Datasets:

matic-korosec
/

cankar-parallel

language:
  - sl
license: cc-by-sa-4.0
pretty_name: CankarParallel
size_categories:
  - 10K<n<100K
task_categories:
  - text-generation
tags:
  - slovene
  - style-transfer
  - ivan-cankar
  - public-domain

CankarParallel - plain Slovene to Ivan Cankar's prose voice

9,950 parallel passages for Slovene literary style transfer. Each row pairs a passage of Ivan Cankar's prose with a plain modern Slovene rendering of the same content.

Built for CankarGPT, a from-scratch Slovene micro-LLM. The pairs train the direction plain -> cankar.

Fields

passage_id - sha256 of the Cankar text, first 16 hex chars
cankar - the original passage, unmodified public-domain text
plain - a plain modern Slovene rendering, generated
title, url - the source work
model, in_tokens, out_tokens - generation provenance

How it was made

The Cankar side comes from Wikivir (Slovene Wikisource), segmented into 2-6 sentence paragraph-bounded passages. Only prose genres were used - Cankar's plays are excluded, because a paragraph in drama is a speaker turn and segmenting them as prose glues speaker labels into the text.

The plain side was generated with claude-sonnet-5 via the Batch API, using one fixed "plain Slovene register" prompt (sha256 cec36eb07e77c6d9) shared with the inference-time drafting stage, so the styler never meets a distribution at inference it was not trained on.

Model choice was measured, not assumed. A 50-passage pilot against a cheaper model found broken Slovene grammar - wrong dual verb forms, gender disagreement, a reflexive-only verb used transitively - which a char-ngram style classifier scored identically to the stronger model's output. On a low-resource language the cheap option corrupts the data invisibly.

9,950 of 10,049 responses passed the quality filters. Rejections are quarantined, never silently dropped.

Held-out works are excluded

Passages from the frozen held-out evaluation set (holdout sha256 07ccca96a4fe339d) were excluded before generation. Pairs built from held-out works would contaminate any evaluation of a model trained on them, one-way and undetectably.

Licensing and attribution

CC BY-SA 4.0. Attribute Slovene Wikisource (Wikivir) contributors, whose volunteer transcriptions the cankar side reproduces, and this dataset.

Ivan Cankar died in 1918, so the underlying work is public domain. The transcriptions are a separate layer: whether a faithful transcription of a public-domain text attracts rights of its own is unsettled in EU law - it lacks the originality copyright requires, but the sui generis database right protects substantial investment regardless. CC BY-SA is applied deliberately rather than because the question was answered: it is what Wikivir asks for, it costs nothing here, and the alternative required being right about an open question.

Share-alike applies to redistribution of the dataset. Whether it reaches model weights trained on it is unsettled everywhere and is not a position this card takes.

Reproducing

uv run cankar pairs segment
uv run cankar pairs destyle --parse-only   # re-derives pairs from the raw log
uv run cankar pairs publish

The middle step re-parses committed raw responses and costs nothing. Generating them from scratch is destyle --limit N, which submits paid batches.

Provenance: corpus sha256 d9b05bf04db96db6, passages sha256 e1cad3a2643782f9, pairs sha256 c387b33c7d036b95, generated at 2026-07-28T12:39:40+00:00 from git acf0d2c.