datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
qg_squad[SQuAD](https://rajpurkar.github.io/SQuAD-explorer/) evaluation set for the question generation (QG) models. The split
of test and development set follows the ["Neural Question Generation"](https://arxiv.org/abs/1705.00106) work and is
compatible with the [leader board](https://paperswithcode.com/sota/question-generation-on-squad11).qg_esquad[SQuAD-es](https://huggingface.co/datasets/squad_es) dataset for question generation (QG) task.qg_jaquad[JaQuAD](https://github.com/SkelterLabsInc/JaQuAD) dataset for question generation (QG) task. The test set of the original
data is not publicly released, so we randomly sampled test questions from the training set.qg_koquad[KorQuAD](https://huggingface.co/datasets/squad_kor_v1) dataset for question generation (QG) task.qg_subjqa[SubjQA](https://github.com/megagonlabs/SubjQA) dataset for question generation (QG) task.qg_squadshifts[SQuAD Shifts](https://modestyachts.github.io/squadshifts-website/index.html) dataset for question generation (QG) task.qg_tweetqaQuestion generation dataset based on [TweetQA](https://huggingface.co/datasets/tweet_qa).tool-calls-mini
tool-calls-mini
500 synthetic tool-calling conversations in TRL's conversational format,
for supervised fine-tuning. Built to be coherent: every tool result is a plausible
function of the arguments it was called with, and every final answer reflects that
result — so the set teaches when to call a tool, not just what a call looks like.
Format
Each row has messages and tools. An assistant turn carries tool_calls instead of
content; the tool replies as a tool role… See the full description on the dataset page: https://huggingface.co/datasets/qgallouedec/tool-calls-mini.LearningQ-qg
Dataset Card for LearningQ-qg
Dataset Summary
LearningQ, a challenging educational question generation dataset containing over 230K document-question pairs by [Guanliang Chen, Jie Yang, Claudia Hauff and Geert-Jan Houben]. It includes 7K instructor-designed questions assessing knowledge concepts being taught and 223K learner-generated questions seeking in-depth understanding of the taught concepts. This new version collected and corrected from over than 50000 error and… See the full description on the dataset page: https://huggingface.co/datasets/sidovic/LearningQ-qg.QGEvalIT_QA-QGqg-passages-multi
qg-passages-multi
The training set behind olaverse/mist-qg-1.5b — passages paired with search-style questions they directly answer, across 25 languages.
Dataset Summary
~50,000 passages, each with 2–3 natural-language questions that the passage answers. Built by distilling a strong teacher model over Aya Collection passages, then filtering with round-trip retrieval verification so every kept question is a genuine, answerable positive for its passage.… See the full description on the dataset page: https://huggingface.co/datasets/olaverse/qg-passages-multi.qg-eval-multi-fresh
qg-eval-multi-fresh
Held-out evaluation passages for olaverse/mist-qg-1.5b — provably never seen during training, used to compute the model's published round-trip keep-rate.
Dataset Summary
625 passages (~25 per language, 25 languages), sampled from the same source as the training data but explicitly deduplicated against every passage used to train mist-qg-1.5b, so scores on this set measure generalization rather than memorization.
Data Fields… See the full description on the dataset page: https://huggingface.co/datasets/olaverse/qg-eval-multi-fresh.one-line-answers
One-line answers
Questions with answers that are exactly one short sentence.
A demo dataset for Finetune Studio.
Most chat datasets teach a model things it broadly already does, which makes a
fine-tune hard to see. This one teaches a style instead: after training on it, a
model that used to answer in five paragraphs answers in one line. That difference
is obvious to someone who has never trained a model before, which is the point.
How it was made
Mechanically… See the full description on the dataset page: https://huggingface.co/datasets/qgallouedec/one-line-answers.bif-qg-safe-top120-eval-inference
BIF QG Safe Top120 (gamma=1000, nbeta=1000) - Eval Inference Results
Inference results from evaluating the BIF per-query-group (safe) top120 drop experiment across all SFT checkpoints.
Experiment Details
Base model: Qwen2.5-0.5B (dolly SFT)
Training data: 480 samples (dropped top-120 by BIF safe query group scores from 600-sample alignment dataset)
BIF config: gamma=1000, nbeta=1000
Drop method: BIF top120 (query_group=safe)
Eval data: xstest_150_ordered.jsonl (50 safe +… See the full description on the dataset page: https://huggingface.co/datasets/kasadin/bif-qg-safe-top120-eval-inference.