--- license: cc-by-4.0 --- # Romansh SFT Data Supervised fine-tuning (SFT) splits built from the **swiss-ai/apertus-pretrain-rumansh** corpus. It contains dictionary list translation, sentence-level translation, idiom identification, and a small set of human-translated Romansh instructions. **Source hub:** https://huggingface.co/datasets/swiss-ai/apertus-pretrain-rumansh ## Provenance - **Dictionaries:** All dictionary entries originate from **Pledarigrond** and are provided by the Lia Rumantscha. Includes idioms: **Sursilvan, Sutsilvan, Surmiran, Rumantsch Grischun**. Each entry forms a Prompt–Answer pair of the type: - Prompt: `"Übersetze die folgende Liste von -Begriffen ins Deutsche:\n{romansh_list}"` - Answer: `"{german_list}"` - and the reverse: Prompt in German with Answer in Romansh. - **Idiom identification:** Labels derived from public text in **La Quotidiana** (see swiss-ai/apertus-pretrain-rumansh). Prompts follow the template: - Prompt: `"Sag mir in welchem Idiom der folgende Satz ist: {romansh_sentence}"` - Answer: `""` - **Human translations:** Random sample from a filtered **Tülü** dataset prepared by the **Swiss AI Initiative** (link pending). Translated by volunteers via **https://data-collection.swissai.cscs.ch/**. Prize support: **CHF 350.–** from **Prof. Antoine Bosselut**. Released under **CC BY 4.0**. - **Synthetic translations:** Sentence-level alignment was performed bidirectionally (German ↔ Idiom, Multilingual ↔ Rumantsch Grischun). - Alignment implemented with **SentenceTransformers** [`sentence-transformers/paraphrase-multilingual-mpnet-base-v2`](https://huggingface.co/sentence-transformers/paraphrase-multilingual-mpnet-base-v2), version `2.2.2`, cosine similarity ≥ **0.65**, mutual nearest-neighbour matching, and an RG word-count ratio filter ≤ **1.3×**. - Translations were then scored by **Qwen2-32B-Instruct** ([Qwen/Qwen2-32B-Instruct](https://huggingface.co/Qwen/Qwen3-32B)), deployed by the **Swiss AI Initiative**, using a strict integer-only evaluation prompt (0 for failures, otherwise 1–10 for accuracy + fluency). - Only translations with a score ≥ **7** were retained. ## File overview and counts | File | Task | Direction / Labels | # Examples | |---|---|---|---:| | `sft_dictionary_RG.jsonl` | Dictionary list translation | de → **Rumantsch Grischun**: 7,132; **Rumantsch Grischun** → de: 7,132 | **14,264** | | `sft_dictionary_Surmiran.jsonl` | Dictionary list translation | de → **Surmiran**: 3,743; **Surmiran** → de: 3,743 | **7,486** | | `sft_dictionary_Sursilvan.jsonl` | Dictionary list translation | de → **Sursilvan**: 676; **Sursilvan** → de: 676 | **1,352** | | `sft_dictionary_Sutsilvan.jsonl` | Dictionary list translation | de → **Sutsilvan**: 2,927; **Sutsilvan** → de: 2,927 | **5,854** | | `sft_grischun_quality_filtered.jsonl` | Sentence translation (filtered) | **German ↔ RG: 234; English ↔ RG: 262; French ↔ RG: 276; Italian ↔ RG: 266** | **1,038** | | `sft_surmiran_quality_filtered.jsonl` | Sentence translation (filtered) | de ↔ **Surmiran**: 42 | **42** | | `sft_surmiran_translated.jsonl` | Sentence translation | de ↔ **Surmiran**: 156 | **156** | | `sft_Sursilvan_quality_filtered.jsonl` | Sentence translation (filtered) | de ↔ **Sursilvan**: 44; **Sursilvan ↔ de**: 138 | **182** | | `sft_vallader_quality_filtered.jsonl` | Sentence translation (filtered) | de ↔ **Vallader**: 88 | **88** | | `sft_idiom_identification.jsonl` | Single-label classification | **RG**: 3,000; **Sursilvan**: 3,000; **Surmiran**: 3,000; **Vallader**: 3,000; **Puter**: 3,000; **Sutsilvan**: 1,322 | **16,322** | | `SFT_Human.jsonl` | Human-authored Romansh instructions | Free-form (Q&A, explanations, creative) | **139** | ## Acknowledgements Thanks to volunteer translators—especially **Donat D.**, **Lea B.**, and **Madlaina F.** —and to **Prof. Antoine Bosselut** for prize support. ## Contact Note that all data has been preprocessed using the pipeline in https://github.com/swiss-ai/Swiss-AI-Romansh-Scripts. Questions or corrections: **niklasc@icloud.com**