datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
MMLU-Pro
MMLU-Pro Dataset
MMLU-Pro dataset is a more robust and challenging massive multi-task understanding dataset tailored to more rigorously benchmark large language models' capabilities. This dataset contains 12K complex questions across various disciplines.
|Github | 🏆Leaderboard | 📖Paper |
🚀 What's New
[2026.03.11] Added more cutting-edge frontier models to the leaderboard, including the Claude-4.6 series, Seed2.0 series, Qwen3.5 series, and Gemini-3.1-Pro… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/MMLU-Pro.WebInstructSub
🦣 MAmmoTH2: Scaling Instructions from the Web
Project Page: https://tiger-ai-lab.github.io/MAmmoTH2/
Paper: https://arxiv.org/pdf/2405.03548
Code: https://github.com/TIGER-AI-Lab/MAmmoTH2
WebInstruct (Subset)
This repo contains the partial dataset used in "MAmmoTH2: Scaling Instructions from the Web". This partial data is coming mostly from the forums like stackexchange. This subset contains very high-quality data to boost LLM performance through instruction tuning.… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/WebInstructSub.TheoremQA
Dataset Card for "TheoremQA"
Introduction
We propose the first question-answering dataset driven by STEM theorems. We annotated 800 QA pairs covering 350+ theorems spanning across Math, EE&CS, Physics and Finance. The dataset is collected by human experts with very high quality. We provide the dataset as a new benchmark to test the limit of large language models to apply theorems to solve challenging university-level questions. We provide a pipeline in the following to… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/TheoremQA.VisualWebInstruct
VisualWebInstruct: Scaling up Multimodal Instruction Data through Web Search
VisualWebInstruct is a large-scale, diverse multimodal instruction dataset designed to enhance vision-language models' reasoning capabilities. The dataset contains approximately 900K question-answer (QA) pairs, with 40% consisting of visual QA pairs associated with 163,743 unique images, while the remaining 60% are text-only QA pairs.
Please also checkout our more recent verified version at Huggingface.… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/VisualWebInstruct.HRVideoBench
HRVideoBench
This repo contains the test data for HRVideoBench, which is released under the paper "VISTA: Enhancing Long-Duration and High-Resolution Video Understanding by Video Spatiotemporal Augmentation". VISTA is a video spatiotemporal augmentation method that generates long-duration and high-resolution video instruction-following data to enhance the video understanding capabilities of video LMMs.
🌐 Homepage | 📖 arXiv | 💻 GitHub | 🤗 VISTA-400K | 🤗 Models | 🤗 HRVideoBench… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/HRVideoBench.VisualWebInstruct-Recall
Introduction
This is the dataset recalled from Google Search from the seed images.
Links
Github|
Paper|
Website
Citation
@article{visualwebinstruct,
title={VisualWebInstruct: Scaling up Multimodal Instruction Data through Web Search},
author = {Jia, Yiming and Li, Jiachen and Yue, Xiang and Li, Bo and Nie, Ping and Zou, Kai and Chen, Wenhu},
journal={arXiv preprint arXiv:2503.10582},
year={2025}
}
WebInstruct-verified
General-Reasoner: Advancing LLM Reasoning Across All Domains
💻 Code |
📄 Paper |
📊 Dataset |
🤗 Model |
🌐 Project Page
Overview
Figure: Effectiveness of General-Reasoner trained with diverse verifiable reasoning questions using model-based verifier compared to baseline methods on various reasoning tasks.
General-Reasoner is a training paradigm for large language models (LLMs), designed to robustly enhance reasoning abilities across… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/WebInstruct-verified.VISTA-400K
VISTA-400K
This repo contains all subsets for VISTA-400K. VISTA is a video spatiotemporal augmentation method that generates long-duration and high-resolution video instruction-following data to enhance the video understanding capabilities of video LMMs.
This repo is under construction. Please stay tuned.
🌐 Homepage | 📖 arXiv | 💻 GitHub | 🤗 VISTA-400K | 🤗 Models | 🤗 HRVideoBench
Video Instruction Data Synthesis Pipeline
VISTA leverages insights from… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/VISTA-400K.VisualWebInstruct-verified
🧠 VisualWebInstruct-Verified: High-Confidence Multimodal QA for Reinforcement Learning
VisualWebInstruct-Verified is a high-confidence subset of VisualWebInstruct, curated specifically for Reinforcement Learning (RL) and Reward Model training.
It contains verified multimodal question–answer pairs where correctness, reasoning quality, and image–text alignment have been explicitly validated.
This dataset is ideal for RLVR training pipelines.
📘 Dataset Overview… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/VisualWebInstruct-verified.VisualWebInstruct-Seed
Introduction
This is the seed dataset we used to conduct Google Search.
Links
Github|
Paper|
Website
Citation
@article{visualwebinstruct,
title={VisualWebInstruct: Scaling up Multimodal Instruction Data through Web Search},
author = {Jia, Yiming and Li, Jiachen and Yue, Xiang and Li, Bo and Nie, Ping and Zou, Kai and Chen, Wenhu},
journal={arXiv preprint arXiv:2503.10582},
year={2025}
}
Mantis-Eval
Overview
This is a newly curated dataset to evaluate multimodal language models' capability to reason over multiple images. More details are shown in https://tiger-ai-lab.github.io/Mantis/.
Statistics
This evaluation dataset contains 217 human-annotated challenging multi-image reasoning problems.
Leaderboard
We list the current results as follows:
Models
Size
Mantis-Eval
LLaVA OneVision
72B
77.60
LLaVA OneVision
7B
64.20
GPT-4V
-
62.67… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/Mantis-Eval.mle-benchSimple fork of MLE-Bench Lite.
cat * > data.zip
unzip data.zip
Then move all the folders to .cache/mle-bench/
ClawBench
ClawBench Dataset
ClawBench is an open benchmark for AI web agents — the systems that drive a real browser to complete a user's task end-to-end. It scores agents on real, everyday online tasks (booking flights, ordering groceries, submitting job applications) across live websites.
|💻 Github | 🏆 Leaderboard | 📖 Paper | 🌐 Website |
🚀 What's New
[2026.05.12] Added the V2 corpus (130 newer tasks across 63 platforms) and 7 new models judged with… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/ClawBench.ViRL39K
1. Overview of ViRL39K
ViRL39K (pronounced as "viral") provides a curated collection of 38,870 verifiable QAs for Vision-Language RL training.
It is built on top of newly collected problems and existing datasets (
Llava-OneVision,
R1-OneVision,
MM-Eureka,
MM-Math,
M3CoT,
DeepScaleR,
MV-Math)
through cleaning, reformatting, rephrasing and verification.ViRL39K lays the foundation for SoTA Vision-Language Reasoning Model VL-Rethinker. It has the following merits:
high-quality and… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/ViRL39K.SWE-QA-Pro-Bench
SWE-QA-Pro Bench (A Repository-level QA Benchmark Built from Diverse Long-tail Repositories)
💻 GitHub | 📖 Paper | 🤗 SWE-QA-Pro
📢 News
🚀 [2026-5-19] The evaluation code is released on GitHub.
🔥 [2026-3-23] SWE-QA-Pro Bench is publicly released! The model and code will be released soon.
Introduction
SWE-QA-Pro Bench is a repository-level question answering dataset designed to evaluate whether models can perform grounded, agentic reasoning… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/SWE-QA-Pro-Bench.AIME25The AIME25 part 1 exam from the website.
ScreenSpot-ProSimplied Version of ScreenSpot-Pro dataset.
MEGA-Bench
MEGA-Bench: Scaling Multimodal Evaluation to over 500 Real-World Tasks [ICLR 2025]
🌐 Homepage | 🏆 Leaderboard | 🤗 Dataset | 🤗 Paper | 🔎 Visualiaztion | 📖 arXiv | GitHub
🔔 News
[2025-01]: Paper accepted by ICLR 2025.
[2024-10-18]: Initial release of the evaluation code on our Github repo.
[2024-10-14]: Paper released on arXiv.
❗❗ Data Information
We put the file path of images/videos in HF datasets. Please download the zipped data here.
We chose… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/MEGA-Bench.VideoFeedback2📃Paper |
🌐Website |
💻Code |
🛢️Dataset (VideoFeedback2) |
🤗Model (VideoScore2) |
🤗Space (VideoScore2) |
🤗50K videos cache
Overview
VideoFeedback2 is a large-scale, human-annotated dataset designed for training and evaluating multi-dimensional video evaluator 🤗Model (VideoScore2). It contains 27,168 AI-generated videos paired with fine-grained human feedback scores and reasoning traces across three evaluation dimensions:
(1) Visual Quality; (2) Text Alignment; (3)… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/VideoFeedback2.WebInstructFull
🦣 MAmmoTH2: Scaling Instructions from the Web
Project Page: https://tiger-ai-lab.github.io/MAmmoTH2/
Paper: https://arxiv.org/pdf/2405.03548
Code: https://github.com/TIGER-AI-Lab/MAmmoTH2
WebInstruct (Full)
This repo the entire dataset used in "MAmmoTH2: Scaling Instructions from the Web". The dataset is licensed under CC BY-NC 4.0, which means that the dataset can only be used for non-commercial purposes; any use that involves profit-making or commercial advantage is… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/WebInstructFull.One-Shot-CFT-Data
One-Shot-CFT: Unleashing the Reasoning Potential of Pre-trained LLMs by Critique Fine-Tuning on One Problem
💻 Code |
📄 Paper |
📊 Dataset |
🤗 Model |
🌐 Project Page
🧠 Overview
One-Shot Critique Fine-Tuning (CFT) is a simple, robust, and compute-efficient training paradigm for unleashing the reasoning capabilities of pretrained LLMs in both mathematical and logical domains. By leveraging critiques on just one problem, One-Shot CFT enables models… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/One-Shot-CFT-Data.Fineweb-InstructWe convert the pre-training corpus from Fineweb-Edu (https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) to instruction following format. We select a subset with quality filter and then use GPT-4 to extract instruction-following pairs. The dataset contains roughly 16M instruction pairs. The basic concept is similar to MAmmoTH2 (https://arxiv.org/abs/2405.03548).
Citation
If you use dataset useful, please cite the following paper:
@article{yue2024mammoth2… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/Fineweb-Instruct.SWE-QA-Pro-SFT-Trajectories
SWE-QA-Pro SFT Trajectories
💻 GitHub | 📖 Paper | 🤗 SWE-QA-Pro
Introduction
SWE-QA-Pro SFT Trajectories is a set of agentic tool-use trajectories for repository-level question answering, used as the supervised fine-tuning (SFT) data in the SWE-QA-Pro training recipe.
Each item is a multi-turn trajectory in which an agent answers a repository-grounded question by exploring the codebase with read-only tools rather than relying on memorized knowledge. The… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/SWE-QA-Pro-SFT-Trajectories.BrowserAgent-SeedData
BrowserAgent-Data
Dataset used in https://github.com/TIGER-AI-Lab/BrowserAgent.
Summary
Total rows: 230,015
Total size: ~29.47 MB
Subsets: 2wiki, bamboogle, hotpot, musique, nq, popqa
Splits and Sizes
2wiki: 22,576 rows (~2.73 MB)
bamboogle: 125 rows (~0.03 MB)
hotpot: 97,852 rows (~14.20 MB)
musique: 12,417 rows (~0.92 MB)
nq: 82,778 rows (~10.14 MB)
popqa: 14,267 rows (~1.45 MB)
Files are stored as Parquet under each subset directory with dev/train/test… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/BrowserAgent-SeedData.WebInstruct-verified-unfilteredThis repo contains the unfiltered version WebInstruct-verified in the General Reasoner work.
General-Reasoner: Advancing LLM Reasoning Across All Domains
💻 Code |
📄 Paper |
📊 Dataset |
🤗 Model |
🌐 Project Page
Overview
Figure: Effectiveness of General-Reasoner trained with diverse verifiable reasoning questions using model-based verifier compared to baseline methods on various reasoning tasks.
General-Reasoner is a training paradigm… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/WebInstruct-verified-unfiltered.AceCode-V2-122K
🂡 AceCode-V2-122K
Paper |
Github
RM/RL Models
AceCoder-V2-122k is an improved version of V1, TIGER-Lab/AceCode-87K. Each question and test cases in the dataset is rewritten by OpenAI's o1-mini, then filtered using Qwen Coder 2.5 32B Instruct.
Dataset statistics of AceCoderV2-123K before and after test-case filtering
Subset
Evol
OSS
Stack Python
Overall
Before Filtering
# Questions
41,548
35,933
69,739
147,220
# Avg Test Cases
20.51
22.04
20.56… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/AceCode-V2-122K.tigqa-extractive-qa
TIGQA — Extractive QA Subset
Dataset Summary
This repository contains a cleaned, extractive-QA-only (SQuAD-style) subset of the TIGQA
question–answering data: 1,665 question–answer pairs in Tigrinya, each with a single answer
located as an exact character span inside its context passage (answer_start and the span
length), ready for direct use in span-extraction QA training and evaluation.
Every row carries a match_status field (exact or fuzzy) recording how… See the full description on the dataset page: https://huggingface.co/datasets/Hailay/tigqa-extractive-qa.tigrinya-squad
Tigrinya-SQuAD: Machine-Translated Training Dataset
Tigrinya-SQuAD is a machine-translated and filtered version of the English SQuAD 1.1 training dataset, automatically converted to Tigrinya for training question-answering models in low-resource settings.
This silver dataset serves as training data for Tigrinya question-answering systems. For evaluation and benchmarking, please use the gold-standard TiQuAD dataset, which contains human-annotated validation and test sets.
Published… See the full description on the dataset page: https://huggingface.co/datasets/fgaim/tigrinya-squad.Code-170k-tigrinya
Dataset Description
Code-170k-tigrinya is a groundbreaking dataset containing 176,999 programming conversations, originally sourced from glaiveai/glaive-code-assistant-v2 and translated into Tigrinya, making coding education accessible to Tigrinya speakers.
🌟 Key Features
176,999 high-quality conversations about programming and coding
Pure Tigrinya language - democratizing coding education
Multi-turn dialogues covering various programming concepts
Diverse topics:… See the full description on the dataset page: https://huggingface.co/datasets/michsethowusu/Code-170k-tigrinya.tigle-source-code
TIGLE
The interface is built as a prototype based on Dzogchen, Atiyoga teachings available in English and sourced, compiled by a practitioner exploring how Dharma language and current global AI could intersect. The architecture, the pipeline works. The answers are useful for orientation — learning key terms, lineages, main practices, understanding the view.
This is why it is accessible as repository rather than a product:
Digital Bardo - is the current state of samsara.… See the full description on the dataset page: https://huggingface.co/datasets/Tigle/tigle-source-code.