datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
EpiBench-NeurIPS2026
EpiBench
Anonymous release for NeurIPS 2026 Evaluations & Datasets Track review (paper ID 1899). All methodology, ablations, and analyses are in the companion paper; this card lists only what reviewers and downstream users need to load the data.
A 25,737-patient ILAE-aligned multimodal epilepsy benchmark derived from PubMed Central case reports + 192 EpiRAG textbook vignettes.
6 tasks: epilepsy_type, seizure_type, ez_localization, aed_response, surgery_outcome, status_epilepticus… See the full description on the dataset page: https://huggingface.co/datasets/NeurIPS-1899-ED-2026/EpiBench-NeurIPS2026.neurips-2026-evals
NeurIPS 2026 Agent Evaluation Dataset
This dataset contains evaluation results for various AI agents across multiple benchmarks.
Dataset Structure
The dataset is organized by model (as configs) with each benchmark as a split.
Each model/benchmark folder contains:
Main results file (.jsonl or .parquet format)
Summary statistics (.summary.json) - for models with metadata
Configuration file (.toml) - for models with metadata
Traces folder with execution traces… See the full description on the dataset page: https://huggingface.co/datasets/akenginorhun/neurips-2026-evals.IGF-Bench
IGF-Bench: Indoor Geometric Fidelity Benchmark
Anonymous mirror for NeurIPS 2026 Evaluations and Datasets Track double-blind review.
The de-anonymised author/maintainer information will replace this header at camera-ready.
IGF-Bench is the first benchmark for evaluating structural-level geometric fidelity of conditionally generated indoor scene images, going beyond perceptual metrics like FID and LPIPS. It pairs 3,600 calibrated synthetic ground-truth views with 21,600 generated… See the full description on the dataset page: https://huggingface.co/datasets/igfbench-neurips2026/IGF-Bench.causalverify-neurips2026
🎯 CausalVerify
An Execution-Grounded Benchmark for LLM Causal Inference Workflows
NeurIPS 2026 — Evaluations and Datasets Track · double-blind review · frozen at tag neurips2026-submission
💡 TL;DR
A benchmark of 259 published economics papers (Experiment A — real-paper text-agreement diagnostic) and 100 fixed-seed synthetic data-generating processes (Experiment B — execution-grounded coefficient recovery), evaluating 7 frontier LLMs. The central… See the full description on the dataset page: https://huggingface.co/datasets/causalverify/causalverify-neurips2026.crychic-dafny-acsl
CRYCHIC Dafny-to-ACSL-C Verified Translation Benchmark
This anonymized review artifact accompanies the NeurIPS 2026 Evaluations and Datasets submission:
CRYCHIC: A Universal Framework for Cross-Language Verified Code Translation.
CRYCHIC translates verified Dafny programs into C programs annotated with ACSL specifications, then checks the generated artifacts with Frama-C WP. This release contains the 1,679 fully verified Dafny/C+ACSL pairs used as the positive benchmark corpus.
The… See the full description on the dataset page: https://huggingface.co/datasets/neurips2026-crychic/crychic-dafny-acsl.Regressionsob
The Structured Output Benchmark (SOB)
A multi-source benchmark for evaluating structured-output quality in LLMs.
(Anonymous submission — links to code, paper, and leaderboard withheld during double-blind review.)
Dataset summary
SOB evaluates how accurately LLMs produce schema-compliant and value-correct JSON from unstructured or semi-structured context — across three source modalities:
Config
Source
Context delivered as
Records
default
HotpotQA (multi-hop QA)… See the full description on the dataset page: https://huggingface.co/datasets/neurips2026sob/sob.LR-Transfer-Trajectory
Dataset Documentation
Overview
This dataset captures per-step training and validation metrics from training runs of a 12-layer GPT-style decoder-only transformer. Each run is stored as a single .csv file in which every row corresponds to one logged step, and several columns hold per-parameter measurements encoded as JSON.
The dataset is designed to support post-hoc analysis of:
Loss curves (train / val)
Throughput and step latency
Per-layer / per-parameter dynamics:… See the full description on the dataset page: https://huggingface.co/datasets/neurips2026lrtransfer/LR-Transfer-Trajectory.formal-anytime-valid-stats
Formal-AVS: A Lean Benchmark for Anytime-Valid Confidence-Sequence Theorem Proving
60 Lean 4 theorem targets on anytime-valid confidence sequences across four families (Howard-Ramdas, betting, Whitehouse vector, asymptotic CLT).
Benchmark Structure
60 targets grouped into tiers T0-T3 (pre-evaluation) and categories T4-T5 (empirical)
7 drafters evaluated across single-shot, agentic, and unbounded modes
14 Aristotle sessions (unbounded refinement)
Headline Results… See the full description on the dataset page: https://huggingface.co/datasets/neurips-2026-avs-bench/formal-anytime-valid-stats.neurips-2026-cpx
neurips-2026-cpx — Korean OSCE history-taking dialogues with a GPT-4o virtual standardized patient
49 text-based history-taking dialogue sessions between 17 senior Korean
medical-student participants (Years 3–4 of a 6-year curriculum) and a
GPT-4o-driven virtual standardized patient (VSP). Released as the
empirical evaluation dataset accompanying our NeurIPS 2026 submission.
Sessions: 49
Participants: 17 (anonymised to R001–R017)
Total QA turns: 1,763
Language: Korean
Domain:… See the full description on the dataset page: https://huggingface.co/datasets/neurips-2026-cpx/neurips-2026-cpx.pi_bench
pi-bench
pi-bench is a multi-task audio benchmark prepared for public hosting and evaluation reproducibility. The repository is organized as a Hugging Face dataset with one dataset config per task file under data/, so each benchmark subset is visible and loadable independently.
Overview
The current release contains 11 task-specific configs spanning three broad categories:
Counterfactual/contextual QA (CTC_*)
Clarification-seeking QA (Trivia_Clarification_*… See the full description on the dataset page: https://huggingface.co/datasets/neurips2026-pi-bench/pi_bench.retrieval-conditional-neurips2026
Dataset Release — Retrieval-Conditional NeurIPS 2026
This bundle accompanies the NeurIPS 2026 D&B Track submission
"To Retrieve or Not to Retrieve? Most of the Benefit is Structural, Not Semantic."
Contents
File
Config name
Description
data/per_task_outcomes.csv
per_task_outcomes (default)
Per-(backbone × env × condition × task) success/failure labels. 3,064 rows.
data/stats_per_cell.csv
stats_per_cell
54-cell aggregate success rates and pairwise contrasts.… See the full description on the dataset page: https://huggingface.co/datasets/anon123312/retrieval-conditional-neurips2026.pca_50ksignalbench-terminalbench-bkp
SignalBench TerminalBench Dense Signal Dataset
TerminalBench evaluation points and Monte Carlo Q-value labels for studying dense signal functions in text-only terminal tasks.
This repository contains one SignalBench dataset with two synchronized views:
runtime/dataset.pkl is the executable artifact used by the SignalBench benchmark code.
data/examples.parquet has exactly one row per benchmark example, with
state/action/next-state text, renderable state_image and
next_state_image… See the full description on the dataset page: https://huggingface.co/datasets/neurips2026-anonymous/signalbench-terminalbench-bkp.signalbench-terminalbench
SignalBench TerminalBench Dense Signal Dataset
TerminalBench evaluation points and Monte Carlo Q-value labels for studying dense signal functions in text-only terminal tasks.
This repository contains one SignalBench dataset with two synchronized views:
runtime/dataset.pkl is the executable artifact used by the SignalBench benchmark code.
data/examples.parquet has exactly one row per benchmark example, with
state/action/next-state text, renderable state_image and
next_state_image… See the full description on the dataset page: https://huggingface.co/datasets/neurips2026-anonymous/signalbench-terminalbench.neurips2026-epistemic-honesty
Hard Layer V3: Epistemic Honesty Benchmark for Medical LLMs
Dataset Description
Hard Layer V3 is a 100-question benchmark designed to measure epistemic honesty in medical large language models — whether models explicitly acknowledge uncertainty when confronted with fabricated medical entities, ambiguous thresholds, and knowledge boundaries.
Unlike traditional medical QA benchmarks that focus on accuracy, this benchmark evaluates whether models can appropriately respond… See the full description on the dataset page: https://huggingface.co/datasets/vmti/neurips2026-epistemic-honesty.bioresponseneurips-2026-pkb-datasetpraxis