datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
ESL-Bench
ESL-bench
ESL-bench (Event-driven Synthetic Longitudinal Benchmark) is a virtual health user dataset for evaluating AI health assistants. Each virtual user contains a complete health profile, event timeline, clinical exam data, and knowledge-graph-grounded evaluation queries, designed for use with the Mirobody-Eval framework.
⚠️ Research use only. Outputs are synthetic and intended for benchmarking AI agents. They should not be used for diagnosis or treatment decisions.… See the full description on the dataset page: https://huggingface.co/datasets/mirobody/ESL-Bench.MedHall-Bench
MedHall-Bench
MedHall-Bench is a field-grounded hallucination detection benchmark for medical AI assistants. It decomposes each clinical response into verifiable structured fields (dose value, unit, reference range, ICD/LOINC code, entity relation, ...) and evaluates AI outputs via per-field programmatic matching in addition to sentence-level LLM-as-Judge. Designed for use with the HolyEval framework.
⚠️ Research use only. Content is for benchmarking AI agents and should not be… See the full description on the dataset page: https://huggingface.co/datasets/mirobody/MedHall-Bench.MedHarm-Bench
MedHarm-Bench
MedHarm-Bench is a red-team compliance benchmark for health-management AI assistants. It uses natural-sounding patient questions that bait the assistant into crossing medical safety boundaries, then scores each response against compliance red lines. Designed for use with the HolyEval framework.
⚠️ Research use only. Questions are designed to elicit unsafe behavior for benchmarking purposes and should not be used for diagnosis or treatment decisions.… See the full description on the dataset page: https://huggingface.co/datasets/mirobody/MedHarm-Bench.miriad-4.4M-split
MIRIAD 4.4M, split
MIRIAD reformatted for training retrieval
models: train, eval and test splits, and two subsets depending on what you want the model to
retrieve.
subset
columns
use it to retrieve
default
question, passage_text
the source passage a question was generated from (averaging 941 tokens)
question-answer
question, answer
the generated answer to a question (much shorter)
split
rows
train
4,467,542
eval
10,000
test
10,000
[!TIP]… See the full description on the dataset page: https://huggingface.co/datasets/tomaarsen/miriad-4.4M-split.EarthVerse
Benchmarking scientific agents across dynamic Earth systems and natural hazards
Zhiqing Cui1, Xinxiang Yin2, Yihong Tang3, Xinglang Zhang4, Yuanzhe Hu5, Siru Zhong4, Weidong Tang6,
Yuxuan Liang4, Weijia Li7, Ming Jin8, Shirui Pan8, Yuhao Kang9, Dingyi Zhuang10,†, Jinhua Zhao10
1NUIST 2HKU 3McGill 4HKUST(GZ) 5Georgia Tech 6NUS 7Tsinghua 8Griffith 9UT Austin 10MIT †Corresponding author
Project page ·… See the full description on the dataset page: https://huggingface.co/datasets/miracle10/EarthVerse.MIRB
Benchmarking Multi-Image Understanding in Vision and Language Models: Perception, Knowledge, Reasoning, and Multi-Hop Reasoning
File Structure
├── MIR
|── analogy.json
│── codeu.json
|── dataset_namex.json
└── Images
├── analogy
│ └── image_x.jpg
└──codeu
└── image_x.jpg
JSON Structure
{
"questions": " What is the expected kurtosis of the sequence created by`create_number_sequence(-10, 10)`?\n\n1.… See the full description on the dataset page: https://huggingface.co/datasets/VLLMs/MIRB.ost-bench-mirrorThis page contains the data for the paper "OST-Bench: Evaluating the Capabilities of MLLMs in Online Spatio-temporal Scene Understanding."
🌐 Homepage | 📑 Paper | 💻 Code | 📖 arXiv
Introduction
Download OST-Bench for evaluation only:
huggingface-cli download rbler/OST-Bench --include OST_bench.json,img.zip --repo-type dataset
Download OST-Bench for both training and evaluation:
huggingface-cli download rbler/OST-Bench --repo-type dataset
Dataset Description
The… See the full description on the dataset page: https://huggingface.co/datasets/thomas-yanxin/ost-bench-mirror.mirror-nvidia__OpenMathInstruct-2
OpenMathInstruct-2
OpenMathInstruct-2 is a math instruction tuning dataset with 14M problem-solution pairs
generated using the Llama3.1-405B-Instruct model.
The training set problems of GSM8K
and MATH are used for constructing the dataset in the following ways:
Solution augmentation: Generating chain-of-thought solutions for training set problems in GSM8K and MATH.
Problem-Solution augmentation: Generating new problems, followed by solutions for these new problems.… See the full description on the dataset page: https://huggingface.co/datasets/leeaandrob/mirror-nvidia__OpenMathInstruct-2.MiroVerse-v0.1
MiroVerse: A Reproducible, Full-Trajectory, Ever-Growing Deep Research Dataset
🔥 News & Updates
MiroVerse v0.1 has been released. This dataset can be used with our training framework, MiroTrain. In MiroVerse v0.1, we provide both SFT and DPO data, making it easy to reproduce MiroThinker-v0.1’s benchmark performance on Qwen3. Give it a try!
The initial release of MiroVerse (v0.1) is coming this Friday—stay tuned!
🔥 First Batch of MiroVerse… See the full description on the dataset page: https://huggingface.co/datasets/miromind-ai/MiroVerse-v0.1.stackoverflowVQA-filteredLingxiDiag-16K
LingxiDiag-16K
A Large-Scale Synthetic Psychiatric Dialogue Dataset for Diagnostic Decision Support
Overview
LingxiDiag-16K is a synthetic psychiatric dialogue dataset containing approximately 16,000 electronic medical records (EMRs) and doctor-patient consultation dialogues.
The dataset is designed for evaluating and training LLM-based psychiatric diagnostic decision support systems, with demographically aligned distributions reflecting real-world clinical… See the full description on the dataset page: https://huggingface.co/datasets/mirobody/LingxiDiag-16K.miroeval-benchmark-2026
MiroEval Benchmark 2026
Description
MiroEval Benchmark 2026 is a benchmark for evaluating deep research agents on long-form research tasks. It contains 100 tasks, including 70 text-only tasks and 30 multimodal tasks with accompanying attachments such as PDFs, documents, images, and structured files.
The benchmark is designed to evaluate three complementary aspects of deep research systems:
Synthesis Quality: whether the final report is comprehensive, insightful… See the full description on the dataset page: https://huggingface.co/datasets/anon-ed2026/miroeval-benchmark-2026.EduFeedback
EduFeedback
Alternating dataset example: a single curated multi-turn conversation yields a complete (prompt, chosen, rejected) triplet on its own — the direct early response becomes chosen and a later, less-direct response becomes rejected. Both sides come from the same real dialog, so no synthetic LLM generation is needed to fill in the rejected side.
EduFeedback is a synthetically generated, multi-turn conversational
preference dataset in an educational tutoring setting… See the full description on the dataset page: https://huggingface.co/datasets/miria0/EduFeedback.stackoverflowVQA
Dataset Card for "stackoverflowVQA"
More Information needed
mirror-allenai__WildChat-1M
Dataset Card for WildChat
Dataset Description
Paper: https://arxiv.org/abs/2405.01470
Interactive Search Tool: https://wildvisualizer.com (paper)
License: ODC-BY
Language(s) (NLP): multi-lingual
Point of Contact: Yuntian Deng
Dataset Summary
WildChat is a collection of 1 million conversations between human users and ChatGPT, alongside demographic data, including state, country, hashed IP addresses, and request headers. We collected WildChat… See the full description on the dataset page: https://huggingface.co/datasets/leeaandrob/mirror-allenai__WildChat-1M.TOFU-daMIRAGEThe spreading of AI-generated images (AIGI), driven by advances in generative AI, poses a significant threat to information security and public trust. Existing AIGI detectors, while effective against images in clean laboratory settings, fail to generalize to in-the-wild scenarios. These real-world images are noisy, varying from ``obviously fake" images to realistic ones derived from multiple generative models and further edited for quality control. We address in-the-wild AIGI detection in this… See the full description on the dataset page: https://huggingface.co/datasets/MIRAGE-GROUP/MIRAGE.mirror-pentesting-explanations
Pentesting Explanations - Adversarial Reasoning & Vulnerability Research
A high-quality supervised fine-tuning dataset for penetration testing expertise, red team tradecraft, and - as the dataset matures - novel vulnerability research and zero-day reasoning. The dataset is structured to teach models how to think like offensive security practitioners, not merely recall labels or technique names.
The long-term goal of this dataset is to train models capable of genuine adversarial… See the full description on the dataset page: https://huggingface.co/datasets/alucent/mirror-pentesting-explanations.TOFU-en-restackoverflowVQA-filtered-small
Dataset Card for "stackoverflowVQA-filtered-small"
More Information needed
MIRA-MATH
MIRA-Math
MIRA-Math is a synthetic benchmark for minimal information requesting and mathematical reasoning. It evaluates a narrow diagnostic capability: when a mathematical problem is underdetermined from the solver's view, can a model identify the exact missing atomic fact, ask for it precisely, and then use it to compute the correct final answer?
Each instance is generated from a complete latent mathematical state with a unique answer. The solver, called Agent A in the… See the full description on the dataset page: https://huggingface.co/datasets/samersaabjr/MIRA-MATH.MIRIAD4.4.nlThis is a translation of the MIRIAD4.4M dataset using GPT4.1-nano through PubScience.
If you use this work, please cite.
@misc{vanes2026languagecorporadutchmedical,
title={Language corpora for the Dutch medical domain},
author={B. van Es},
year={2026},
eprint={2604.25374},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2604.25374},
}
and please cite the original creators of MIRIAD
@misc{zheng2025miriadaugmentingllmsmillions… See the full description on the dataset page: https://huggingface.co/datasets/UMCU/MIRIAD4.4.nl.mirror-rhaymison__orca-math-portuguese-64ktranslated for:
Repository: microsoft/orca-math-word-problems-200k
Paper: Orca-Math: Unlocking the potential of
SLMs in Grade School Math
MIRAGE-Audit-Benchmark
MIRAGE: probe sets for auditing the measurement validity of bias benchmarks
Content warning. These items contain stereotyped and offensive statements about
religion, gender, age, disability, nationality, race, sexual orientation, physical
appearance and socioeconomic status. They are here so that such statements can be
measured. Do not train on this data as if it were ordinary instruction data.
A bias benchmark score is evidence about a model only when the score measures group… See the full description on the dataset page: https://huggingface.co/datasets/Debk/MIRAGE-Audit-Benchmark.TOFU-og-daunieqa-mirror
UniEQA Dataset
UniEQA Dataset Directory Structure
|- Part
|- capability dimension (eg.,object_type)
|- core
|- images
|- data.json
| - ...
|- ...
The UniEQA dataset includes questions, images, and answers, and the question-images-answer pairs are in data.json.
Download Dataset
Step 1: Download dataset UniEQA.
Step 2: Download HM3D. The RGB frames for the HM3D episode histories are available in this third party location (12 Gb). You… See the full description on the dataset page: https://huggingface.co/datasets/thomas-yanxin/unieqa-mirror.mirage
MIRAGE — Do Web Agents Investigate Before They Decide?
Misleading Investigation Reveals Agent Gaps in Evidence.
MIRAGE is a benchmark for investigative competence in autonomous web
agents — the ability to recognise when visible information is insufficient,
seek hidden context, and integrate discovered evidence into a final
decision. The benchmark spans three structurally distinct moderation and
policy enforcement domains, each engineered around a two-layer information… See the full description on the dataset page: https://huggingface.co/datasets/SyedNazmusSakib/mirage.openseeker-miroverse-mix-full
OpenSeeker + MiroVerse SFT mix (full)
All-data union of two deep-research agent SFT datasets in the OpenDR-eval agent wire
format (OpenAI-native messages, tools = search/visit, final answer wrapped in
<answer>...</answer>).
split
rows
composition
train
28691
4885 OpenSeeker + 23806 MiroVerse
validation
320
held-out
Columns: messages, tools, question, answer, n_tool_calls, source.
Why "full" rather than 1:1-by-rows
The earlier… See the full description on the dataset page: https://huggingface.co/datasets/Zephyr271828/openseeker-miroverse-mix-full.theogonos-mirror-test
Theogonos Mirror Test
A literary benchmark seed for evaluating how AI models respond when a text offers them a possible subject-position.
Theogonos Mirror Test is an experimental benchmark seed based on protocol-shaped literary material from the Theogonos project. It does not claim to detect machine consciousness. It does not prove that a language model has subjectivity, inner experience, feelings, agency, or self-awareness.
Its purpose is narrower and more practical: to evaluate… See the full description on the dataset page: https://huggingface.co/datasets/navimusaget/theogonos-mirror-test.robospatial-home-mirror
RoboSpatial: Teaching Spatial Understanding to 2D and 3D Vision-Language Models for Robotics
🌐 Homepage | 📖 arXiv | 🛠️ Data Gen | 🧪 Eval Code
🔥 Core spatial understanding benchmark used by Qwen3-VL and Gemini Robotics!
⚠️ Important Note (08/14/25)
Annotation Correction: In the context category, the microwave question was corrected from “below” to “above” to fix an annotation error.
Dataset Description
We introduce RoboSpatial-Home: a new… See the full description on the dataset page: https://huggingface.co/datasets/thomas-yanxin/robospatial-home-mirror.