CoolFace
Datasetpublic

Finnish-NLP/dolci-instruct-sft-fi

Dolci Instruct SFT — Finnish (machine-translated) Finnish machine translation of the Dolci Instruct SFT mixture (dolci-instruct-sft-filtered-v1, no-math / no-latex subset), produced for SFT of Finnish LLMs. Translation model: translategemma-27b (Gemma-based 27B translation model) Rows: 234,745 (multi-turn chat, mostly single Q→A) Language: Finnish (fi) Format: chat messages (role / content) Provenance & filtering Translated from the English… See the full description on the dataset page: https://huggingface.co/datasets/Finnish-NLP/dolci-instruct-sft-fi.

sourceHugging Faceodc-byupdated 2mo agoView on Hugging Face
0likes30downloads
Dataset Card

Dolci Instruct SFT — Finnish (machine-translated)

Finnish machine translation of the Dolci Instruct SFT mixture (dolci-instruct-sft-filtered-v1, no-math / no-latex subset), produced for SFT of Finnish LLMs.

  • —Translation model: translategemma-27b (Gemma-based 27B translation model)
  • —Rows: 234,745 (multi-turn chat, mostly single Q→A)
  • —Language: Finnish (fi)
  • —Format: chat messages (role / content)

Provenance & filtering

Translated from the English dolci-instruct-sft-filtered-v1-no-math-no-latex mixture. During translation, oversize prompts were skipped, leaving some rows with empty content in the raw output. Those rows were removed here: 246,444 raw rows → 234,745 rows with non-empty content on every message. The all-null function_calls / functions fields present in the raw parquet were dropped; each message keeps only role and content.

Schema

columntypedescription
idstringsource example id
messageslist<struct<role: string, content: string>>chat turns
source_datasetstringoriginating sub-dataset (Aya, FLAN, WildGuardMix, …)
domainstringdomain label (Safety, Chat, Reasoning, Science, …)

Source-dataset composition (top)

Aya, FLAN, WildGuardMix, WildJailbreak, Verifiable Reasoning, Logic Puzzles, Wildchat, Dolci Instruct Precise IF, CoCoNot, OpenAssistant, OpenThoughts3+ Science, SciRiff, and a small hardcoded set.

Usage

python
from datasets import load_dataset
ds = load_dataset("Finnish-NLP/dolci-instruct-sft-fi", split="train")
print(ds[0]["messages"])

Caveats

Machine-translated; translation artifacts are possible. Safety/jailbreak subsets contain adversarial prompts by design. Content inherits the license and terms of the upstream Dolci/Tulu sources (ODC-BY). Verify licensing of individual sub-datasets before downstream redistribution.