LieUr/vivit-b16x2-k400-postblock5-bf16-activations
ViViT-B/16x2 K400 post-block-5 bf16 activations This data-only repository contains the frozen 10,400-clip activation cache used by NeonByte L5: 3,200 train, 800 dev, and 6,400 eval activations. Each .safetensors file stores the 3,137 × 768 bfloat16 hidden state after ViViT block 5 in CLS|tubelet(t,y,x)-row-major order. The repository contains derived activations and provenance metadata only. It contains no raw video, frames, audio, model weights, executable scripts, or NeonByte… See the full description on the dataset page: https://huggingface.co/datasets/LieUr/vivit-b16x2-k400-postblock5-bf16-activations.
ViViT-B/16x2 K400 post-block-5 bf16 activations
This data-only repository contains the frozen 10,400-clip activation cache used by NeonByte L5: 3,200 train, 800 dev, and 6,400 eval activations. Each .safetensors file stores the 3,137 × 768 bfloat16 hidden state after ViViT block 5 in CLS|tubelet(t,y,x)-row-major order.
The repository contains derived activations and provenance metadata only. It contains no raw video, frames, audio, model weights, executable scripts, or NeonByte source code. split.json, the annotation CSVs, label map, and diagnostic_sources.json preserve the pinned selection and source identities needed for reproducibility.
Source model: google/vivit-b-16x2-kinetics400 at revision 8a7171a57f79b9aaa58bc8d977c002a0ea0f0d42. Source media identities follow the official CVDF Kinetics-400 shard receipts. See LICENSE.md for the source-dependent rights notice.