datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
apertus_multiblimpapertus-8b-greek-cpt-modern-greek-train
Exact Modern-Greek training content for Apertus 8B Greek CPT
This is the public Modern-Greek, train-only document snapshot selected for the full 8B D0 continued-pretraining run. It preserves the upstream v2 schema and metadata; text is reproduced as its exact training-time Apertus-parity PII-masked value. Selection is reconstructed from immutable post-mask training catalogs and content hashes. It contains no replay payload.
Exact selected content
HPLT Modern… See the full description on the dataset page: https://huggingface.co/datasets/fffoivos/apertus-8b-greek-cpt-modern-greek-train.apertus-pretrain-romanshThis dataset consist of three differnt parts. Monolingual Romansh Data, Polylingual data or more precisely translated data from Romansh into either German, French, Italian or English and Sythetic Data.
The Polylingual data consists of aligned and non aligned data. The synthetic data was created by interweaving the translational data and prefacing it with the sentence " This is a text translated from SOURCE LANGUAGE to Rumantsch Grischun".
The data has a metadata "idiom" if the if specific… See the full description on the dataset page: https://huggingface.co/datasets/swiss-ai/apertus-pretrain-romansh.apertus-pretrain-swiss
Swiss Pretrain Data
This dataset provides a large collection of open-access and license-compliant Swiss data sources for language model training.
The dataset includes the following sources:
Name
Internal ID
Tokens (B)
Description
Curia Vista
curiavista
0.5
Legal and administrative documents from the Swiss database of parliamentary proceedings.
enscheidsuche
enscheidsuche_html
4.5
Swiss court decisions, sampled at 50% for balance.
FineWeb-2… See the full description on the dataset page: https://huggingface.co/datasets/swiss-ai/apertus-pretrain-swiss.Apertus_v1.5_Preference_Data
Apertus 1.5 Preference Dataset
This is the preference dataset used for the offline DPO stage of Apertus v1.5 alignment training, applied to the 70B model.
The prompts come from Ai2's Olmo 3 Dolci-Instruct-DPO dataset. We only reuse the prompts from Dolci-Instruct-DPO; all chosen / rejected responses in this dataset were generated by us.
How this dataset was built
Prompts. Taken from Dolci-Instruct-DPO (ODC-BY).
Response generation and annotation. Every prompt was… See the full description on the dataset page: https://huggingface.co/datasets/swiss-ai/Apertus_v1.5_Preference_Data.apertus-sft-mixture
Apertus Supervised Finetuning Data
Our supervised finetuning data contains a carefully curated blend of instruction-following datasets,
developed through eight iterations of empirical evaluation. This final mixture comprises approximately
3.8 million examples from diverse sources, balancing generalinstruction-following, mathematical reasoning,
code generation, and multilingual capabilities.
More details about data provenance, preparation, and statistics can be found in our tech… See the full description on the dataset page: https://huggingface.co/datasets/swiss-ai/apertus-sft-mixture.apertus-v1.1-1.5b-atlasApertus-8B-2509-microQAT-logitsThis dataset provides a small sample of TOP-K logits computed using swiss-ai/Apertus-8B-2509 on samples from Data Phase 5 of Apertus pre-training.
Format
This data represents documents packed into chuncks of 4096 tokens separated by EOS. The provided fields are as follows:
input_ids: Input tokens.
index: Positions of top-256 highest-probability next-token predictions for each token.
exp_logits: Normalized probabilities of top-256 highest-probability next-token predictions for each… See the full description on the dataset page: https://huggingface.co/datasets/daslab-testing/Apertus-8B-2509-microQAT-logits.Apertus-v1.5-QAT-10K
mlx-community/Apertus-v1.5-QAT-10K
This is a 2000 sample subset of the chosen pairs inside swiss-ai/Apertus_v1p5_Preference_Data for MLX-LM-LoRA and MLX-LoRA-Studio and the Quantization Aware Trained Appertus models.
apertus-pretrain-poisonandcanariesThis dataset was used as part of Apertus v1 training for poisoning experiments. See our technical report for details, as well as the dedicated study.
apertus-v1.1-0.5b-atlas
apertus-v1.1-0.5b-atlas
EAGLE3-Apertus-8B-Instruct-2509-Data
EAGLE3-Apertus-8B-Instruct-2509-Data
Training dataset for the thomaskiefer/EAGLE3-Apertus-8B-Instruct-2509 speculative decoding draft model.
Dataset Description
This dataset contains ~375k multi-turn conversations used to train an Eagle3 draft model for swiss-ai/Apertus-8B-Instruct-2509.
Data Sources
The prompts are sourced from:
UltraChat - Large-scale multi-turn dialogue dataset
ShareGPT - Real user conversations
OpenThoughts-114k-math - Mathematical… See the full description on the dataset page: https://huggingface.co/datasets/thomaskiefer/EAGLE3-Apertus-8B-Instruct-2509-Data.apertus-posttrain-romansh
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… See the full description on the dataset page: https://huggingface.co/datasets/swiss-ai/apertus-posttrain-romansh.Apertus-8B-Instruct-2509-microQAT-logitsapertus-pretrain-romansh-backtranslatedVersion of https://hf.co/datasets/swiss-ai/apertus-pretrain-romansh (monolingual split only) that includes MT-generated translations into German.
The intended purpose of this dataset is to train MT systems or LLMs on the task of idiom-specific German→Romansh translation. Note that the German translations in this dataset might contain errors, since they have been automatically generated by an MT system.
Composition of the dataset and Romansh data sources
See… See the full description on the dataset page: https://huggingface.co/datasets/jvamvas/apertus-pretrain-romansh-backtranslated.train_prm800k_apertus_70b_gpt-oss-120b_annotated_256tok_part6greek-apertus-sft
Greek Apertus SFT datasets
The supervised fine-tuning data of the Greek Apertus project: the GlossAPI team of EELLAK (Open Technologies Alliance) continues the pre-training of swiss-ai/Apertus-8B-2509 on Greek text and then trains it on instructions, with a grant from the Swiss AI Initiative. This repository holds every training arm we assembled, exactly as it went (or goes) to the trainer: one messages list per row, chat format, no system turn.
Access is gated: request it and… See the full description on the dataset page: https://huggingface.co/datasets/fffoivos/greek-apertus-sft.apertus-v1.1-4b-atlasfiltered_apertus_pretrainedUsed to create this dataset:
import json
import os
from datasets import load_dataset
from tqdm import tqdm
# --- Configuration ---
DATASET_NAME = "swiss-ai/apertus-pretrain-swiss"
SUBSTRING_TO_FILTER = "entscheidsuche_html"
COLUMN_TO_CHECK = "id"
OUTPUT_FILENAME = "filtered_apertus_pretrain_swiss.jsonl"
def filter_function(example):
"""
Returns True to keep the example, False to discard it.
We keep the row only if the substring is NOT in the 'id' column.
"""
return… See the full description on the dataset page: https://huggingface.co/datasets/liechticonsulting/filtered_apertus_pretrained.emotion_stories_Apertus_8B_Instruct
Emotion Stories — Apertus-8B-Instruct
Synthetic short stories that convey a target emotion implicitly — without ever
naming the emotion or its direct synonyms. Each story expresses the emotion only
through actions, body language, dialogue, internal reactions, and situational
context. The dataset was built to study emotion representations in language
models (e.g. probing and activation-steering experiments).
Generated with swiss-ai/Apertus-8B-Instruct-2509.
A companion set… See the full description on the dataset page: https://huggingface.co/datasets/snae/emotion_stories_Apertus_8B_Instruct.croco-munin-apertus-8b-da-simpo-fullcroco-munin-apertus-8b-da-50kcroco-munin-apertus-8b-da-simpo-full-50kapertus-pretrain-romansh-backtranslated-sftSubsampled version of https://huggingface.co/datasets/jvamvas/apertus-pretrain-romansh-backtranslated that has been processed as follows:
Balanced subsampling to 30k samples (5k samples per variety)
Samples with higher LID scores are prioritized
Formatted as German-to-Romansh translation instruction pairs in prompt/completion format (prompt: Übersetze den folgenden Text nach {variety}:\n\n{german_backtranslation}; completion: the Romansh text)
Normalized linebreaks to have clear text… See the full description on the dataset page: https://huggingface.co/datasets/jvamvas/apertus-pretrain-romansh-backtranslated-sft.train_prm800k_apertus_70b_math_texts_part1apertus_70b_3899_0.2_0.75_correctness_no_refuse_sfttrain_prm800k_apertus_70b_math_texts_part5train_prm800k_apertus_70b_math_texts_256tok_part6train_prm800k_apertus_70b_gpt-oss-120b_annotated_256tok_part4train_prm800k_apertus_70b_gpt-oss-120b_annotated_256tok_part7