14 de→en	NLLB-600M sacreBLEU	33.37

Full tables and honesty notes: model card + metrics_snapshot.json.

License

Apache-2.0 for packaging. Sample rows may derive from Wiktionary-class data — respect upstream licenses.

PFLT FSOT — sample densify + benchmark snapshot

Contents

What this is not

Headline numbers (v0.2.0)

License