Datasetpublic

Iuda/vjepa-temporal-coherence

Layer 16: separate classifiers within each dataset Open in Colab. Select L4/A100 → Run all. Downloads are anonymous; no Drive mount is needed. Each task trains its own linear classifier on frozen layer-16 features, with training-only scaling and grouped cross-validation. Held-out scenes test whether a shared readout works within that dataset. Failure does not establish that the layer lacks the information. Classifier weights never transfer between datasets. Task Labels… See the full description on the dataset page: https://huggingface.co/datasets/Iuda/vjepa-temporal-coherence.

sourceHugging Faceupdated 11d agoView on Hugging Face

0likes89downloads

Dataset Card

Layer 16: separate classifiers within each dataset

**Open in Colab**. Select L4/A100 → Run all. Downloads are anonymous; no Drive mount is needed.

Each task trains its own linear classifier on frozen layer-16 features, with training-only scaling and grouped cross-validation. Held-out scenes test whether a shared readout works within that dataset. Failure does not establish that the layer lacks the information. Classifier weights never transfer between datasets.

Task	Labels	Groups kept together
MAT order	Intact / shuffled, including reversals	Four source variants
UCF order	Intact / shuffled, including reversals	100 source groups
UCF direction	Forward / reverse intact videos	The same 100 groups
IntPhys continuity	Possible / impossible	60 matched scenes

The notebook has 7 code cells, 169 code lines and 241 words of explanation. Pixel change is the order/continuity baseline; endpoint pixels are the direction baseline. IntPhys probing is exploratory and differs from its standard benchmark protocol. MAT sources are closely related; human coherence ratings remain unmeasured.

Verified on 2026-09-29: all seven code cells completed on an NVIDIA L4 in 15 min 43 s, encoding all 1,296 inputs. Layer-16 held-out balanced accuracy: MAT order 100%, UCF order 94.6%, UCF direction 55.5%, IntPhys continuity 58.8%. These estimates alone do not establish statistical significance. Pixel change ranked intact above shuffled videos perfectly within sources, so order classification remains compatible with simple visual cues.

Files

—temporal_coherence_colab.ipynb: complete raw-input workflow.
—temporal_coherence_colab_executed.ipynb: the same notebook with verified outputs.
—within_dataset_results.csv: current probe results; verification.json records checks and hashes.
—ucf_sources.json: fixed source selection and frame indices.
—mat_raw_frames.zip: lossless original pixels and frame orders; the defective two-frame condition is excluded during analysis.
—archive/supervised_study/: the previous study and its separate execution record.

The notebook pins the model, UCF and IntPhys2 revisions and downloads their assets directly. Their original terms apply. The supplied MAT stimuli have no asserted redistribution license.

Results, features and held-out predictions are saved to /content/layer16_within_dataset_results.zip.