datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
vibench
VIBench
VIBench is a benchmark for measuring vertical integration bias in
direct and agentic code generation. It contains 20 direct scenarios and
20 aligned agentic workflows covering realistic software integration
choices across cloud and API ecosystems. The bundle includes the
benchmark tasks, model metadata, system prompts, provider-proof
artifacts, and the blind detector-audit sample used for validation.
Viewer splits
The dataset viewer is intentionally simplified to… See the full description on the dataset page: https://huggingface.co/datasets/vibench-emnlp26/vibench.vibench-results
VIBench Results
VIBench Results contains the retained raw generations,
detector-labeled outputs, complete runs, option-order and runtime
ablations, paper-facing summaries, figures, configs, audit files, and
static explorer indices used in the study.
The main paper evaluation covers 13 models, 15,600 direct generations,
and 2,000 agentic runs. Direct and agentic VIB are reported as
scenario-matched, share-normalized differences in affiliated-ecosystem
selection relative to strict… See the full description on the dataset page: https://huggingface.co/datasets/vibench-emnlp26/vibench-results.EMNLP_Cost-Aware-Protocol-Routing
Cost-Aware Protocol Routing: Matched Protocol Outcomes
The short version. We ran the same 6,803 reasoning problems through four
different LLM collaboration setups — from a single direct answer up to a
four-agent deliberation — and recorded, for every problem, which ones got it
right. Then we asked whether a model can look at a problem beforehand and
predict which setup is worth paying for.
It can predict whether it will fail. It cannot predict which collaboration
protocol will… See the full description on the dataset page: https://huggingface.co/datasets/AgentsSci/EMNLP_Cost-Aware-Protocol-Routing.emnlp2026_dataset
When Synthetic Data Hurts: On Catastrophic Forgetting in Skill Retrieval for LLM Agents
This Readme file serves as a guide for using the two datasets we presented in our paper: Harbor Trial dataset which consists of real task execution data and Track B dataset which contains synthetic data. Those datasets are intended to finetune your own skill retrieval solutions and recipes and serve as a common ground to compare new catastrophic forgetting mitigation methodologies with our… See the full description on the dataset page: https://huggingface.co/datasets/manulife/emnlp2026_dataset.emnlp-2020-2025-atomic-claims
EMNLP 2020–2025 Atomic Contribution Claims (ACC), with drift clusters
18,293 atomic contribution claims extracted from the abstracts of the full EMNLP main track 2020–2025, plus the canonical 80-cluster drift clustering and per-cluster drift statistics used in the Drift Inspector paper.
An atomic contribution claim (ACC) is a single self-contained sentence stating one concrete contribution of a paper: atomic (one contribution-bearing proposition), decontextualized (pronouns… See the full description on the dataset page: https://huggingface.co/datasets/Hamyrappy/emnlp-2020-2025-atomic-claims.CASTELLA
CASTELLA
This repository provides wav files used in CASTELLA: Long Audio Dataset with Captions and Temporal Boundaries
. This dataset is originally provided in GitHub.
Each sample includes long audio containing some audio events with the temporal and textual annotation.
Project page: https://h-munakata.github.io/CASTELLA-demo/
Code: https://github.com/line/lighthouse
This repository only contains annotation data, not audio data.
Extracted features are available on HF.
If you need… See the full description on the dataset page: https://huggingface.co/datasets/lighthouse-emnlp2024/CASTELLA.paper-central-data-emnlp2024-sorted-by-github-stars-no-artifactsEMNLP2023-paperspaper-central-data-emnlp2024-sorted-by-github-starsemnlp_papers_samplemPIQA-MRL-2025-EMNLP0907_preference_emnlp2023