strive-vid/strive-materialized

STRIVE MolmoPoint prepared data (gated backup)

This repository keeps copies of prepared MolmoPoint data. With it, a new machine, or this one after /scratch is wiped, can skip the SceneSample builds and the InputSample materialization. The contents are derived from the gated strive-vid datasets. This repository is gated too: anyone can see this page, but only accounts that an admin of strive-vid approves by hand can download the files. Send the access request on this page. The STRIVE repository writes and reads this repository with makefiles/molmopoint/hub_backup.mk (code: strive_datasets/molmopoint/hub_backup.py).

Layout

Each top folder is one materialization date. Inside it, the layout is the same as in the data root (data/scratch, a link to /scratch/ac2538/strive_data):

<date>/
├── raw/<repository>/acquisition.json             raw repository, pinned revision, and prefix
├── raw/<repository>/catalogs/                    mesh-caption catalogs that the scenes record
└── molmopoint_prepared/
    ├── <source>_v1/scene_samples/                SceneSample parent
    ├── <source>_v1/input_samples/                collection registry and the children (<label>/<sample_spec_key>/)
    ├── pack_plans/<preset>/                      pack plans built from these children (added when made)
    └── logs/                                     build and materialization logs

The raw scenes are not here. The videos, masks, and tracks (about 2.7 TB) are the pinned revisions of the strive-vid datasets. A child row only points at them, for example at train/00000/frames/rgb.mp4, frames 0–96, with the file's SHA-256. restore downloads them again at the revision that raw/<repository>/acquisition.json records, so each pointer finds the same bytes.

Use

Run from the STRIVE repository root. HF_HOME must hold a Hugging Face token with access to this repository (and to the strive-vid datasets for a restore).

# Restore DATE into data/scratch, then download the raw scenes again (2.5 hours for about 2.7 TB on snavely-compute-03).
make -f makefiles/molmopoint/hub_backup.mk restore DATE=2026-10-19 PYTHON="$PYTHON"
# Back up the prepared data of the three new sources under DATE.
make -f makefiles/molmopoint/hub_backup.mk push DATE=2026-10-19 PYTHON="$PYTHON"
# Add a pack plan built from the children of DATE.
make -f makefiles/molmopoint/hub_backup.mk push-pack-plan DATE=2026-10-19 PRESET=<preset> PYTHON="$PYTHON"

restore never overwrites a file: a file that already exists must have the same bytes, else it stops before it moves anything. A rerun after an interruption continues where it stopped. On a host with another GPU type, new packing budgets and pack plans are needed; the parents and children are reused as they are.

The restore depends on the strive-vid repositories keeping the recorded revisions.

Contents

Date	Sources (revision; scenes per role train / validation / test)	Children	Record
2026-10-19	TexVerse-CLEVRER a5057644, Breaking-Bad 67d60e9b, TCOW d2456918 (10,000 / 1,000 / 1,000 each)	mp_dc arm A (grounded_dense_drop_contacts_v2) and arm D (numeric_sparse_drop_contacts_v2), drop_contacts, 512 × 512, Qwen3.5-0.8B processor 2fc0636	lab-notes/molmopoint/2026-10-19-full-new-source-build/ in the STRIVE repository
2026-10-27	The same sources and revisions as 2026-10-19	The parents and children of 2026-10-19 again (same bytes), and the Arm D streaming children numeric_sparse_drop_contacts_stream_<view>_w8_s4_v2 (views window, window_prefix, merge, prefix; windows of 8 frames at a stride of 4) under the label main, with the A6000 pack plans stream_<view>_4b_arm_d_a6000. Each view is added when it is built. The label scenes20 holds small development children (the first 20 parents of each role).	lab-notes/molmopoint/2026-10-22-streaming-baselines/ in the STRIVE repository