Temporal-SAE-VideoMAE — reproducibility artifacts

Weights, cached activations, and the synthetic NFP ball dataset for the project Detecting Temporal Concepts in Video Transformers via Sparse Autoencoders (code: https://github.com/AndrewRqy/temporal-sae-videomae).

These artifacts reproduce the full study — the trained SAEs plus the PCA / ICA / raw linear-decomposition baselines (monosemanticity score and the no-false-positives temporal-feature test) — without re-training, re-running the VideoMAE forward passes, refitting PCA/ICA, or re-rendering the dataset.

VideoMAE (MCG-NJU/videomae-base-finetuned-ssv2) and DINOv2 (facebook/dinov2-base) backbone weights download automatically from the Hub; the SAE / PCA / ICA weights here are the ones trained/fit for this project. To re-extract activations from scratch you also need the public Something-Something-v2 dataset.

The cluster_nfp_results/ tensors are verification-only — every experiment is reproducible from the hosted weights + data, so these saved cluster outputs simply let you diff against the published numbers without recomputing (the results are regenerable by re-running the NFP test).

Contents

Path	Size	What it is
weights/sae/videomae_sae.pt	37 MB	VideoMAE SAE (6144 feat, ×8, l1=0.1, dead-pen 0.03) — the main result (MS 0.475, NFP 75/6144)
weights/sae/dino_sae.pt	37 MB	DINOv2 SAE — negative control
weights/sae/synthetic100_sae.pt	37 MB	Synthetic SAE, 100 static dirs — positive control
weights/sae/synthetic763_sae.pt	37 MB	Synthetic SAE, 763 static dirs — robustness variant
synthetic/matrices_100.pt	0.3 MB	Synthetic ground truth (n_static=100): W_τ, W_static, seed=42
synthetic/all_videos_100.pt	72 MB	Synthetic reps h [3000,8,768] + tau + tau_norm — §3 positive control
synthetic/matrices_763.pt	2.3 MB	Synthetic ground truth (n_static=763): W_τ, W_static, seed=42
synthetic/all_videos_763.pt	72 MB	Synthetic reps (763-static) — §4 robustness variant
weights/pca.pt	1.6 MB	Fitted PCA decomposition (mean + components), 256-comp fit on layer-11 acts
weights/ica.pt	1.6 MB	Fitted FastICA decomposition (mean + un/mixing matrices)
reproducibility/ssv2_val_dinov2.pt	2.4 MB	DINOv2 max-pooled embeddings for 800 SSv2-val clips (MS reference)
reproducibility/pca_sign_split_val_acts.pt	1.6 MB	Max-pooled PCA features for the 800 val clips
reproducibility/ica_sign_split_val_acts.pt	1.6 MB	Max-pooled ICA features for the 800 val clips
reproducibility/nfp_results/sae_nfp.pt	370 MB	VideoMAE SAE NFP tensors (local run; reproduces cluster)
reproducibility/nfp_results/raw_nfp.pt	45 MB	Raw-layer NFP tensors (C, t_stat, p_val, C_mean, tau, mask)
reproducibility/nfp_results/pca_sign_split_nfp.pt	30 MB	PCA NFP tensors
reproducibility/nfp_results/ica_sign_split_nfp.pt	30 MB	ICA NFP tensors
cluster_nfp_results/videomae_deadpen_0p03.pt	370 MB	Cluster VideoMAE SAE NFP result (the paper's 75/6144)
cluster_nfp_results/videomae_deadpen_0p03_avgtau.pt	370 MB	Avg-frames tau variant (§4.2)
cluster_nfp_results/videomae_deadpen_0p03_analyzed.pt	593 MB	Post-processed: per-feature/tau correlations (nfp_analyze.py)
cluster_nfp_results/dino_patch_nfp.pt, dino_negative_control.pt	370 MB ea	DINOv2 negative-control NFP results (§2)
cluster_nfp_results/synthetic_nfp.pt, synthetic_763_nfp.pt	370 MB ea	Synthetic positive-control NFP results (§3/§4)
activations/ssv2_train_layer11_post_mlp_residual_part1.pt	1.8 GB	Cached VideoMAE layer-11 (post-MLP residual) activations over 400 SSv2-train videos (~627k tokens) — refit PCA/ICA at any D without re-running VideoMAE
nfp_ball_dataset.tar.gz	271 MB	3000 synthetic ball videos (16 frames, 224×224 PNGs + metadata.json with tau ground truth + ball-token indices). Extracts to nfp/v00000 … v02999/

SHA256SUMS.txt (in this repo and the GitHub repo's PR branch) has integrity checksums.

How these map to the code

All consumed by scripts in the GitHub repo (SAEs: EXPERIMENTS.md §1–§4; baselines: §5b–§5d):

• weights/sae/videomae_sae.pt → AutoEncoder.from_pretrained; --sae_model standard --sae_path … in training/extract_activations.py (MS) and analysis/nfp_test.py (NFP). Run via jobs/local/run_sae_local.ps1.
• weights/sae/{dino,synthetic100,synthetic763}_sae.pt → same loader, for the §2 negative control and §3/§4 positive controls (point nfp_test.py at the matching dataset).
• synthetic/{matrices,all_videos}_{100,763}.pt → ground truth + data for the synthetic positive control (§3) and robustness variant (§4): analysis/nfp_test_synthetic.py, proj_fraction.py.
• cluster_nfp_results/*.pt → the published NFP output tensors (the paper's numbers), to diff against re-run results; load with torch.load (each has C, t_stat, p_val, C_mean).
• weights/{pca,ica}.pt → loaded by PCADict/ICADict.from_pretrained; used by jobs/local/run_nfp_local.ps1 and analysis/sweep_pca_ica_dim.py.
• activations/…part1.pt → place under local_runs/train_acts/; the fitting (analysis/fit_pca_ica.py) and the sweep (--train_dir) read it.
• reproducibility/ssv2_val_dinov2.pt → --embeds_path for the MS metric and the sweep.
• reproducibility/nfp_results/*.pt → the raw NFP outputs behind the tables in results/pca_ica_baselines/ (load with torch.load).
• nfp_ball_dataset.tar.gz → tar xzf into data/output/ (gives data/output/nfp/), the --nfp_dir / --dataset_dir for the NFP test and sweep.

Quick start

# Dataset (for the NFP test)
huggingface-cli download AndrewRqy/temporal-sae-videomae nfp_ball_dataset.tar.gz --repo-type dataset --local-dir .
tar xzf nfp_ball_dataset.tar.gz -C <sae-for-vlm>/data/output/

# Weights + caches (skip refitting / re-extracting)
huggingface-cli download AndrewRqy/temporal-sae-videomae --repo-type dataset --local-dir hf_dl
# then move weights/ -> local_runs/decomp/, activations/ -> local_runs/train_acts/,
# reproducibility/ssv2_val_dinov2.pt -> local_runs/embeds/, etc.

Numbers were produced on a single RTX 5070 (256-component PCA/ICA fit on ~627k tokens). They are the valid relative comparison to the SAE (whose scores are the cluster reference in the paper).