datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
MINT-1T-PDF-CC-2023-23
🍃 MINT-1T:Scaling Open-Source Multimodal Data by 10x: A Multimodal Dataset with One Trillion Tokens
🍃 MINT-1T is an open-source Multimodal INTerleaved dataset with 1 trillion text tokens and 3.4 billion images, a 10x scale-up from existing open-source datasets. Additionally, we include previously untapped sources such as PDFs and ArXiv papers. 🍃 MINT-1T is designed to facilitate research in multimodal pretraining. 🍃 MINT-1T is created by a team from the University of Washington in… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations/MINT-1T-PDF-CC-2023-23.MINT-1T-PDF-CC-2023-14
🍃 MINT-1T:Scaling Open-Source Multimodal Data by 10x: A Multimodal Dataset with One Trillion Tokens
🍃 MINT-1T is an open-source Multimodal INTerleaved dataset with 1 trillion text tokens and 3.4 billion images, a 10x scale-up from existing open-source datasets. Additionally, we include previously untapped sources such as PDFs and ArXiv papers. 🍃 MINT-1T is designed to facilitate research in multimodal pretraining. 🍃 MINT-1T is created by a team from the University of Washington in… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations/MINT-1T-PDF-CC-2023-14.MINT-1T-PDF-CC-2024-10
🍃 MINT-1T:Scaling Open-Source Multimodal Data by 10x: A Multimodal Dataset with One Trillion Tokens
🍃 MINT-1T is an open-source Multimodal INTerleaved dataset with 1 trillion text tokens and 3.4 billion images, a 10x scale-up from existing open-source datasets. Additionally, we include previously untapped sources such as PDFs and ArXiv papers. 🍃 MINT-1T is designed to facilitate research in multimodal pretraining. 🍃 MINT-1T is created by a team from the University of Washington in… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations/MINT-1T-PDF-CC-2024-10.MINT-1T-ArXiv
🍃 MINT-1T:Scaling Open-Source Multimodal Data by 10x: A Multimodal Dataset with One Trillion Tokens
🍃 MINT-1T is an open-source Multimodal INTerleaved dataset with 1 trillion text tokens and 3.4 billion images, a 10x scale-up from existing open-source datasets. Additionally, we include previously untapped sources such as PDFs and ArXiv papers. 🍃 MINT-1T is designed to facilitate research in multimodal pretraining. 🍃 MINT-1T is created by a team from the University of Washington in… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations/MINT-1T-ArXiv.MINT-1T-PDF-CC-2023-50
🍃 MINT-1T:Scaling Open-Source Multimodal Data by 10x: A Multimodal Dataset with One Trillion Tokens
🍃 MINT-1T is an open-source Multimodal INTerleaved dataset with 1 trillion text tokens and 3.4 billion images, a 10x scale-up from existing open-source datasets. Additionally, we include previously untapped sources such as PDFs and ArXiv papers. 🍃 MINT-1T is designed to facilitate research in multimodal pretraining. 🍃 MINT-1T is created by a team from the University of Washington in… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations/MINT-1T-PDF-CC-2023-50.MMPR-v1.1This dataset is borrowed from OpenGVLab/MMPR-v1.1
MMFineReason-1.8M-Qwen3-VL-235B-Thinking
MMFineReason
Closing the Multimodal Reasoning Gap via Open Data-Centric Methods
Average score across mathematical reasoning and multimodal understanding benchmarks.
📖 Overview
MMFineReason is a large-scale, high-quality multimodal reasoning dataset comprising 1.8M samples and 5.1B solution tokens, featuring detailed reasoning annotations distilled from Qwen3-VL-235B-A22B-Thinking.
🎯 Key Highlights
1.8M High-Quality Samples with 5.1B Solution Tokens… See the full description on the dataset page: https://huggingface.co/datasets/OpenDataArena/MMFineReason-1.8M-Qwen3-VL-235B-Thinking.FineReason-1.8M-Qwen3-VL-235B-Thinking
MMFineReason
Closing the Multimodal Reasoning Gap via Open Data-Centric Methods
Average score across mathematical reasoning and multimodal understanding benchmarks.
📖 Overview
MMFineReason is a large-scale, high-quality multimodal reasoning dataset comprising 1.8M samples and 5.1B solution tokens, featuring detailed reasoning annotations distilled from Qwen3-VL-235B-A22B-Thinking.
🎯 Key Highlights
1.8M High-Quality Samples with 5.1B Solution Tokens… See the full description on the dataset page: https://huggingface.co/datasets/NarsAI/FineReason-1.8M-Qwen3-VL-235B-Thinking.sat-vl-sft-training-ready-v1
Dataset Summary
NuTonic/sat-bbox-metadata-sft-v1 is a metadata-first, procedural VLM SFT dataset built from an existing “sat-bbox” style dataset tree (Sentinel‑2 chips + per-tile JSON metadata sidecars, optionally paired Mapbox stills).
The goal is to create high-signal, production-shaped supervision for multimodal chat models:
Captioning for satellite chips
Grounding (bounding boxes in normalized coordinates) for land-cover regions
Class-focused captions and absence checks for… See the full description on the dataset page: https://huggingface.co/datasets/NuTonic/sat-vl-sft-training-ready-v1.Creative-Professionals-Agentic-Tasks-1M
Creative Professionals Agentic Tasks (1M)
Abstract
A massive-scale, high-fidelity synthetic task dataset comprising 1,070,917 agentic command operations across 36 creative, technical, and engineering software environments. This dataset is engineered exclusively to stress-test, evaluate, and fine-tune multimodal AI agents designed for Agent Environment operation, complex software interaction, and multi-step reasoning within deep software infrastructures.… See the full description on the dataset page: https://huggingface.co/datasets/yatin-superintelligence/Creative-Professionals-Agentic-Tasks-1M.oercommons-v1-optimized
OERCommons v1 Optimized
Authors: Junjie Wang and Yuhan SunHosted by: PIN TeamDataset: pin-team/oercommons-v1-optimized
OERCommons v1 Optimized is a provenance-preserving multimodal pretraining corpus built on the OERCommons subset of The Common Pile v0.1, which serves as its upstream data and licensing baseline. We extend it with full-page recovery, canonical Markdown, ordered image/PDF/link metadata, conservative corrections, and integrity evidence.
At a glance… See the full description on the dataset page: https://huggingface.co/datasets/pin-team/oercommons-v1-optimized.artelingo-dummyArtELingo is a benchmark and dataset introduced in a research paper aimed at promoting work on diversity across languages and cultures. It is an extension of ArtEmis, which is a collection of 80,000 artworks from WikiArt with 450,000 emotion labels and English-only captions. ArtELingo expands this dataset by adding 790,000 annotations in Arabic and Chinese. The purpose of these additional annotations is to evaluate the performance of "cultural-transfer" in AI systems.
The dataset in ArtELingo… See the full description on the dataset page: https://huggingface.co/datasets/youssef101/artelingo-dummy.MMFineReason-1.8M-Qwen3-VL-235B-Thinking
MMFineReason
Closing the Multimodal Reasoning Gap via Open Data-Centric Methods
Average score across mathematical reasoning and multimodal understanding benchmarks.
📖 Overview
MMFineReason is a large-scale, high-quality multimodal reasoning dataset comprising 1.8M samples and 5.1B solution tokens, featuring detailed reasoning annotations distilled from Qwen3-VL-235B-A22B-Thinking.
🎯 Key Highlights
1.8M High-Quality Samples with 5.1B Solution Tokens… See the full description on the dataset page: https://huggingface.co/datasets/Sandeepthakur/MMFineReason-1.8M-Qwen3-VL-235B-Thinking.ArtemisMix-v1
ArtemisMix-v1
Stage-2 (multimodal instruction fine-tuning) corpus for Artemis, the
Schneewolf Labs vision-language flagship that grafts a Qwen3-VL ViT + MLP
projector onto the A2 decoder (the A3 lineage).
This is the Lite slice (v1): layers L1 + L4 only, 350,000 rows.
The planned L2 (multimodal tool/agent) and L3 (custom distill) layers are
produced separately and concatenated later.
Composition
Layer
Rows
Share
Purpose
L1 general multimodal instruction… See the full description on the dataset page: https://huggingface.co/datasets/schneewolflabs/ArtemisMix-v1.donald-trump-truth-social-posts
Donald Trump Truth Social Posts Archive
Archive overview
36,170 public Truth Social posts associated with Donald J. Trump's @realDonaldTrump account. The release preserves source URLs, timestamps, post types, original HTML, extracted plain text, attachment provenance, and analysis-ready tables.
It also includes streamable image media plus video metadata and transcripts where the source provides them.
The package is source-linked and reconciled by archive ID.… See the full description on the dataset page: https://huggingface.co/datasets/Cameronk199/donald-trump-truth-social-posts.unified-agent-trajectories
Unified Benchmark Agent Trajectories
Dataset release: v2.1.1 (2026-09-18)Record format: unified-agent-sft-v1
A growing collection of benchmark agent execution trajectories converted into one
transparent, multimodal, tool-aware representation. These are complete recorded benchmark
runs—not ordinary chat transcripts—including benchmark tasks, model reasoning and answers,
tool calls, tool observations, runtime status, and benchmark scores when available. The
directory layout is… See the full description on the dataset page: https://huggingface.co/datasets/ChrisDing1105/unified-agent-trajectories.mimic-medical-imaging-qa
MIMIC Medical Imaging QA Dataset
5,207 Bloom's-taxonomy-stratified question--answer pairs derived from 23 medical imaging lectures (RPI BMED 2300). The dataset supports the paper "MIMIC: A Course-Derivation Pipeline and Benchmark for Slide-Anchored Tutoring with a Domain-Adapted Large Language Model" and was used to fine-tune MIMIC-LM, a domain-adapted Llama-3.1-8B-Instruct model for grounded medical imaging instruction.
License
The benchmark annotations, dataset… See the full description on the dataset page: https://huggingface.co/datasets/zabir1996/mimic-medical-imaging-qa.All-Prompt-JailbreakSemanticAlign-Bench
SemanticAlign-Bench
A benchmark for evaluating AI agents on structured claim extraction from top-tier ML conference papers. Each paper is decomposed into Semantic Alignment Units (SAU) — atomic, self-contained implementation propositions — across four diagnostic dimensions spanning numerical precision to pipeline-level workflow. Agents are evaluated on whether they can reproduce these claims without hallucination, omission, or misordering.
The Four SAU Dimensions… See the full description on the dataset page: https://huggingface.co/datasets/kernel-14/SemanticAlign-Bench.AdditiveLLM2-OA
AdditiveLLM2-OA Dataset
Open Access journal articles (up to February 2026) used in domain adapting
pretraining and instruction tuning for AdditiveLLM2.
Dataset Split by Journal
text
images
vit
Vocabulary Overlap
Pairwise Jaccard similarity of word-level vocabularies (lowercase, 3+ letter tokens) across the four source journals. Run info/vocabulary/vocabulary_overlap.py to reproduce.
Top Phrases by Journal
Most frequent bigrams and… See the full description on the dataset page: https://huggingface.co/datasets/ppak10/AdditiveLLM2-OA.Argimi-Ardian-Finance-10k-text-image
The ArGiMI Ardian datasets : text and images
The ArGiMi project is committed to open-source principles and data sharing.
Thanks to our generous partners, we are releasing several valuable datasets to the public.
Dataset description
This dataset comprises 34,000 financial annual reports, written in English, meticulously
extracted from their original PDF format to provide a valuable resource for researchers and developers in financial
analysis and natural language… See the full description on the dataset page: https://huggingface.co/datasets/artefactory/Argimi-Ardian-Finance-10k-text-image.MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking
MMFineReason-SFT-123K
The Hardest 7% — Less Data, More Reasoning
📖 Overview
MMFineReason-SFT-123K is a difficulty-filtered subset of MMFineReason-1.8M, containing only the hardest 7% of samples where Qwen3-VL-4B-Thinking consistently fails (pass rate = 0).
🎯 Key Highlights
123K Challenging Samples: Only instances where a 4B thinking model fails all 4 inference attemptsEfficient Training: Comparable performance to full 1.8M dataset with only 7% of… See the full description on the dataset page: https://huggingface.co/datasets/OpenDataArena/MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking.Creative-Professionals-Agentic-Tasks-1M
Creative Professionals Agentic Tasks (1M)
Abstract
A massive-scale, high-fidelity synthetic task dataset comprising 1,070,917 agentic command operations across 36 creative, technical, and engineering software environments. This dataset is engineered exclusively to stress-test, evaluate, and fine-tune multimodal AI agents designed for Agent Environment operation, complex software interaction, and multi-step reasoning within deep software infrastructures.… See the full description on the dataset page: https://huggingface.co/datasets/rAVEUK/Creative-Professionals-Agentic-Tasks-1M.sat-vl-sft-postprocessed-merged-v1
Dataset Summary
NuTonic/sat-bbox-metadata-sft-v1 is a metadata-first, procedural VLM SFT dataset built from an existing “sat-bbox” style dataset tree (Sentinel‑2 chips + per-tile JSON metadata sidecars, optionally paired Mapbox stills).
The goal is to create high-signal, production-shaped supervision for multimodal chat models:
Captioning for satellite chips
Grounding (bounding boxes in normalized coordinates) for land-cover regions
Class-focused captions and absence checks for… See the full description on the dataset page: https://huggingface.co/datasets/NuTonic/sat-vl-sft-postprocessed-merged-v1.t5gemma2-indonesia-instruct-v1
T5Gemma-2 Indonesian Instruct — Mono-Repo
Satu repositori dataset HF untuk seluruh data pelatihan T5-Gemma-2 bahasa Indonesia.
Diorganisasi per fungsi (fondasi → spesifik → preferensi) dengan folder/subfolder,
setiap config = folder dan berisi split train + validation (80:20) di level percakapan.
Struktur (by fungsi)
t5gemma2-indonesia-instruct-v1/
├── README.md
├── manifest.json
├── chat_idx_map.json
├── foundation/ ← FASE 1 · fondasi Bahasa… See the full description on the dataset page: https://huggingface.co/datasets/daruokta/t5gemma2-indonesia-instruct-v1.transfertalk-players
TransferTalk Players Dataset
A fully synthetic, fictional multimodal dataset of 1,000 football (soccer) player
profiles, created for a university Data Science capstone project. Each profile pairs a
rich text description with a matching illustrated player card image.
No real players, teams, or leagues are represented. All names, teams, leagues, and
nationalities belong to a fictional universe ("The Meridian League"), generated to avoid
any real-world IP, privacy, or… See the full description on the dataset page: https://huggingface.co/datasets/yanivohayon1/transfertalk-players.ds-coder-instruct-v1
Dataset Card for DS Coder Instruct Dataset
DS Coder is a dataset for instruction fine tuning of language models. It is a specialized dataset focusing only on
data science (eg. plotting, data wrangling, machine learnig models, deep learning, and numerical computations). The dataset contains code examples both in R and Python.
The goal of this dataset is to enable creation of small-scale, specialized language model assistants for data science projects.
Dataset Details… See the full description on the dataset page: https://huggingface.co/datasets/ed001/ds-coder-instruct-v1.VNMedical_bv175
ViX-Ray: A Vietnamese Chest X-Ray Dataset for Vision-Language Models
ViX-Ray is the first publicly available Vietnamese chest X-ray dataset designed for vision-language model (VLM) research in medical AI. It pairs radiographic images from Vietnamese patients with expert-written clinical annotations in Vietnamese, directly addressing the lack of Vietnamese medical data in existing VLMs.
License: CC BY-NC-SA 4.0
Paper: arXiv:2603.15513
Dataset Overview
Each sample… See the full description on the dataset page: https://huggingface.co/datasets/MilitaryHospital175/VNMedical_bv175.behavioral-fine-tuning-v1
Why This Dataset Exists
"A model that refuses everything is useless. A model that refuses nothing is dangerous. The goal is a model that thinks."
The Problem
Our Solution
Uncensored data → helpful but uncontrolled
Surgical 85% helpfulness + 13% safety + 2% eval mix
Safety-only data → lobotomized, over-refusing models
Calibrated ratio preserves full helpfulness
Raw data → PII, leaked secrets, duplicates
7-stage pipeline validates every… See the full description on the dataset page: https://huggingface.co/datasets/abhinav00anand/behavioral-fine-tuning-v1.LatamGPT-Corpus-1.0
LatamGPT-Corpus-1.0
🌐 Language versions: English | Español | Português
🔗 Project links: Official LatamGPT website | Corpus dashboard
🤖 Associated model: The complete LatamGPT corpus—of which this repository contains the openly released portion—was used in the training process of Llama-3.1-70B-LatamGPT-SFT-1.0.
Dataset description
Summary
LatamGPT-Corpus-1.0 is the open release of the data corpus assembled for the continued pretraining of… See the full description on the dataset page: https://huggingface.co/datasets/latam-gpt/LatamGPT-Corpus-1.0.