datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
multihop_qamultihop_qa_sftmultihop_qa_sft_doc4096_seq1024_v2zoology_multihop
Zoology Multihop Associative Retrieval
irodkin/zoology_multihop is a deterministic synthetic associative-retrieval
dataset with multiple queries per context. It uses integer token arrays rather
than natural-language text.
Configurations
Every configuration is named N{N}-H{H}-V4096 and uses one of:
N: 8, 16, 32, 64, or 128 total key-value edges.
H: 1, 2, 4, or 8 edges per chain.
V: exactly 4,096 total tokens.
The complete release contains all 20 combinations. A… See the full description on the dataset page: https://huggingface.co/datasets/irodkin/zoology_multihop.MultihopSpatial
[ECCV 2026] MultihopSpatial: Multi-hop Compositional Spatial Reasoning Benchmark for Vision-Language Models
Project Page |
Paper |
Model
Overview
MultihopSpatial is a benchmark designed to evaluate whether vision-language models (VLMs) demonstrate robustness in multi-hop compositional spatial reasoning. Unlike existing benchmarks that only assess single-step spatial relations, MultihopSpatial features queries with 1 to 3 reasoning hops paired with… See the full description on the dataset page: https://huggingface.co/datasets/etri-vilab/MultihopSpatial.Multi-Hop-Objects-Countingar_multihop
AR Multihop Associative Retrieval
irodkin/ar_multihop is a deterministic synthetic associative-retrieval
dataset with multiple queries per context. Hub rows are raw structured integer
arrays; model-specific structural tokens and sequence formatting are added by a
training collator.
Configurations
Configurations are named N{N}-H{H}-V16 for every combination of:
N: 8, 16, 32, 64, or 128 total key-value edges.
H: 1, 2, 4, or 8 edges per chain; H must divide N.
V16:… See the full description on the dataset page: https://huggingface.co/datasets/irodkin/ar_multihop.multihopqa
MultiHopQA
This dataset contains the MultiHopQA data along with intermediate retrieval and generation steps, as well as final predictions generated in the paper Chain-of-Retrieval Augmented Generation.
Fields
The dataset includes the following fields for each data point:
query: The multi-hop question.
query_id: A unique identifier for the query.
answers: A list of correct answer(s) to the multi-hop question.
context_doc_ids: A list of document IDs retrieved by the… See the full description on the dataset page: https://huggingface.co/datasets/corag/multihopqa.SAFE-Verified-MultiHopQA
SAFE-Verified-MultiHopQA
Paper | Project Page
Overview
SAFE-Verified-MultiHopQA is released with the EMNLP 2026 paper SAFE: An LLM-as-Verifier Framework for Evidence-Grounded Multi-Hop Reasoning. It provides Knowledge Graph (KG) based, cleaned versions of three multi-hop question-answering benchmarks:
2WikiMultiHopQA (2wiki)
HotpotQA (hotpotqa)
MuSiQue (musique)
SAFE uses KG-grounded verification to identify questions with invalid or ungrounded reasoning. This… See the full description on the dataset page: https://huggingface.co/datasets/Daeyongkwon98/SAFE-Verified-MultiHopQA.cql_gen-browsecomp_plus_qa_gen-oai_gpt5_low-multihop_2-v3187MegaWikiQA-v1-multihop
MegaWikiQA v1 Multihop Dataset
Combined and shuffled Wiki5M-based synthetic multihop QA dataset for hypernetwork / knowledge-injection research.
Sources
Hop
Source dataset
Rows
1
nace-ai/wiki5m_1hop_qa_pairs_1M_stratified_with_domain
1,000,000
2
nace-ai/wiki5m_2hop_qa_pairs_noun_v10_domain
1,894,483
3
nace-ai/wiki5m_3hop_qa_pairs_noun_v10_domain
2,393,688
Total
shuffled (seed=42)
5,288,171
Schema
question, answer
hop — 1 / 2 /… See the full description on the dataset page: https://huggingface.co/datasets/nace-ai/MegaWikiQA-v1-multihop.multihop-qa-testwiki-multihop-qa-500k
wiki-multihop-qa-500k
500,000 synthetic multi-hop QA pairs generated from Wikipedia.
Built for training the think-in-silence latent reasoning model — a model that reasons entirely in vector space without generating chain-of-thought tokens.
Why This Dataset Exists
Most publicly available QA datasets have two problems for reasoning research:
Too small. HotpotQA has 113K samples. StrategyQA has 2.8K. Not enough diversity to train a generalizable reasoning module.
Too many… See the full description on the dataset page: https://huggingface.co/datasets/rajat5039/wiki-multihop-qa-500k.LMLM-multihop-mquake-sftmath-visual-multihop-qamultihopfakepedia_GTmultihopRAG-1024multihop_preferencesmultihop_qa_sft-hard-negatives
multihop_qa_sft — hard negative IDs
Hard negatives mined for ragrawal36/multihop_qa_sft
(train split). IDs only — they index the document corpus at
mihir-1999/multihop_qa_sft-doc-corpus.
Schema
column
type
meaning
row_id
int32
source row index in ragrawal36/multihop_qa_sft train
pos_doc_ids
list[int32]
the row's own supporting docs (positives)
neg_doc_ids
list[int32]
200 mined hard negatives
pos_doc_ids and neg_doc_ids are disjoint by… See the full description on the dataset page: https://huggingface.co/datasets/mihir-1999/multihop_qa_sft-hard-negatives.multihopqa
MultiHopQA
This dataset contains the MultiHopQA data along with intermediate retrieval and generation steps, as well as final predictions generated in the paper Chain-of-Retrieval Augmented Generation.
Fields
The dataset includes the following fields for each data point:
query: The multi-hop question.
query_id: A unique identifier for the query.
answers: A list of correct answer(s) to the multi-hop question.
context_doc_ids: A list of document IDs retrieved by the… See the full description on the dataset page: https://huggingface.co/datasets/KOJKO/multihopqa.MultiHopVQAmultihop-qa-mixedMultiHop-Filtered-VQAmultihopfakepedia_OLMoE-1B-7B-0125-Instruct_allmultihop_qa_sft-doc-corpus
multihop_qa_sft — deduplicated document corpus
Flat document corpus extracted from
ragrawal36/multihop_qa_sft
(train split), built as stage 0 of a hard-negative mining pipeline.
Each source row packs its supporting paragraphs into one paragraphs string joined by
<|doc_separator|>. This dataset splits them into individual documents, deduplicates
exactly (whitespace-collapsed, casefolded), and assigns stable integer doc_ids.
source rows
1,341,045
document instances… See the full description on the dataset page: https://huggingface.co/datasets/mihir-1999/multihop_qa_sft-doc-corpus.multihopRAG-256multihop_qa_testmultihop_qa_sft-hard-neg-cotnemotron-tool-n1-multihop-corrected
Nemotron Tool-N1 Multi-hop Dataset (Corrected)
This dataset contains multi-hop tool calling examples with corrected prompt formatting.
Dataset Info
Total samples: 5,555
Train samples: 4,999
Test samples: 556
Columns: prompt, answer, id, dataset, task
Source: Downloaded from arcee-train/nemotron-tool-n1-multihop and corrected
Corrections Applied
Converted string prompts to chat message format
Added proper system messages
Ensured consistent structure across… See the full description on the dataset page: https://huggingface.co/datasets/Anna4242/nemotron-tool-n1-multihop-corrected.multi-hop-psychiatric-medical-assessments-test-data