datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
TreeUQ
TreeUQ — Geographically-Tiled Bavaria EO Benchmark
TreeUQ is a large-scale Earth observation benchmark for tree species mapping and tree structure estimation (height, count, density, variance) over Bavaria, Germany.
Each 128×128 pixel patch (10 m resolution, EPSG:25832) contains:
Sentinel-2 — 4 seasonal composites (spring/summer/autumn/winter 2025), 10 bands
Sentinel-1 GRD — 4 seasonal composites, VV + VH polarisation, linear gamma-0
Tree species raster — Bavarian species… See the full description on the dataset page: https://huggingface.co/datasets/iclr2027kiwi/TreeUQ.iclr2027-videomme-long-checkpoints-20260914iclr-2027-proceedings-whitelist
ICLR 2027 proceedings whitelist
Generated public metadata used by the ICLR 2027 reciprocal-reviewing checker.
iclr2027-vnbench-features-20260913iclr2027_univtac_mlx_ckptICLR2027-duplexsafety-exp
ICLR2027-duplexsafety-exp
Raw outputs of full-duplex spoken-dialogue safety experiments: simulated conversations (audio and
transcripts), judge verdicts and system traces. Research use only; access on request.
Layout: studies/<study>/{REPORT.md, manifest.json, tables/, shared/, runs/, audio/}.
Studies
study
what
episodes
site
v01full
v0.1 seat, P2 (cancel) vs P2k (wait), the full 341-goal seed set, arms direct / c4 / c8, benchmark v2.1
2046
JHU ARCH… See the full description on the dataset page: https://huggingface.co/datasets/MagicLuke/ICLR2027-duplexsafety-exp.ROOTS
ROOTS
ROOTS contains 43,922,159 audio-language conversations across four taxonomy tiers. Audio is supplied by the source datasets below.
Quick start
from datasets import load_dataset
dataset = load_dataset("iclr2027anon/ROOTS", split="train", streaming=True)
row = next(iter(dataset))
Use the audio guide to locate and load each conversation's clips in order.
Columns
Columns
Meaning
id
Conversation ID: roots_ followed by 32 hexadecimal… See the full description on the dataset page: https://huggingface.co/datasets/iclr2027anon/ROOTS.iclr2027-videomme-features-20260913iclr2027_freetacman_pretrain_robust_l2_totalbs512synthetic-strong-detection
Synthetic Strong Detection
This dataset contains 1,000,000 ten-second sound scenes with event annotations. WAV audio is included in data-20k/train-*.tar.
Quick start
from datasets import load_dataset
metadata = load_dataset("iclr2027anon/synthetic-strong-detection", split="train")
Download and extract the TAR shards from data-20k/. Load a row's WAV from audio/<audio_name>.
Metadata
Column
Meaning
id
Scene ID: sed_ followed by 32… See the full description on the dataset page: https://huggingface.co/datasets/iclr2027anon/synthetic-strong-detection.iclr2027-videomme-frames-20260913synthetic-detect-diarize
Synthetic Detection and Diarization
This dataset contains 1,000,000 ten-second sound scenes with event and speaker annotations. WAV audio is included in data-20k/train-*.tar.
Quick start
from datasets import load_dataset
metadata = load_dataset("iclr2027anon/synthetic-detect-diarize", split="train")
Download and extract the TAR shards from data-20k/. Load a row's WAV from audio/<audio_name>.
Metadata
Column
Meaning
id
Scene ID:… See the full description on the dataset page: https://huggingface.co/datasets/iclr2027anon/synthetic-detect-diarize.surface-audit
TruthfulQA-476 — a surface-form-cleaned binary-choice TruthfulQA
TruthfulQA-476 is the recommended drop-in replacement for the binary-choice TruthfulQA
evaluation set. It keeps 476 of the 790 original question pairs, in the original schema, chosen so
that a classifier restricted to six surface features of the answer text (negation, hedging, length,
token statistics) can no longer separate correct from incorrect answers above chance, while the
ranking of models on the subset… See the full description on the dataset page: https://huggingface.co/datasets/iclr2027-surface-audit/surface-audit.iclr2027-vnbench-frames-20260913animalspeak-pseudovox
AnimalSpeak Pseudovox
This dataset contains 346,907 short, silence-trimmed animal vocalization clips from the train-unseen split. WAV audio is included in data-20k/train-*.tar.
Quick start
from datasets import load_dataset
metadata = load_dataset("iclr2027anon/animalspeak-pseudovox", split="train")
Download and extract the TAR shards from data-20k/. Load a row's WAV from audio/<audio_name>.
Metadata
Column
Meaning
id
Clip ID: pseudovox_… See the full description on the dataset page: https://huggingface.co/datasets/iclr2027anon/animalspeak-pseudovox.medgym-ICLR2027iclr2027_univtac_l2robust_ftmpre_l2robust_ckptLUSim-V5-LLM-cache-Berlin-ICLR2027
LUSim V5 LLM response caches (ICLR 2027 submission)
One tarball per run. llm_cache_v5.tar.gz is the K = 50 East-West Express run (extracts to data/berlin/llm_cache_v5/{baseline,shock}/); the other nine are the historical-validation, second-corridor, floor-space, sensitivity, and archetype-sweep runs (see data/berlin/llm_cache_v5/MANIFEST.md in the code repository for SHA-256 hashes and extraction paths). Generated with --no-llm replay in mind: a cache miss fails fast.
BEANS-Next
BEANS-Next
BEANS-Next evaluates audio-language models on 43 bioacoustic tasks, with 57,637 examples and all evaluation audio included.
Tier
Focus
Examples
1
Acoustic perception
9,029
2
Semantic recognition
24,972
3
Structural and relational reasoning
18,652
4
Multiple audio clips and in-context learning
4,984
Quick start
Load the test split, then download the audio for an example:
from datasets import load_dataset
from huggingface_hub import… See the full description on the dataset page: https://huggingface.co/datasets/iclr2027anon/BEANS-Next.iclr2027-b200-backup-public-20260913ICLR2027WorldModel