datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
sokoban-10k-vjepa2-tokenizedcs2-10k-vjepa2-latents-300
CS2-10k V-JEPA2 latent cache (300 matches)
Frozen facebook/vjepa2-vitl-fpc64-256 embeddings of single-POV Counter-Strike 2
gameplay, from 300 matches of RekaAI/CS2-10k
(mirage + dust2). This is the training substrate for the
scale300 hierarchical world model.
Rebuilding it from source takes ~82 hours of wall-clock (elapsed_min 4942.7),
almost all of it network-bound, which is why it is published here.
Contents
file
shape / rows
notes
latents.npy
[6,942… See the full description on the dataset page: https://huggingface.co/datasets/cbctr/cs2-10k-vjepa2-latents-300.vjepa2-temporal-order-blindspots
V-JEPA2 Temporal Order Blind Spots
This dataset documents blind spots for the pretrained video world model facebook/vjepa2-vith-fpc64-256.
The model produces nearly identical embeddings for videos whose temporal order has been severely corrupted, indicating weak sensitivity to temporal directionality and causal motion structure.
Model Tested
Model name: facebook/vjepa2-vith-fpc64-256Type: Self-supervised video world model (JEPA-style joint embedding predictive… See the full description on the dataset page: https://huggingface.co/datasets/Nuntea/vjepa2-temporal-order-blindspots.sokoban-10k-vjepa2-tokenized-shards