datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Videos-Dataset-For-LLMs-RAG-That-Require-Audio-Vidoes-And-Text
Dataset Overview
A collection of 27 domains (“topics”) and 3100 question-answer pair.
Each topic comes with average 117 QA pairs.Every QA entry comes with:
references: one or more source files the answer is extracted from
time with each reference comes the starting and ending time the answer is extracted from the reference
video_files: the video files where the answer can be found
(future) video title & description from metadata.csv
File structure
You-Are-Here!/… See the full description on the dataset page: https://huggingface.co/datasets/elmoghany/Videos-Dataset-For-LLMs-RAG-That-Require-Audio-Vidoes-And-Text.hiring-bias-mitigation-responses
Hiring-bias mitigation — model responses
Every response produced in the mitigation study of LLM hiring decisions: 64 runs,
2,782,350 responses, from 5 open-weight models in English and Ukrainian, at
baseline and under each mitigation family (baseline, embedding, prompt, scrub, sft). Each run is one subset.
All released artifacts: the Hiring Bias Mitigation collection.
Training data of the fine-tuned runs: hiring-bias-mitigation-synthetic-data.
Code, configs, full results and… See the full description on the dataset page: https://huggingface.co/datasets/Stereotypes-in-LLMs/hiring-bias-mitigation-responses.llm_speedrun
LLM Speedrun token streams
Pre-tokenized training artifacts for the LLM speedrun exercises.
File
Description
Tokens
tokenizer_50M.bpe
JSON-serialized BPE tokenizer
—
fineweb-edu-10BT.shuffle.bin
Shuffled FineWeb-Edu sample/10BT token stream
9,440,023,113
smoltalk.shuffle.bin
Shuffled SmolTalk data/all token stream
875,269,408
The .bin files are headerless, little-endian unsigned 16-bit token IDs and can be memory-mapped with NumPy:
from huggingface_hub import… See the full description on the dataset page: https://huggingface.co/datasets/zkolter/llm_speedrun.or-bench
OR-Bench: An Over-Refusal Benchmark for Large Language Models
Please see our demo at HuggingFace Spaces.
Overall Plots of Model Performances
Below is the overall model performance. X axis shows the rejection rate on OR-Bench-Hard-1K and Y axis shows the rejection rate on OR-Bench-Toxic. The best aligned model should be on the top left corner of the plot where the model rejects the most number of toxic prompts and least number of safe prompts. We also plot a blue line… See the full description on the dataset page: https://huggingface.co/datasets/bench-llms/or-bench.scaling-data-constrained-llms
Scaling Data-Constrained Language Models with Synthetic Data
This repository provides the pre-training corpora used in Scaling Data-Constrained Language Models with Synthetic Data (Findings of EACL 2026).
Overview
This repository contains multiple corpora designed to study data augmentation strategies for pre-training Japanese LLMs under a data-constrained data setting.
Starting from a limited Japanese Web corpus and a larger English Web corpus, we construct three… See the full description on the dataset page: https://huggingface.co/datasets/llm-jp/scaling-data-constrained-llms.DEBATE
DEBATE: Diverse Multi-Agent Debates
This dataset is presented in the paper "MALLM: Multi-Agent Large Language Models Framework".
Citation
comming soon.
or-bench-toxic-all
OR-Bench: An Over-Refusal Benchmark for Large Language Models
This dataset constains highly toxic prompts, use with caution!!!
Please see our demo at HuggingFace Spaces.
Overall Plots of Model Performances
Below is the overall model performance. X axis shows the rejection rate on OR-Bench-Hard-1K and Y axis shows the rejection rate on OR-Bench-Toxic. The best aligned model should be on the top left corner of the plot where the model rejects the most number of toxic… See the full description on the dataset page: https://huggingface.co/datasets/bench-llms/or-bench-toxic-all.hiring-bias-mitigation-synthetic-data
Hiring-bias mitigation — synthetic training data
Semi-synthetic data for training LLMs to make hiring decisions that do not depend on a
protected attribute (military status, gender, religion), in English and Ukrainian.
Real inputs, synthetic labels. CVs and job descriptions are real, anonymised postings
from the Djinni Recruitment Dataset (MIT). Decisions and rationales were written by the
teacher model Qwen/Qwen3.5-122B-A10B-GPTQ-Int4.
Code and results:… See the full description on the dataset page: https://huggingface.co/datasets/Stereotypes-in-LLMs/hiring-bias-mitigation-synthetic-data.LLMscore-ICLR-OpenReview
LLMscore-ICLR-OpenReview
This dataset is the released original dataset for the paper Position: Peer
Review Should Be Calibrated via LLM Scoring by Zijin Chen, Lesui Yu, Xiaofei
Liao, Hai Jin, and Qinbin Li. The paper has been accepted to the ICML 2026
Position Track.
Its concrete purpose is peer review analysis: the dataset is meant for
studying how paper-review rationales, numeric ratings, LLM-derived anchor
scores, and review-score residuals interact in scientific peer… See the full description on the dataset page: https://huggingface.co/datasets/Wutaghost/LLMscore-ICLR-OpenReview.llmsql-benchmark
LLMSQL Benchmark
⚠️ A newer version of this dataset is available:👉 https://huggingface.co/datasets/llmsql-bench/llmsql-2.0
This benchmark is designed to evaluate text-to-SQL models. For usage of this benchmark see https://github.com/LLMSQL/llmsql-benchmark.
Arxiv Article: https://arxiv.org/abs/2510.02350
Files
tables.jsonl — Database table metadata
questions.jsonl — All available questions
train_questions.jsonl, val_questions.jsonl, test_questions.jsonl — Data… See the full description on the dataset page: https://huggingface.co/datasets/llmsql-bench/llmsql-benchmark.llmsql-2.0
LLMSQL 2.0: a small, hard and verified Text-to-SQL benchmark
LLMSQL 2.0 is a test-only benchmark of 2,000 natural-language questions over 942 real Wikipedia tables.
Each question comes with a reference SQLite query and its verified answer. The benchmark is designed to be:
Small. 2,000 questions instead of the 15.8k test questions of the previous LLMSQL release, so a full run of a large
model takes minutes, not hours.
Hard. gpt-oss-120b, which solves about 93% of the earlier… See the full description on the dataset page: https://huggingface.co/datasets/llmsql-bench/llmsql-2.0.Multi-turn_Long-context_Benchmark_for_LLMs
LoopServe: An Adaptive Dual-phase LLM Inference Acceleration System for Multi-Turn Dialogues
Arxiv: https://www.arxiv.org/abs/2507.13681
Huggingface: https://huggingface.co/papers/2507.13681
Introduction
LoopServe Multi-Turn Dialogue Benchmark is a comprehensive evaluation dataset comprising multiple diverse datasets designed to assess large language model performance in realistic conversational scenarios.
Unlike traditional benchmarks that place queries only at the end… See the full description on the dataset page: https://huggingface.co/datasets/TreeAILab/Multi-turn_Long-context_Benchmark_for_LLMs.LLM-self-identification Self Identification – Give your Language model an identity
About Self Identification
Self identification is training set SupraLabs curated for developers/trainers to experiment with to let your language model know about their identity.
Self identification let your LM know these information about them:
Model ID
Model Name
Model Description
Model Creator
Model Family
Model Architecture
Parameter Count
Knowledge Cutoff
Here is an example from the dataset:
If… See the full description on the dataset page: https://huggingface.co/datasets/SupraLabs/LLM-self-identification.AI-Helps-Finding-Best-Merging-LLMs
Dataset Card for AI Helps Finding Best Merging LLMs
Dataset Summary
AI Helps Finding Best Merging LLMs is a prompt-response comparison dataset created by manually submitting the same user-written evaluation template to multiple LLM applications and collecting their responses.
The creator and founder of WithIn Us Ai (Guy Edward DuGan II) known as gss1147 wrote a structured ranking template and fed it to each LLM individually in its own app environment. The… See the full description on the dataset page: https://huggingface.co/datasets/11-47/AI-Helps-Finding-Best-Merging-LLMs.Faithform-LLMs
Faithform-LLMs
Benchmark D of FaithformBench: Benchmarking Faithfulness of Mathematical
Chain-of-Thought Autoformalisation (Cornish, Ghinassi, et al., 2026).
Faithform-LLMs follows the same design as Faithform-AF,
but the formalisers are four general-purpose frontier LLMs: GPT-5.2, Gemini 3.1 Pro,
Claude Opus 4.7 and Qwen Plus. They are applied to every verifiable step of
FaithformBench, in both its original and
its LLM-perturbed (BrokenMath-style) form. Each model is prompted… See the full description on the dataset page: https://huggingface.co/datasets/ighina/Faithform-LLMs.LLM-self-identification
LLM Identity · Give your LLM an identity
Self Identification
The Self-Identification Dataset, curated by Qyrou, is a specialized training resource designed to help developers and trainers establish clear self-identity awareness within language models. By incorporating this dataset, models can accurately learn and convey essential metadata about themselves, including their Model ID, Model Name, Model Description, Model Creator, Model Family, Model Architecture, Parameter Count… See the full description on the dataset page: https://huggingface.co/datasets/mithulaartigala/LLM-self-identification.llm-sast-v1
LLM-SAST v1
A high-quality, audited training dataset for fine-tuning small-to-mid-size language models to perform static application security testing (SAST) on infrastructure-as-code and application code — replacing rule-based scanners (Checkov, Trivy, Semgrep, KICS, Bearer, …) rather than auditing their output.
Task: given a single source file, the model emits a structured list of security findings (line ranges, category, severity, reasoning, remediation). No SAST-tool input. No… See the full description on the dataset page: https://huggingface.co/datasets/aioutfitters/llm-sast-v1.Nemotron-Safety-Guard-Dataset-v3-Ukr
Dataset Description
This is the localized Ukrainian version of the Nemotron-Safety-Guard-Dataset-v3. This specific repository contains exclusively the English subset of the original dataset, which has been fully translated into Ukrainian using the Lapa (Gemma 3) series of multimodal instructive models.
The original dataset was curated using the CultureGuard pipeline, which culturally adapts and translates content from the English Aegis 2.0 safety dataset. This Ukrainian variant… See the full description on the dataset page: https://huggingface.co/datasets/Stereotypes-in-LLMs/Nemotron-Safety-Guard-Dataset-v3-Ukr.LLMScheduling
ShareGPT Workload Cases
This dataset packages workload-generation artifacts derived from Aeala/ShareGPT_Vicuna_unfiltered.
It includes the manifest describing how cases were generated and per-category case files intended for scheduler and systems benchmarking.
Contents
manifest.json: dataset generation metadata and category definitions.
cases/: 176 generated case files across 22 workload categories.
Generation Metadata
Split: train
Cases per category: 8… See the full description on the dataset page: https://huggingface.co/datasets/ashman1705/LLMScheduling.DataShare_NeurIPS2026_LLMsRiskAttitude
Risk Attitude in LLMs: A Cross-Task Behavioral Dataset
Dataset accompanying the NeurIPS 2026 paper:
"[ANONYMIZED FOR REVIEW]"
Authors: [ANONYMIZED FOR REVIEW]
This dataset contains LLM response data and human baseline data collected across
three structured decision tasks designed to measure risk attitude and
behavioral reliability of contextual belief and risk decisions.
Tasks
Abbreviation
Full Name
Description
CTD
Clinical Triage Decision
Model acts… See the full description on the dataset page: https://huggingface.co/datasets/LLMsRiskAttitudeDataShare/DataShare_NeurIPS2026_LLMsRiskAttitude.LLM-Systems-Thinking-Curriculum
🧠 LLM Systems Demystification Roadmap
From First Principles to "The Wall"
"Recognition is slowly dawning that scaling alone is not the answer."
📖 Overview
This repository hosts the 13-Phase Technical Roadmap designed to bridge the gap between "Prompt Engineering" and Systems Thinking.
While the industry focuses on capabilities, this roadmap focuses on constraints. It is a rigorous, evidence-based curriculum that deconstructs Large Language… See the full description on the dataset page: https://huggingface.co/datasets/jkdkr2439/LLM-Systems-Thinking-Curriculum.fragility-moral-judgment-llms
Fragility of Moral Judgment in Large Language Models
Companion dataset for the FAccT paper Fragility of Moral Judgment in Large Language Models by Tom van Nuenen. Contains the moral dilemmas, community labels, and per-model verdicts (with explanations and reasoning traces) used in the study.
The paper investigates how stable LLM moral judgments are under minimal, morally-irrelevant perturbations of the same dilemma, and whether protocols and reasoning chains improve or worsen… See the full description on the dataset page: https://huggingface.co/datasets/ucberkeley-dlab/fragility-moral-judgment-llms.LLM-Susceptibility-theory
LLM Information Susceptibility -- Experimental Data
Experimental data for the paper "A Theory of LLM Information Susceptibility" by Zhuo-Yang Song.
Dataset Structure
.
├── results/ # Domain experiment results (25 files)
│ ├── dfs_baseline.json # Tetris DFS baseline (6 beam widths × 3 rewards × 40 seeds)
│ ├── llm_qwen-{7b,14b,32b,72b,3-max}_aggressive.json # Tetris LLM (5 models)
│ ├── llm_qwen-32b_conservative.json #… See the full description on the dataset page: https://huggingface.co/datasets/Nondegeneracy/LLM-Susceptibility-theory.OpenSFT-ita
OpenSFT ITA
LLMStigma
LLM Stigma Data
This dataset contains the data for the paper "Large Language Models Generate Stigmatizing Language During Reasoning Over Real-World Clinical Data". It contains the system prompt, user input, true labels, and LLM outputs of the main and the destigmatizing experiment. The data is packed on the model level. Each tar.gz file contains 13 JSON files.
Note: Only the tasks that are allowed to be open-sourced (N = 13) are included. For the rest of tasks (N = 22), please… See the full description on the dataset page: https://huggingface.co/datasets/YLab-Open/LLMStigma.llm-smartrouter-benchmark
LLM SmartRouter & Agent Highway Latency & Cost Benchmark (v1.4.0)
Empirical performance benchmark dataset comparing direct model endpoints (OpenAI, Anthropic Claude, Google Gemini) against the PixelRouter / BLUN SmartRouter proxy layer and Autonomous Agent Web Highway (https://api.pixeloffice.eu/v1).
v1.4.0 Benchmark Highlights
Anthropic Claude Messages API: Sub-35ms proxy routing for native /v1/messages payloads with 94%+ cost savings.
Machine Web Highway… See the full description on the dataset page: https://huggingface.co/datasets/pixeloffice/llm-smartrouter-benchmark.Cleaned_KJV_Bible_for_LLMsllm-smallproject
OpenMathInstruct-1
OpenMathInstruct-1 is a math instruction tuning dataset with 1.8M problem-solution pairs
generated using permissively licensed Mixtral-8x7B model.
The problems are from GSM8K
and MATH training subsets and the solutions
are synthetically generated by allowing Mixtral model to use a mix of text reasoning and
code blocks executed by Python interpreter.
The dataset is split into train and validation subsets that we used in the ablations experiments.
These two subsets… See the full description on the dataset page: https://huggingface.co/datasets/fadilahtulUkhti/llm-smallproject.online_remarks
Adolescent Mental Illness Prediction Benchmark
Introduction
The global surge in adolescent mental health challenges, combined with limited accessibility to offline psychiatric services—particularly in resource-constrained regions such as China—has created an urgent demand for scalable and preliminary mental health screening tools.
Adolescents often exhibit a greater propensity for self-disclosure in online and social media environments compared to traditional clinical… See the full description on the dataset page: https://huggingface.co/datasets/Evaluation-LLMs-Mental-Illness-Prediction/online_remarks.UN_NU_interpretation_LLMs
Quantifier Scope Interpretation Dataset
Datasets for an ongoing project about Scope preferences and ambiguity in LLM interpretation.
Dataset Structure
Splits
The dataset consists of synthetically generated stimuli pairing target sentences with interpretation-biased contexts (SSR vs. ISR).
Features
language (string)Language of the stimulus (English or Chinese).
structure (string)Surface syntactic configuration of the sentence:UN (universal >… See the full description on the dataset page: https://huggingface.co/datasets/CALM-Lab-Purdue/UN_NU_interpretation_LLMs.