datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
casimedicos-exp
Antidote CasiMedicos Dataset - Possible Answers Explanations in Resident Medical Exams
We present a new multilingual parallel medical dataset of commented medical exams which includes not only explanatory arguments
for the correct answer but also arguments to explain why the remaining possible answers are incorrect.
This dataset can be used for various NLP tasks including: Medical Question Answering, Explanatory Argument Extraction or Explanation Generation.
The… See the full description on the dataset page: https://huggingface.co/datasets/HiTZ/casimedicos-exp.MedExpQA
MexExpQA: Multilingual Benchmarking of Medical QA with reference gold explanations and Retrieval Augmented Generation (RAG)
We present a new multilingual parallel medical benchmark, MedExpQA, for the evaluation of LLMs on Medical Question Answering.
This benchmark can be used for various NLP tasks including: Medical Question Answering or Explanation Generation.
Although the design of MedExpQA is independent of any specific dataset, for the first version of the… See the full description on the dataset page: https://huggingface.co/datasets/HiTZ/MedExpQA.BertaQA
Dataset Card for BertaQA
BertaQA is a trivia dataset comprising 4,756 multiple-choice trivia questions, with one single correct answer and 2 additional distractors. Crucially, questions are distributed between local and global topics. Whereas answering questions in the latter group requires general world knowledge, local questions require specific knowledge about the Basque Country and its culture. Additionally, questions are classified into eight categories, namely Basque and… See the full description on the dataset page: https://huggingface.co/datasets/HiTZ/BertaQA.EusProficiency
Dataset Card for EusProficiency
EusProficiency comprises 5,169 exercises on different topics from past EGA exams, the official C1-level certificate of proficiency in Basque.
We collected the atarikoa exercises from EGA exams through the years 1998 to 2008. Atarikoa is the first qualifying test of EGA, which measures different aspects of language competency, such as reading comprehension, grammar, vocabulary, spelling, and writing. Each test generally has 85 multiple-choice questions… See the full description on the dataset page: https://huggingface.co/datasets/HiTZ/EusProficiency.EusReading
Dataset Card for EusReading
EusReading consists of 352 reading comprehension exercises (irakurmena) sourced from the set of past EGA exams from 1998 to 2008. Each test generally has 10 multiple-choice questions, with 4 choices and a single correct answer. These exercises are more challenging than Belebele due to the complexity and length of the input texts. As a result, EusReading is useful to measure long context understanding of models.
Curated by: HiTZ Research Center & IXA… See the full description on the dataset page: https://huggingface.co/datasets/HiTZ/EusReading.EusTrivia
Dataset Card for EusTrivia
EusTrivia consists of 1,715 trivia questions from multiple online sources. 56.3% of the questions are elementary level (grades 3-6), while the rest are considered challenging. A significant portion of the questions focus specifically on the Basque Country, its language and culture. Each multiple-choice question contains two, three or four choices (3.84 on average) and a single correct answer. Five areas of knowledge are covered:
Humanities and Natural… See the full description on the dataset page: https://huggingface.co/datasets/HiTZ/EusTrivia.Hi-ToM_Dataset
Hi-ToM Dataset
This is the dataset for the paper "Hi-ToM: A Benchmark for Evaluating Higher-Order Theory of Mind Reasoning in Large Language Models".
The Hi-ToM_data folder
Contains ToMh data consisting of story-question pairs and the corresponding answers.
The names of subfolder branches have the following meanings:
Tell / No_Tell: whether or not the stories contain communications among agents.
MC / CoT: the prompting style. MC corresponds to Vanilla Prompting (VP) in… See the full description on the dataset page: https://huggingface.co/datasets/Hi-ToM/Hi-ToM_Dataset.PIQA-eu
Dataset Card for PIQA-eu
Point of Contact: hitz@ehu.eus
Dataset Description
Dataset Summary
PIQA-eu is the professional translation to Basque of the PIQA's
(Bisk et al., 2020) validation partition.
PIQA is a commonsense QA benchmark for naive physics reasoning focusing on how we interact with everyday
objects in everyday situations.
Languages
eu-ES
Dataset Structure
Data Instances
PIQA-eu examples look like this:
{… See the full description on the dataset page: https://huggingface.co/datasets/HiTZ/PIQA-eu.elkarhizketak-RAG
Dataset Card for ElkarHizketak RAG and its Disruptor Variants
Base and disruptor variants of ElkarHizketak, built to stress-test conversational RAG systems in Basque under realistic interaction patterns (conversational openings, topic shifts).
Dataset Details
Dataset Description
This dataset extends ElkarHizketak with a base variant (rewritten opening queries, retrieval-needed labels, retrieved chunks) and disruptor variants that inject… See the full description on the dataset page: https://huggingface.co/datasets/HiTZ/elkarhizketak-RAG.f-cov-offset1024-qwen3-1.7b-base-math12k-c50b0bba-rollouts
f_cov_Qwen3-1.7B-Base_math12k_offset1024_token_mean rollouts
This dataset contains one compressed JSONL shard for every completed training
step. The step and rollout_index columns uniquely locate a rollout within
this training run. Run metadata and per-step row counts are recorded in
rollout_manifest.json.
f-cov-offset2048-qwen3-1.7b-base-math12k-4a578b3f-rollouts
f_cov_Qwen3-1.7B-Base_math12k_offset2048_token_mean rollouts
This dataset contains one compressed JSONL shard for every completed training
step. The step and rollout_index columns uniquely locate a rollout within
this training run. Run metadata and per-step row counts are recorded in
rollout_manifest.json.
fixed-n-rb-offset-cost-aware-marginrl-qwen3-1.7b-base-math12k-offset2048-token-mean-rollouts
fixed_n_rb_offset_cost_aware_marginrl_Qwen3-1.7B-Base_math12k_offset2048_token_mean rollouts
This dataset contains one compressed JSONL shard for every completed training
step. The step and rollout_index columns uniquely locate a rollout within
this training run. Run metadata and per-step row counts are recorded in
rollout_manifest.json.
IBERtaQA
Dataset Card for IBERtaQA
Curated by:
HiTZ Center, University of the Basque Country (EHU)
Barcelona Supercomputing Center (BSC)
CiTIUS, University of Santiago de Compostela (USC)
GPLSI, University of Alicante (UA)
Funded by: Project Desarrollo de Modelos ALIA; Projecte AINA
License: CC-BY-4.0
Dataset Summary
IBERtaQA is a multilingual benchmark designed to evaluate the cultural and factual knowledge of language models across Iberian languages
and cultural… See the full description on the dataset page: https://huggingface.co/datasets/HiTZ/IBERtaQA.fixed-n-rb-offset1024-qwen3-1.7b-base-math12k-d8311fda-rollouts
fixed_n_rb_offset_cost_aware_marginrl_Qwen3-1.7B-Base_math12k_offset1024_token_mean_rerun rollouts
This dataset contains one compressed JSONL shard for every completed training
step. The step and rollout_index columns uniquely locate a rollout within
this training run. Run metadata and per-step row counts are recorded in
rollout_manifest.json.
BasPhyCo
BasPhyCo
This repository contains the data and code for the paper Physical Commonsense Reasoning for Lower-Resourced Languages and Dialects: a Study on Basque.
Physical commonsense reasoning represents a fundamental capability of human intelligence, enabling individuals to understand their environment, predict future events, and navigate physical spaces. Recent years have witnessed growing interest in reasoning tasks within Natural Language Processing (NLP). However, no prior… See the full description on the dataset page: https://huggingface.co/datasets/HiTZ/BasPhyCo.hitori-9x9-unique
Hitori 9×9: uniquely solvable, difficulty-graded
1,950 machine-generated 9×9 Hitori puzzles,
every one verified to have exactly one solution: 1,000 train + 200 test
(standard config) and a 750-puzzle stratified hard test set (hard config).
Built as the training/eval domain for PRISM: an experiment testing whether
HRM-style recurrent reasoning models (27M params) generalize to a puzzle domain
they have never been trained on. A 27M-parameter HRM trained on the 1,000 standard… See the full description on the dataset page: https://huggingface.co/datasets/siaglass/hitori-9x9-unique.hitokoto这里收集了 Hitokoto 中所有的语句。
数据来自 https://sentences-bundle.hitokoto.cn。
HABE-HiTZ_C1_Dataset
HABE-HiTZ C1 Dataset
Paper: Accepted in LREC 2026Contact: ekhi.azurmendi@ehu.eus
Dataset Summary
The HABE-HiTZ C1 dataset is a collection of essays and annotations designed for training and evaluating models on scoring and feedback generation.
The dataset is organized into three distinct subsets. To ensure traceability, every essay is assigned a unique ID that remains consistent across all three subsets, allowing you to easily map scores, feedbacks, generations and… See the full description on the dataset page: https://huggingface.co/datasets/EkhiAzur/HABE-HiTZ_C1_Dataset.per-context-rb-l0-0-qwen3-1.7b-compression-bs32-32k-146103-rollouts
per_context_rb_l0_0_Qwen3-1.7B_compression_bs32_n16_32k_1epoch rollouts
This dataset contains one compressed JSONL shard for every completed training
step. The step and rollout_index columns uniquely locate a rollout within
this training run. Run metadata and per-step row counts are recorded in
rollout_manifest.json.