datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
ImageNetV2vaihingen-cr
Paper
This dataset is released as part of our ECCV 2026 paper:
Interpretation-Oriented Cloud Removal via Observation-Anchored Residual Flow with Geo-Contextual Alignment
Paper page: https://huggingface.co/papers/2607.02471
arXiv: https://arxiv.org/abs/2607.02471
Code: https://github.com/wzy6055/GACR
multitabqa_pretraining
Usage
import pandas as pd
from datasets import load_dataset
multitableQA_pretraining = load_dataset("vaishali/multitabqa_pretraining")
for sample in multitableQA_pretraining['train']:
sql_query = sample['query']
input_table_names = sample["table_names"]
input_tables = [pd.read_json(table, orient='split') for table in sample['tables']]
answer = pd.read_json(sample['answer'], orient='split')
# flattened input/output
input_to_model = sample["source"]
target =… See the full description on the dataset page: https://huggingface.co/datasets/vaishali/multitabqa_pretraining.vai-seller-admin-index
bep40/vai-seller-admin-index
Generated by ML Intern
This dataset repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.
Try ML Intern: https://smolagents-ml-intern.hf.space
Source code: https://github.com/huggingface/ml-intern
Usage
from datasets import load_dataset
dataset = load_dataset('bep40/vai-seller-admin-index')
CaST-Bench
CaST-Bench: Benchmarking Causal Chain-Grounded Spatio-Temporal Reasoning for Video Question Answering
This is the official repository for the CaST-Bench dataset, introduced in the paper
"CaST-Bench: Benchmarking Causal Chain-Grounded Spatio-Temporal Reasoning for Video Question Answering".
CaST-Bench is the first benchmark to evaluate Vision-Language Models (VLMs) on causal chain
reasoning grounded in fine-grained spatio-temporal evidence. Given a video and a causal question… See the full description on the dataset page: https://huggingface.co/datasets/wovenbytoyota-vai/CaST-Bench.spider-tableQA
Dataset Card for "spider-tableQA"
Usage
import pandas as pd
from datasets import load_dataset
spider_tableQA = load_dataset("vaishali/spider-tableQA")
for sample in spider_tableQA['train']:
question = sample['question']
sql_query = sample['query']
input_table_names = sample["table_names"]
input_tables = [pd.read_json(table, orient='split') for table in sample['tables']]
answer = pd.read_json(sample['answer'], orient='split')
# flattened input/output… See the full description on the dataset page: https://huggingface.co/datasets/vaishali/spider-tableQA.banglaTabQA
Dataset Card for "banglaTabQA"
Usage
import pandas as pd
from datasets import load_dataset
banglatableQA = load_dataset("vaishali/banglaTabQA")
for sample in banglatableQA['train']:
question = sample['question']
input_table = pd.read_json(sample['table'], orient='split')
answer = pd.read_json(sample['answer'], orient='split')
BibTeX entry and citation info
@inproceedings{pal-etal-2024-table,
title = "Table Question Answering for Low-resourced… See the full description on the dataset page: https://huggingface.co/datasets/vaishali/banglaTabQA.InstVL
InstVL: A Large-Scale Instance-Aware Vision-Language Dataset
This is the official repository for the InstVL dataset, introduced in the paper InstAP: Instance-Aware Vision-Language Pre-Train for Spatial-Temporal Understanding.
InstVL is a large-scale dataset of images and videos designed to bridge the gap between holistic scene understanding and fine-grained, instance-level comprehension. Current vision-language pre-training (VLP) paradigms excel at global scene understanding but… See the full description on the dataset page: https://huggingface.co/datasets/wovenbytoyota-vai/InstVL.split_legacysurvey_hsc_embeddingsTrain, calibration and test sets across models for Legacy Survey ↔ HSC embeddings (source: UniverseTBD/legacysurvey_hsc_embeddings).
vai-speech-text-parallel
Vai Speech-Text Parallel Dataset
Dataset Description
This dataset contains 23286 parallel speech-text pairs for Vai, a language spoken primarily in Ghana. The dataset consists of audio recordings paired with their corresponding text transcriptions, making it suitable for automatic speech recognition (ASR) and text-to-speech (TTS) tasks.
Dataset Summary
Language: Vai - vai
Task: Speech Recognition, Text-to-Speech
Size: 23286 audio files > 1KB (small/corrupted… See the full description on the dataset page: https://huggingface.co/datasets/michsethowusu/vai-speech-text-parallel.cricket-kg
Cricket Knowledge Graph
36,619 nodes. 1,392,017 edges. 21,324 matches — Tests, ODIs, T20s, IPL, BBL and more,
men's and women's, from December 2001 to March 2026.
Built with Samyama Graph.
Loader and ETL: samyama-ai/cricket-kg.
Attribution — required
This dataset is CC-BY-4.0, which means attribution is a condition of use, not a
courtesy. If you use it, credit the source:
Source data from Cricsheet.org, licensed
CC-BY-4.0.
That requirement travels with the data… See the full description on the dataset page: https://huggingface.co/datasets/VaidhyaMegha/cricket-kg.hindiTabQA
Dataset Card for "hindiTabQA"
Usage
import pandas as pd
from datasets import load_dataset
hinditableQA = load_dataset("vaishali/hindiTabQA")
for sample in hinditableQA['train']:
question = sample['question']
input_table = pd.read_json(sample['table'], orient='split')
answer = pd.read_json(sample['answer'], orient='split')
BibTeX entry and citation info
@inproceedings{pal-etal-2024-table,
title = "Table Question Answering for Low-resourced {I}ndic… See the full description on the dataset page: https://huggingface.co/datasets/vaishali/hindiTabQA.vaigai-dataset
Vaigai Dataset
aakkaasshh/vaigai-dataset is an Indic-focused multilingual text corpus aggregated for tokenizer training, built with the goal of reaching tokenizer quality on par with dedicated Indic tokenizers (e.g. Sarvam AI's).
One Parquet file per language is stored at the root of this repo (e.g. bn.parquet), with no nested folders.
Every time new data is fetched for a language, it is merged with that language's existing file and deduplicated on the text column, so re-running… See the full description on the dataset page: https://huggingface.co/datasets/aakkaasshh/vaigai-dataset.edge-ai-kg
Edge AI Deployment Knowledge Graph
25,152 nodes. 76,306 edges. Boards, kernels and neural networks in one graph — so you can
ask what actually runs on your silicon.
Built with Samyama Graph.
Loader and generator: samyama-ai/edge-ai-kg.
Part real, part synthetic — and every node says which
Every node carries a provenance property ("real" or "synthetic") and a source. No
node is unstamped:
provenance
Nodes
synthetic
23,910
real
1,242
Do not… See the full description on the dataset page: https://huggingface.co/datasets/VaidhyaMegha/edge-ai-kg.spider-tableQA-pretraining
Dataset Card for "spider-tableQA-pretraining"
Usage
import pandas as pd
from datasets import load_dataset
spider_tableQA_pretraining = load_dataset("vaishali/spider-tableQA-pretraining")
for sample in spider_tableQA_pretraining['train']:
sql_query = sample['query']
input_table_names = sample["table_names"]
input_tables = [pd.read_json(table, orient='split') for table in sample['tables']]
answer = pd.read_json(sample['answer'], orient='split')
# flattened… See the full description on the dataset page: https://huggingface.co/datasets/vaishali/spider-tableQA-pretraining.clinicaltrials-kg
Clinical Trials Knowledge Graph
7,628,735 nodes. 15,531,427 edges. 575,778 clinical trials registered on
ClinicalTrials.gov — with their arms, outcomes, sites, sponsors, conditions, interventions,
MeSH codes and linked PubMed publications.
Built with Samyama Graph.
Loader and ETL: samyama-ai/clinicaltrials-kg.
Stored as Parquet — 623 MB, against roughly 7 GB as raw JSON.
Attribution — required
Two acknowledgements travel with this data:
Courtesy of the U.S.… See the full description on the dataset page: https://huggingface.co/datasets/VaidhyaMegha/clinicaltrials-kg.bank-model-risk-kg
Bank Model-Risk Knowledge Graph
520 nodes. 2,391 edges. A synthetic bank's entire model-risk inventory as a graph —
models, the data and assumptions behind them, their validations and findings, the regulations
that govern them, and the submissions and decisions they drive.
Built with Samyama Graph.
Generator and loader: samyama-ai/bank-model-risk-kg.
Entirely synthetic — this is the point, not a caveat
Every record here is generated. It represents no real… See the full description on the dataset page: https://huggingface.co/datasets/VaidhyaMegha/bank-model-risk-kg.football-kg
Football World Cup Knowledge Graph
16,150 nodes. 12,384 edges. 92 years of FIFA World Cup football as a graph — 30
tournaments, 1,248 matches, 3,637 goals, 10,401 players, 88 national teams and 240 stadiums.
Built with Samyama Graph.
Loader and ETL: samyama-ai/football-kg.
What this is
Every FIFA World Cup from 1930 to 2022 — 22 men's tournaments (1930–2022) and 8 women's
tournaments (1991–2019) — modelled as a property graph rather than a set of tables.
Because… See the full description on the dataset page: https://huggingface.co/datasets/VaidhyaMegha/football-kg.megaunscene
Emergent Extreme-View Geometry in 3D Foundation Models
Yiwen Zhang¹ Joseph Tung² Ruojin Cai³ David Fouhey² Hadar Averbuch-Elor¹
¹Cornell University ²New York University ³Kempner Institute, Harvard University
MegaUnScene Benchmark
Overview
MegaUnScene is a dataset of Internet scenes unseen by existing 3DFMs for benchmarking. There are three test splits split across two evaluation tasks:
Relative Pose Estimation: UnScenePairs and UnScenePairs-t… See the full description on the dataset page: https://huggingface.co/datasets/cornell-vailab/megaunscene.pathways-kg
Pathways Knowledge Graph — partial build
138,725 nodes. 601,322 edges. Human biological pathways, protein interactions, Gene
Ontology annotations, genes, drugs and diseases in one graph.
Built with Samyama Graph.
Loader: samyama-ai/pathways-kg.
Known gaps
Measured against the engine directly, not inferred:
#
Gap
Detail
1
PARTICIPATES_IN is 2% of its reported size
The Reactome loader reports participates_in_edges=143723. The graph holds 3,423. 140,300… See the full description on the dataset page: https://huggingface.co/datasets/VaidhyaMegha/pathways-kg.mental-health-kg
Mental Health Knowledge Graph
113,710 nodes. 1,665,153 edges. US behavioural-health provision as a graph: which
facilities exist and what they offer, which clinicians are licensed to practise, where the
federal government designates a shortage, and a simulated population to measure coverage
against.
Built with Samyama Graph.
Loader and ETL: samyama-ai/mental-health-kg.
Built for referral routing — which help exists where, for whom, in what language, at what
price — and for… See the full description on the dataset page: https://huggingface.co/datasets/VaidhyaMegha/mental-health-kg.tulu-3-sft-mixture-geminiedtech-kg
Education-to-Career Pathways Knowledge Graph
1,098 nodes. 1,287 edges. One US school district's published high-school course catalogue
as a graph you can walk — which courses exist, what subject they sit in, what each requires,
and which specialty pathways include them.
Built with Samyama Graph.
Loader and ETL: samyama-ai/edtech-kg.
Why a graph
A student picks courses four times in high school. The decisions compound, they are close to
irreversible, and they are… See the full description on the dataset page: https://huggingface.co/datasets/VaidhyaMegha/edtech-kg.echo-data-rewritten-queries-hard-negative-qwen3-4bgeoQuery-tableQA
Dataset Card for "geoQuery-tableQA"
Usage
import pandas as pd
from datasets import load_dataset
geoQuery_tableQA = load_dataset("vaishali/geoQuery-tableQA")
for sample in geoQuery_tableQA['train']:
question = sample['question']
input_table_names = sample["table_names"]
input_tables = [pd.read_json(table, orient='split') for table in sample['tables']]
answer = pd.read_json(sample['answer'], orient='split')
# flattened input/output
input_to_model =… See the full description on the dataset page: https://huggingface.co/datasets/vaishali/geoQuery-tableQA.regulatory-affairs-kg
Medical-Device Regulatory Affairs Knowledge Graph
28,520 nodes. 25,332 edges. The device-to-law backbone of US medical-device regulation:
every 510(k) clearance for cardiovascular devices, the product code each was cleared under,
and the Code of Federal Regulations section that governs it.
Built with Samyama Graph.
Loader and ETL: samyama-ai/regulatory-affairs-kg.
The question it answers
When a rule or a recognised standard changes, a manufacturer has to work out… See the full description on the dataset page: https://huggingface.co/datasets/VaidhyaMegha/regulatory-affairs-kg.tulu-3-sft-mixture-hard-negative-llama-3becho-data-rewritten-queriesatis-tableQA
Dataset Card for "atis-tableQA"
Usage
import pandas as pd
from datasets import load_dataset
atis_tableQA = load_dataset("vaishali/atis-tableQA")
for sample in atis_tableQA['train']:
question = sample['question']
sql_query = sample['query'],
answer = pd.read_json(sample['answer'], orient='split')
# flattened input
input_to_llm = sample["source"]
target = sample["target"]
BibTeX entry and citation info
@inproceedings{pal-etal-2023-multitabqa… See the full description on the dataset page: https://huggingface.co/datasets/vaishali/atis-tableQA.mesh-snomed-entity-alignment-15k
MeSH-SNOMED Entity Alignment 15K
MeSH-SNOMED Entity Alignment 15K is a biomedical heterogeneous knowledge graph alignment benchmark for cross-ontology matching between MeSH and SNOMED CT. It is designed to evaluate entity alignment systems under realistic large-graph conditions, where gold-aligned concepts are embedded in much larger biomedical graphs containing many structurally relevant but non-aligned background entities. This release is intended for the accompanying EMNLP… See the full description on the dataset page: https://huggingface.co/datasets/vaibhavalakshmiravideshik/mesh-snomed-entity-alignment-15k.