Jevvideo long-video backup (private)

Personal backup, not for redistribution. CG-Bench and Ego4D (LV-Haystack) videos stay under their own licenses.

checkpoints/<TAG>/: runs/<TAG>_experiment.json (settings), runs/<TAG>_ps3_r2.json (final policy set, absolute node paths), ps3_r2.relative.json (same, paths relative to this folder), and the LoRA snapshots it names under runs/<run>/snapshots/step_XXXX/.
- onejev_sft_videomme_mix_v1: 长短混合 (code 6b4d95c)
- onejev_sft_videomme_mix_mseg_v1: 32帧长短混合（新代码） (code 18cf7dd)
- onejev_sft_videomme_pad3_v1_f128: 128帧正式 (code 6b4d95c + infra fixes)
manifests/<TAG>/cgbench_lvh_dev_vts.jsonl: the CG-Bench + LV-Haystack dev split each run used (video_path = the original Blob path).
videos/cgbench/: from /blob/t-xinyangliu/data/cgbench/videos/; videos/lvhaystack_ego4d_full_scale/: from /blob/t-xinyangliu/data/lvhaystack/ego4d/v2/full_scale/ (the union over the three dev manifests).

2026-11-04: the full LV-Haystack train split (from ARC)

videos/lvhaystack_ego4d_full_scale/: every video of the LongVideoHaystack train split that ARC has, 857 of its 858 videos (516.9 GiB, Ego4D full scale; never downloaded: 0ca23a40-6daf-4503-bfa2-f315a79b7317). It covers the train and dev rows of every ARC run (they all come from this split), including the 5.2 h and 3.09 h videos the fused run drops.
lvhaystack/: ARC's LV-Haystack root without its videos (scripts/prepare_lvhaystack.py):
- annotations/data/{train,val,test}-00000-of-00001.parquet: MLL-Lab/LongVideoHaystack @ 2ad05dee (train 12,892 questions on 858 videos).
- manifests/lvhaystack_{all,train,dev,test}.jsonl and lvhaystack_stats.json: the prepared rows; train and dev both come from the LVH train split, test is LVH val (this upload adds none of its videos). video_path = the ARC path /projects/swan_research_dlm/czwang26/jev_data/lvhaystack/ego4d/v2/full_scale/.
- PREPARED.json, video_uids.txt, missing_video_uids.txt.