Datasets:

ArXiv:
License:
eastha's picture
Complete and seal full-frame canonical D/seed42 dataset
604e5e6 verified
|
Raw History Blame Contribute Delete
2.47 kB
---
license: mit
tags:
- robotics
- vision-language-action
- full-frames
configs:
- config_name: fullframes_d_seed42
default: true
data_files:
- split: train
path: data/train-part-*.parquet
- split: validation
path: data/validation-part-*.parquet
- split: test
path: data/test-part-*.parquet
---
# LIBERO+ full frames · canonical D seed42
Status: COMPLETE — sealed and verified.
Derived from [Sylvest/libero_plus_rlds](https://huggingface.co/datasets/Sylvest/libero_plus_rlds/tree/fb0c7029b076030d5d57227229e4f7460def1f7c), retaining its MIT license and attribution. Original benchmark: [LIBERO+](https://arxiv.org/abs/2510.13626).
All 14,347 episodes / 2,238,036 timesteps are retained. Each row is one timestep with both original 256x256 JPEG byte payloads; no resizing or JPEG re-encoding is performed. Raw 7-D actions, 8-D state, 7-D joint state, language, reward, discount, first/last/terminal flags, and original source episode file metadata are preserved. `source_episode_index` records source TFRecord order. Local Drive paths, credentials and private model weights are not added.
Canonical D episode counts: train 11,477; validation 1,428; test 1,442. Frame counts: train 1,791,269; validation 223,485; test 223,282. Duplicate content keeps its sealed split and calibration labels; identical duplicates consume sealed episode IDs deterministically. The 728 calibration episodes remain part of Train.
`reference/splits.json` and `reference/normalization.json` preserve the existing D/seed42 reference bytes. Numeric data is raw: gripper conversion and train-only normalization are applied by the training consumer, not during storage. Do not frame-shuffle before constructing action chunks or cross episode boundaries.
Stored frame interval is 1. Training query stride (1/4/8) is independent of action horizon (8). Existing sparse datasets/checkpoints are not replaced. Use immutable revisions and check `COMPLETE` / `dataset_manifest.json` before training. Source units are uploaded atomically and remotely hash-verified; `_progress` supports restart without rereading completed source units.
After completion, standard `datasets.load_dataset("VLADE-konkuk/libero-plus-fullframes", revision="<completed commit>", split="train", streaming=True)` reads the Parquet data. `vlade.data.tfrecord_frames.iter_action_queries` provides raw action chunks with an explicit padding mask; apply the sealed normalization before model training.