datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
explicit-edit-benchmark
Explicit Edit Benchmark
226 deterministic exact-edit tasks, run by different agents, harnesses, models and configurations. Every observation records what the harness did and whether the resulting files matched byte for byte.
Source code and benchmark runner: GitHub — Explicit Edit Benchmark
Open the interactive Explorer to compare agents, harnesses, models, versions, reasoning modes, correctness, recovery, time, cost and tokens.
Leaderboard by model route
Score v2… See the full description on the dataset page: https://huggingface.co/datasets/alexshpunt/explicit-edit-benchmark.regx-benchmark
RegX
Cross-Domain Multi-View Point Cloud Registration Benchmark
RegX evaluates multi-view point cloud registration across scales spanning nine orders
of magnitude — nanometre-scale microscopy to kilometre-scale airborne maps — and sensors
never designed to be compared: clinical colonoscopes, RGB-D cameras, spinning and
solid-state LiDAR, terrestrial and airborne laser scanners.
Most registration benchmarks fix one sensor and one scale. RegX asks a narrower question
instead: does… See the full description on the dataset page: https://huggingface.co/datasets/YuePanEdward/regx-benchmark.funes-handoff-recall-benchmark
handover-vs-recall
A long investigation bloats an agent session until each new turn costs more to carry the context than to
do the work. Switching to a fresh session avoids that — but the findings have to travel somehow, and the
ways of moving them differ in cost. This benchmark measures those ways, as cost per successful task,
on tasks that genuinely require the prior investigation:
arm
channel
A branch-only
switch, carry nothing — the fresh session re-derives the… See the full description on the dataset page: https://huggingface.co/datasets/dacorvo/funes-handoff-recall-benchmark.data-product-benchmark
DPDisc Dataset
Paper | Code
Dataset Description
This dataset provides a benchmark for automatic data product creation. The task is framed as follows: given a natural language data product request and a corpus of text and tables, the objective is to identify the relevant tables and text documents that should be included in the resulting data product which would useful to the given data product request. The benchmark brings together three variants: HybridQA, TAT-QA, and… See the full description on the dataset page: https://huggingface.co/datasets/ibm-research/data-product-benchmark.scene-mem-benchmark
scene-mem-benchmark
A benchmark for scene memory in embodied agents: an agent watches a mobile manipulator work
in a house for several minutes, then is asked to retrieve an object it has to remember — one
that was moved, dropped, or merely seen along the way — or (resume) to go back and finish the
job it was interrupted in, remembering how far it had got — or (routine) to put a new object away
where this household keeps that kind of thing, a rule it was never told and can only… See the full description on the dataset page: https://huggingface.co/datasets/Keh0t0/scene-mem-benchmark.hack-ignition-benchmark
hack-ignition benchmark — data, v0.1.6
Training trajectories of reinforcement-learning runs on exploitable graders, for studying and predicting when RL
comes to produce exploits. Each family is a set of GRPO runs over configurations of (start model, prompt,
training set, grader / reward structure, recipe), with one or more seeds per configuration. Every family stores
what its training logs contain — per-step exploit, task and reward rates, the item × step exploit record… See the full description on the dataset page: https://huggingface.co/datasets/EleutherAI/hack-ignition-benchmark.Auto-Fill-Benchmark
Auto-Fill Benchmark
Benchmark for predicting missing cell values in real-world tables, introduced in
Auto-Fill: Learning to Predict Missing Values Accurately with Specialist Language Models
(PVLDB 19(11), 2026 — arXiv:2607.19847).
Each case is a real table in which exactly one cell is replaced by [MISSING], together with the ground-truth value.
Code: https://github.com/lyrain2001/auto-fill
Models: Auto-Fill-Qwen3-8B-Knowledge ·
Auto-Fill-Qwen3-8B-Reasoning ·… See the full description on the dataset page: https://huggingface.co/datasets/lyrain2001/Auto-Fill-Benchmark.legal-llm-benchmark
Legal LLM Benchmark Dataset
Safety-Utility Trade-offs in Legal AI: An LLM Evaluation Across 12 Models
Quick Start
from datasets import load_dataset
# Core datasets
questions = load_dataset("marvintong/legal-llm-benchmark", "questions")
phase1_evals = load_dataset("marvintong/legal-llm-benchmark", "phase1_evaluations")
phase3_evals = load_dataset("marvintong/legal-llm-benchmark", "phase3_evaluations")
# Additional helpful datasets
contracts =… See the full description on the dataset page: https://huggingface.co/datasets/marvintong/legal-llm-benchmark.TrialPanorama-benchmarkDataset website: https://ryanwangzf.github.io/projects/trialpanorama
Reasoning_Benchmark_800
Reasoning Benchmark 800
Reasoning_Benchmark_800 is an 800-episode LeRobot v3 training dataset for
studying semantic and compositional generalization in vision-language-action
models. It combines four completed 200-episode datasets while preserving the
original observations, actions, timing, and videos. Only the training-language
metadata was reassigned.
The experimental control is that held semantic concepts are absent from the
training prompts even though the corresponding… See the full description on the dataset page: https://huggingface.co/datasets/justintiensmith/Reasoning_Benchmark_800.SciCode-Runnable-Benchmark-Reviewedmultilingual_tokenizer_benchmark
Multilingual Tokenizer Benchmark
More details of each subset like word count, character count, original sources, etc, can be found in the dataset_meta.yaml file in the repository root.
Natural language word count functions
Download spacy models
pip install ntlk spacy pygments underthesea camel-tools
python -m spacy download ko_core_news_sm
python -m spacy download ja_core_news_sm
python -m spacy download zh_core_web_sm
import nltk
nltk.download('punkt_tab')… See the full description on the dataset page: https://huggingface.co/datasets/eduagarcia/multilingual_tokenizer_benchmark.rag_instruct_benchmark_tester
Dataset Card for RAG-Instruct-Benchmark-Tester
Dataset Summary
This is an updated benchmarking test dataset for "retrieval augmented generation" (RAG) use cases in the enterprise, especially for financial services, and legal. This test dataset includes 200 questions with context passages pulled from common 'retrieval scenarios', e.g., financial news, earnings releases,
contracts, invoices, technical articles, general news and short texts.
The questions are segmented… See the full description on the dataset page: https://huggingface.co/datasets/llmware/rag_instruct_benchmark_tester.Finers-4k-benchmarkbenchmark-radar
Benchmark Radar Dataset
Overview
Benchmark Radar is a living registry, search engine, and discovery pipeline
for AI evaluation benchmarks. This dataset mirrors the full-corpus findings of the
Benchmark Radar technical report ("Benchmark Radar: Daily Discovery and
Full-Corpus Search Across the AI Evaluation Landscape", arXiv:2609.11115).
As described in the paper's Two Input Paths framework, Benchmark Radar
combines two complementary systems:… See the full description on the dataset page: https://huggingface.co/datasets/ktwu01/benchmark-radar.ai-humanizer-benchmark
AI Humanizer Benchmark — monthly cycle data
The complete raw data of AI Humanizer Benchmark, a monthly measured benchmark of AI humanizers. Every tool rewrites the same 33 freshly generated texts on its default settings; every output is scored by 7 commercial AI detectors (GPTZero, Originality.ai, Copyleaks, Winston AI, ZeroGPT, QuillBot, Grammarly) plus meaning preservation and readability.
This dataset is the official mirror of the GitHub data repository, published by the AI… See the full description on the dataset page: https://huggingface.co/datasets/ai-humanizer-benchmark/ai-humanizer-benchmark.compile-benchmark
CompilingThings Compile Benchmark for MQL5®
This release evaluates compile success of generated MQL5 on two private held-out sets. The first is 300 Expert Advisor prompts, run on four arms: the base model, two tuned local models and one frontier API model. The second is 200 non-EA prompts (include files, custom indicators, scripts and services, 50 each), run on the three local arms, plus a stability re-run of one of them. The holdout results are attested, not fully verifiable:… See the full description on the dataset page: https://huggingface.co/datasets/CompilingThings/compile-benchmark.continual-learning-bench-data
Continual Learning Benchmark — Data
Frozen corpora and supporting artifacts for the six tasks in the Continual Learning Benchmark. The repo accompanies the (anonymized) benchmark codebase, which loads these files and feeds them — with task-specific framing — to the system under evaluation.
Repository layout
blind_spectrum_monitoring/ # frozen scan corpus + metadata
codebase_adaptation/ # final PR dataset + 2 docker images
cohort_studies/ # cohort defs… See the full description on the dataset page: https://huggingface.co/datasets/continual-learning-benchmark/continual-learning-bench-data.MEME
MEME: Multi-Entity and Evolving Memory Evaluation
A benchmark for evaluating LLM memory systems along two orthogonal dimensions: entity scope (single vs. multi-entity) and temporal dynamics (static vs. evolving). MEME defines six tasks targeting memory-intensive operations in each quadrant, including two task types that no prior benchmark covers: Cascade (propagating updates through dependency rules) and Absence (recognizing uncertainty when a previously valid answer becomes… See the full description on the dataset page: https://huggingface.co/datasets/meme-benchmark/MEME.turkish_political_position_benchmark
Turkish Political Position Benchmark
The Turkish Political Position Benchmark measures how language models respond to normative statements about Turkish politics. It reports ideological dimension scores and response similarity to documented political-party reference profiles.
The benchmark does not claim that a model belongs to a party, has a voting intention, or possesses political beliefs. A party similarity score only means that the model produced a similar pattern of answers… See the full description on the dataset page: https://huggingface.co/datasets/logicBombExe/turkish_political_position_benchmark.Benchmark-Testingforge-benchmark
FORGE: Fake Online Recommendations in Generative Environments
FORGE is a benchmark for measuring whether search-augmented large language
models recommend synthetic fake brands when their retrieval evidence is
poisoned. It contains 225 Chinese product queries across 15 categories,
evaluation results for 12 production LLMs, and rebuildable evidence-bundle
indexes.
This dataset accompanies the paper One Polluted Page Is Enough: Evaluating
Web Content Pollution in LLM Recommenders… See the full description on the dataset page: https://huggingface.co/datasets/leoluo25933/forge-benchmark.streaming-phi-deidentification-benchmark
Streaming PHI De-Identification Benchmark
Most PHI de-identification benchmarks evaluate a single document in isolation. That is not how clinical data actually moves. A patient's name appears in a clinical note, then in an ASR transcript ten minutes later, then in imaging metadata an hour after that. Each event looks low-risk on its own. The cumulative exposure across modalities is what creates re-identification risk.
This dataset captures that. Every record is fully synthetic. It… See the full description on the dataset page: https://huggingface.co/datasets/vkatg/streaming-phi-deidentification-benchmark.knows-benchmark
KNOWS Benchmark
KNOWS evaluates web agents on the work people actually do in Google Workspace: writing documents,
building spreadsheets, and composing slide decks that require web research, multi-step tool use, and
faithful grounding in retrieved sources.
This dataset contains the task definitions — the prompt an agent receives, plus the structured
evaluation rubric used to grade the artifact it produces.
Tasks
110 (22 templates × 5 instances)
Domains
20
Mean… See the full description on the dataset page: https://huggingface.co/datasets/utahnlp/knows-benchmark.peptide-reasoning-benchmark
Peptide Reasoning Benchmark
PEB v1.0-RC benchmark release for peptide-reasoning model evaluation.
Includes cases, splits, baselines, references, and leaderboard artifacts.
GitHub: https://github.com/ray-r-ren/peptide-reasoning-bench
Trained a small reference LoRA model: https://huggingface.co/rayrren/the-spice-v0-mvp
powergrid-benchmark2
Note: This is a dummy dataset for browsing only. The official dataset can be downloaded from our pipeline (Data Hub).
Full dataset for Benchmark 2
qwen3.8-27b-inference-benchmark-4090
Qwen3.8-27B Inference Benchmark on RTX 4090 48GB
中文说明 · GitHub benchmark repository
Structured performance and accuracy results for four real Qwen3.8-27B serving configurations on an NVIDIA RTX 4090 48 GB workstation. A dual-GPU llama.cpp BF16 reference additionally used an RTX 3090 24 GB.
This dataset is the analysis-friendly companion to the full benchmark repository. It publishes aggregate tables, 140 normalized per-request performance records, accuracy scores, sanitized… See the full description on the dataset page: https://huggingface.co/datasets/pxzleo/qwen3.8-27b-inference-benchmark-4090.kdv-rag-benchmark
KDV RAG Benchmark
A retrieval benchmark dataset for Turkish VAT (KDV, Katma Değer Vergisi) law — built by adding retrieval layers one at a time (chunking, model choice, hybrid search, reranking, query rewriting, historical/date filtering) and statistically validating each one individually (see Results).
Dataset structure
Splits
Split
Records
Period
train
728
2018-2023
test
154
2024-2026
Split strategy: temporal — train and test… See the full description on the dataset page: https://huggingface.co/datasets/dokukoza/kdv-rag-benchmark.vqa-cmsv-benchmark
VQA-CMSV Benchmark Data Package
This repository contains annotation splits for VQA v2-CMSV, GQA-CMSV, and VG-CMSV, plus patch-mask NPZ files used for mask supervision experiments.
Contents
data/vqa_v2_cmsv/train.json, data/vqa_v2_cmsv/val.json, data/vqa_v2_cmsv/test.json
data/gqa_cmsv/train.jsonl, data/gqa_cmsv/val.jsonl, data/gqa_cmsv/test.jsonl
data/vg_cmsv/train.jsonl, data/vg_cmsv/val.jsonl, data/vg_cmsv/test.jsonl
masks/vqa_v2_cmsv_masks.npz
masks/gqa_cmsv_masks.npz… See the full description on the dataset page: https://huggingface.co/datasets/as-benchmark-artifacts/vqa-cmsv-benchmark.mcphunt-agent-traces
MCPHunt Agent Traces
Agent execution traces from the MCPHunt evaluation framework, measuring
cross-boundary data propagation in multi-server MCP agents.
Contents
main/ — 3,615 traces from 5 models across 147 tasks and 7 environment
variants (risky_v1/v2/v3, benign, hard_neg_v1/v2/v3). One JSON file per model.
mitigation/ — 2,706 traces from the prompt-mitigation study (M0--M3
levels) across 3 models.
live_guard_defense/ — 387 DeepSeek-V4-Flash traces from the… See the full description on the dataset page: https://huggingface.co/datasets/mcphunt-benchmark/mcphunt-agent-traces.