Dataset Viewer
Auto-converted to Parquet Duplicate
id
large_stringclasses
7 values
sentence
large_stringclasses
7 values
source
large_stringclasses
1 value
isv-0001
Ljudi govoręt, že v tamtoj dènj bylo veliko spokojno.
nnnbo
isv-0002
Čisto nebo bylo jasno, a oblåky šli pomalo, nesene legkym větrom.
nnnbo
isv-0003
Ptice, ktore sųt ješče ne poletěli do Raja, pěvali krasno posrěd žòlto-oranževyh drěv.
nnnbo
isv-0004
Jesenj råzlivala sę v světu i pomalo, pomalo prinosila ona na Zemjų hlåd i pŕvy zimovy mråz.
nnnbo
isv-0005
Ljudi govoręt, že v taky dènj prišla Črvena Iskra.
nnnbo
isv-0006
Ja ne zamnogo dobro pamętajų tutoj dènj, ale znajų ja, že to, čto govoręt ljudi, ne jest točno pravda.
nnnbo
isv-0007
Ale takože ja råzumějų, začto oni jego tako pamętajųt.
nnnbo

Interslavic MT Data

A data-only repository: the canonical parallel corpus for training machine translation involving Interslavic. Maintained by the InterSlavic Linguistic Forum Foundation (ISLF).

We provide validated Interslavic sentences and host parallel translations from the community. We do not ship training code; others build their own repos and train on this data. At most, this repo has small scripts for validation, format checks, and conversion.

Goal: Quality governance — so that models marketed as "Interslavic" meet a clear standard and uncertified ones are clearly distinguished.


Contents

Path Description
monolingual/isv_sentences.parquet Single source of truth: Interslavic sentence id, text, and source. All rows verified before merge.
parallel/*.parquet One Parquet per language pair (e.g. isv-eng.parquet). Each row references an id from the monolingual set.
scripts/ Format validation and optional conversion tools.

Format: Parquet only (columnar, compact, streamable).


Data formats

  • Monolingual (monolingual/isv_sentences.parquet): columns id, sentence, source. License per source is documented (e.g. in CONTRIBUTING or SOURCES.md).
  • Parallel (parallel/isv-{lang}.parquet): columns id, isv, target, translator, method (human / machine_raw / machine_postedited), review_status. Every id must exist in the monolingual set; isv must match exactly.

Licensing

  • Data: ISLF Open Data License v1.0 — open use with mandatory certification requirement for models. If you train a model on this data, you must either get ISLF certification OR display a prominent disclaimer. This is a license condition, not optional.
  • Scripts/tooling: MIT License.
  • Quick summary: DATA-USE-AGREEMENT.md explains your obligations in plain language.

How to use

  1. Clone the repo and read CONTRIBUTING.md for format and PR rules.
  2. Use the Parquet files in your own training pipeline. Each parallel file references ids from monolingual/isv_sentences.parquet.
  3. If you publish a model that produces Interslavic output, comply with the DATA-USE-AGREEMENT (attribution, certification or disclaimer).

Validation

From the repo root:

pip install -r requirements.txt
python scripts/validate_corpus.py

Run this before committing changes to monolingual/ or parallel/.


Certification (ISLF)

ISLF certifies output quality of models (like a language exam), not the software.

  • Certified models may use the "Certified by ISLF" mark
  • Uncertified models MUST display a mandatory disclaimer (see LICENSE.md §3.3)
  • Violations result in license termination and public identification

A public registry of certified models will be maintained. Certification criteria will be published separately. Contact ISLF to begin the certification process.

Downloads last month
5