diffusion-gif/training-data-index

Previous training data: versioned location and checksum index

TRAINING_DATA_INDEX.json tells you which retained training inputs belong to which model, their original/current snapshot roles, exact file SHA256/size, and an immutable owner-authorized private backup revision. The same index is attached at the roots of diffusion-gif/any2ego and diffusion-gif/trick-prediction. No original restricted annotation bodies, credentials or original machine paths are in this public index. This index is not a license grant for source data.

Important distinctions

• Legacy mix6 SFT step900: train16,174 / heldout3,228; old multi-point protocol.
• Legacy mix6 GRPO step250: the corresponding original train/heldout Parquet, 8 embedded observation-frame data URIs per record and identical SID sets.
• All19,402 original mix6 NPZ observations and old JSON/Parquet are privately preserved byte-for-byte, with independent tar-member checksum readback.
• The later joint dataset31,202/5,015 and public simulation subset17,421 records must not be substituted as the original training data of those old weights.
• Any2Ego original JSON/CSV, full PhiSeg source manifests and the common32,558 export are separate snapshots. Captioner trainer args identify their source JSON files; generator per-component launch-time subset hashes were not saved.

The index records retained file snapshot hashes, not nonexistent hashes from the original training launch. The large Ego-Exo original RGB/GT-mask pool is NOT remotely backed up here. Original-run data identity and permanent third-party availability cannot be guaranteed merely by a JSON index.

Fetch your original training files

Use an account authorized for the PRIVATE vault. Log in locally; do not put a real token in scripts or chat. This keeps gated/restricted annotations private.

pip install huggingface_hub
hf auth login
python fetch_training_data.py --index TRAINING_DATA_INDEX.json --dataset legacy_mix6_sft --output data_download
# inspect the dry-run plan, then:
python fetch_training_data.py --index TRAINING_DATA_INDEX.json --dataset legacy_mix6_sft --output data_download --execute
python restore_legacy_mix6.py --snapshot-root data_download/snapshots/20261104 --output ready_mix6 --execute
# GRPO: choose --dataset legacy_mix6_grpo; --dataset all_private fetches the entire protected snapshot.

Downloads are pinned to immutable commits, verify exact hashes, and never overwrite a different existing user file. The restore helper changes only npz/frames_root in WORKING JSON copies; all frozen originals remain intact. Keep a separate local/private copy for long-term availability.