Why TF2 Chose Translation Before Romanian-Native Generation
The decision to use controlled English fables as source material for Romanian literary translation.
TF2 takes an apparent detour: it translates controlled English fables before asking a model to generate directly in Romanian. The detour buys a paired object. Every Romanian text has a source whose events, relationships, and moral can be checked.
TF1 already provided three million English stories with structured specifications and provenance. Translating a selected subset creates a parallel resource: each Romanian output can be compared with a known English source, and multiple systems can be evaluated on the same material.
Why the paired design matters
A native Romanian story can be judged for grammar and narrative quality, but there is no source whose meaning it must preserve. Translation adds a constraint: accuracy and cultural adaptation can disagree, and the evaluation must show both.
The paper’s inspected example is blunt: the untuned model turns “skunk” into the invented Fumeg, while the tuned translation preserves sconcs. The sentence can remain fluent while the animal changes species.
What the design gives up
Translated fables are not the same as Romanian-native writing. They can carry source-language syntax, narrative conventions, and model-specific translationese. A large translated corpus may therefore be useful for controlled experiments while remaining a poor sample of Romanian literature.
TF3 later adds Romanian-native generation, which lets the two regimes be studied separately instead of conflated.
The experimental path
TF2 uses a 15K silver-reference set for instruction tuning and evaluation, then builds a three-million-pair corpus at scale. The paper records both artifacts and compares open models with larger proprietary systems under corpus metrics and a five-dimensional rubric.
The TF2 paper separates the 15K set, the large corpus, and the fine-tuned model. The 15K Romanian references are model-generated silver data, not human literary translations. Pairing makes their relationship to the English source inspectable; it does not remove translationese or turn synthetic references into a human standard.