datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
latent-sr-embeddings
Latent-SR Embeddings: Precomputed VAE Latents for Medical Image Super-Resolution
Precomputed VAE latent embeddings from the paper:
"Domain-Specific Latent Representations Improve the Fidelity of Diffusion-Based Medical Image Super-Resolution"Sebastian Cajas, Ashaba Judith, Rahul Gorijavolu, Sahil Kapadia, Hillary Clinton Kasimbazi, Leo Kinyera, Emmanuel Paul Kwesiga, Sri Sri Jaithra Varma Manthena, Luis Filipe Nakayama, Ninsiima Doreen, Leo Anthony Celi.arXiv:2604.12152 (2026)… See the full description on the dataset page: https://huggingface.co/datasets/sebasmos/latent-sr-embeddings.grayscale_image_aesthetic_3M
Dataset Card for "grayscale_image_aesthetic_3M"
More Information needed
bootstrap-latent-thought-dataThis dataset is associated with the paper Reasoning to Learn from Latent Thoughts. It contains data used for pretraining language models with a focus on improving data efficiency by modeling and inferring latent thoughts underlying the text generation process, such as on reasoning-intensive math corpus. An expectation-maximization algorithm is developed for models to self-improve their self-generated thoughts and data efficiency.
animesfw
Dataset Card for "animesfw"
More Information needed
voxcpm2-ghana-speech-ipa-latents
VoxCPM2 Ghana — Precomputed AudioVAE Latents
Precomputed VoxCPM-2 AudioVAE latents for a Ghanaian multilingual TTS fine-tune,
ready for training with the official train_voxcpm_finetune.py
(train_manifest: ghana-latents). No audio decoding or VAE encoding needed at train
time — the latent feat is fed straight to the VoxCPM-2 model with the IPA transcript.
Each language is a dataset subset:
from datasets import load_dataset
ds =… See the full description on the dataset page: https://huggingface.co/datasets/ghanaopenai/voxcpm2-ghana-speech-ipa-latents.video-dit-latents-hq
Video DiT Latents - Animals (HQ)
Pre-computed VAE latents for training video generation models.
Dataset Info
Property
Value
Resolution
256×256 pixels
Latent Shape
(4, 16, 32, 32)
Frames
16 @ 8fps (2 seconds)
VAE
stabilityai/sd-vae-ft-mse
Classes
dog, cat, bird, horse, fish, lion, elephant, monkey, butterfly, deer
Usage
import torch
from pathlib import Path
# Load a single latent
latent = torch.load("dog/12345.pt")… See the full description on the dataset page: https://huggingface.co/datasets/Jnaranjo/video-dit-latents-hq.layergen-eval-latents
LayerGen — Eval-Set Latents (VAE latents + baked text embeddings)
Pre-encoded evaluation-set inputs for the LayerGen layer-decomposition / harmonization models,
so inference can run anywhere (off-AIP) without the raw video → VAE-encode → umT5-encode pipeline.
Each *.parquet is one clip and is fully self-contained:
column group
contents
{composite,mask,fg,bg}_latent_bytes (+ _shape, _dtype)
4-stream Wan-VAE latents, 81f/21 latent-T, fp16, [16,21,60,104]… See the full description on the dataset page: https://huggingface.co/datasets/cs-mshah/layergen-eval-latents.IndicVoice-latent-NEWLatent-Earth
Latent Earth: An Atlas of Architecture in Flux.2
200,000 images of 40,000 places on Earth, each rendered by a single
image model in a single state of its training, with five internal
representations recorded for every image while it was being generated.
Nothing else enters. Each prompt contains only a place's name; no
photographs, no maps, no climate records correct what the model proposes.
This is therefore not a depiction of the world but a probe of the model: a
survey of what… See the full description on the dataset page: https://huggingface.co/datasets/Punktiert/Latent-Earth.got-activations-llama3.1-405b-base
meta-llama/Llama-3.1-405B — Activation Dataset
Cached activations extracted from meta-llama/Llama-3.1-405B (revision unknown).
Contents
Tensor
Layers
Dim
Pooling
Shards
Row Bytes
hidden_layers
0-125
16384
-
12
-
Prompts: 7660
Format version: 1.1
Load with lmprobe
from lmprobe import pull_dataset, load_activation_dataset
# Option 1: Pull into local cache (enables probe training without re-extraction)… See the full description on the dataset page: https://huggingface.co/datasets/latent-lab/got-activations-llama3.1-405b-base.ego10k-vjepa-latents
Ego10k V-JEPA Latents Dataset
This dataset contains compressed, highly-informative Video Joint Embedding Predictive Architecture (V-JEPA) latents extracted from Ego-centric industrial manufacturing videos.
Dataset Structure
The dataset is partitioned into roughly 1GB .parquet chunks using PyArrow.
Data Source and Preprocessing
The latent embeddings in this dataset were systematically extracted from the Ego10k Master Dataset provided by build.ai. The… See the full description on the dataset page: https://huggingface.co/datasets/rookierufus/ego10k-vjepa-latents.SEMM-Latent-Telemetry
SEMM-Latent-Telemetry
Bare-metal hardware telemetry and SNN latent space routing data for neuromorphic quantization research. This dataset documents the discovery of Semantic Attractor Clustering — that a Spiking Neural Network physically routes different semantic concepts (abstract language vs code syntax vs math logic) into distinct, repeatable biological pathways when L2 Normalization is applied to LLM embeddings.
Hub ID: rmems/SEMM-Latent-TelemetryNames: SEMM = Spiking… See the full description on the dataset page: https://huggingface.co/datasets/rmems/SEMM-Latent-Telemetry.voxcpm-ghana-latents
VoxCPM Ghana — Precomputed AudioVAE Latents
The exact training-ready data used to fine-tune
ghananlpcommunity/voxcpm-ghana:
precomputed VoxCPM-0.5B AudioVAE latents (16 kHz) for 42 Ghanaian languages +
filtered Ghanaian English, with language-tagged transcripts. Drop-in for
VoxCPM fine-tuning — no audio decoding or VAE encoding needed at train time.
1,756,157 clips · ~3,400 h · 16 kHz
42 Ghanaian languages (incl. Twi split: twi-asante, twi-akuapem) + en
AudioVAE from… See the full description on the dataset page: https://huggingface.co/datasets/ghananlpcommunity/voxcpm-ghana-latents.midashenglm-gen-training-latents
ModelsLab/midashenglm-gen-training-latents
Precomputed audio latents for fine-tuning
mispeech/midashenglm-gen,
paired with six-view prompts in the exact format the model was trained on.
This is not an audio dataset and not a caption dataset. Each record is the
output of the model's frozen DashengTokenizer encoder — 768-dimensional latents
at 25 Hz, stored float16 — next to the tagged prompt string built from the
source metadata.
Why it exists
The encoder is frozen… See the full description on the dataset page: https://huggingface.co/datasets/ModelsLab/midashenglm-gen-training-latents.got-activations-qwen2.5-0.5b
Qwen/Qwen2.5-0.5B — Activation Dataset
Cached activations extracted from Qwen/Qwen2.5-0.5B (revision 060db6499f32faf8b98477b0a26969ef7d8b9987).
Full-sequence activations (24 layers, 896 dim, float16) and top-100 logits from Qwen/Qwen2.5-0.5B on 7,660 Geometry of Truth statements. Per-layer sharding (v1.2) with independent shard boundaries.
Contents
Tensor
Layers
Dim
Pooling
Shards
Row Bytes
hidden_layers
0-23
896
-
1
-
logits_topk
-
k=100
last_token
1
1200… See the full description on the dataset page: https://huggingface.co/datasets/latent-lab/got-activations-qwen2.5-0.5b.latent-3d-cachewm-latentcorr-vla-recoverylatents-clap-v2danbooru2024-latents-sdxl-1ktar
Danbooru 2024 SDXL VAE latents in 1k tar
Dedicated dataset to align deepghs/danbooru2024-webp-4Mpixel. "4MP-Focus" for average raw image resolution.
Latents are ARB with maximum size of 1024x1024 as the recommended setting in kohyas. Major reason is to make sure I can finetune with RTX 3090. VRAM usage will raise drastically after 1024.
Generated from prepare_buckets_latents_v2.py, modified from prepare_buckets_latents.py.
Used for kohya-ss/sd-scripts. In theory it may replace… See the full description on the dataset page: https://huggingface.co/datasets/6DammK9/danbooru2024-latents-sdxl-1ktar.latent-image-training
squiggles (metadata-fix)
OC-map FEM rebuild at 35 pixels per wavelength, with corrected geometries,
Helmholtz residuals, and the resolved JCMsuite .jcm / .jcmp files used
for each solve.
Configs
metadata (default)
One row per structure folder (sample_XXXX). Geometry comes from published
optical-constant maps (not the old nested-interface metadata).
validation
One row per FEM incidence (theta in {0, 45}). Self-contained pixel map:… See the full description on the dataset page: https://huggingface.co/datasets/als-rixs/latent-image-training.humanego_serve_bread_lingbot_lerobot_with_latents
HumanEgo Serve Bread LingBot LeRobot With Latents
This dataset contains LeRobot-format robot demonstrations for the task:
pick up the bread and place it on the plate
The repository has two standalone LeRobot-style roots:
humanego_serve_bread_lingbot_eef_train: 55 episodes, 41,603 frames, 55 videos.
humanego_serve_bread_lingbot_eef_val: 6 episodes, 5,533 frames, 6 videos.
Each split includes:
data/: episode parquet files.
videos/: MP4 videos for observation.images.ego_rgb.… See the full description on the dataset page: https://huggingface.co/datasets/Coffeecoderss/humanego_serve_bread_lingbot_lerobot_with_latents.sfx-latentssd15-latent-distillation-500k
SD1.5 Latent Distillation Dataset
⚠️ IMPORTANT: Mixed Scaling Warning ⚠️
This dataset contains SD1.5 latents with two different scaling states:
There is no guarantee the system isn't blended as I ran multiple different versions and I'm still uncertain.
It would be a safe bet to omit the first 10 entirely if you are concerned, or stick entirely to the second set as they are all prescaled.
I don't plan to synthesize any more of this poison - 360k is more than enough. My focus has… See the full description on the dataset page: https://huggingface.co/datasets/AbstractPhil/sd15-latent-distillation-500k.flux-schnell-teacher-latents
Flux Schnell Teacher Latents
Pre-computed latents, decoded images, and text embeddings from FLUX.1-schnell for distillation and research.
Usage
from datasets import load_dataset
# Load specific subset
ds = load_dataset("AbstractPhil/flux-schnell-teacher-latents", "train_512")
ds = load_dataset("AbstractPhil/flux-schnell-teacher-latents", "train_2_512")
ds = load_dataset("AbstractPhil/flux-schnell-teacher-latents", "train_3_512")
Subsets
Config… See the full description on the dataset page: https://huggingface.co/datasets/AbstractPhil/flux-schnell-teacher-latents.cache_vae_latents_train_20hq1lq-part-00imagenet1k_invae-latents_dinov2_pcaVJEPA-LATENTS-L2NORMIndicVoice-latent-NEW-parquetlatent_v1_fullrun_alpha2_05ldt-latents
1 Million Image Latents Toy Dataset
A lightweight toy dataset of 1 003 626 image latents paired with CLIP text embeddings.
Raw sources & extraction
LAION‑aesthetic (laion/laion2B-en-aesthetic):
Streamed via 🤗 datasets in 50 k-image blocks.
Filtered for aesthetic > 7.
Skipped PNG/CMYK or images < 32×32 px.
JourneyDB (MidJourney) (JourneyDB/JourneyDB):
Downloaded three zip archives per batch from Hugging Face.
Unzipped locally and selected the first 50 000 valid… See the full description on the dataset page: https://huggingface.co/datasets/shreenithi20/ldt-latents.