datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
libero-dex3da-libero-training-assets
GAM LIBERO Training Assets
This dataset repository contains the assets needed to fine-tune GAM on LIBERO.
Layout
checkpoints/track4world_da3.pth
data/libero_noop/<suite>/*.hdf5
data/libero_noop/_stats/*.json
configs/training/libero_unified/
track4world_da3.pth is the DA3-Giant base checkpoint. The LIBERO HDF5 files contain embedded RGB, proprioception, actions, and depth used by the public GAM training configs.
Code:… See the full description on the dataset page: https://huggingface.co/datasets/SeonghuJeon/3da-libero-training-assets.B2_LIBERO_Longpi05-libero-goal-task-8-evidence
Robium Pi0.5 LIBERO-Goal Task 8 evidence
This is the public evidence bundle for Robium issue #69. It records one fixed,
no-retry evaluation of lerobot/pi05_libero_finetuned_v044 on LIBERO-Goal task
8, put_the_bowl_on_the_plate, using the canonical prompt “put the bowl on the
plate.”
Result
20/20 successful episodes; the predeclared target was 16/20.
Fixed initial states 0–19 map to seeds 1000–1019.
Batch size 1, hard environment/policy reset before every episode… See the full description on the dataset page: https://huggingface.co/datasets/robium/pi05-libero-goal-task-8-evidence.openvla-oft-libero-calib-datapi05-libero-plus-robot-initial-failures
pi0.5 LIBERO-plus Robot Initial State Failures
Failure summary artifacts for evaluating TensorAuto/tPi0.5-libero on the LIBERO-plus Robot Initial States perturbation subset through OpenTau.
Evaluation Setup
Benchmark: LIBERO-plus
Suite: libero_10
Perturbation category: Robot Initial States
Policy: TensorAuto/tPi0.5-libero
Tasks: 393
Episodes per task: 5
Total evaluated episodes: 2430
Seed schedule: episode seeds 1000 to 1004
Episode length: 520 steps
Hardware… See the full description on the dataset page: https://huggingface.co/datasets/d3d3shan/pi05-libero-plus-robot-initial-failures.pi05-libero-plus-language-instructions-failures
pi0.5 LIBERO-plus Language Instructions Failures
Failure summary artifacts for evaluating TensorAuto/tPi0.5-libero on the LIBERO-plus Language Instructions perturbation subset through OpenTau.
Evaluation Setup
Benchmark: LIBERO-plus
Suite: libero_10
Perturbation category: Language Instructions
Policy: TensorAuto/tPi0.5-libero
Tasks: 383
Episodes per task: 5
Metric episodes: 1915
Seed schedule: episode seeds 1000 to 1004
Episode length: 520 steps
Source machine:… See the full description on the dataset page: https://huggingface.co/datasets/d3d3shan/pi05-libero-plus-language-instructions-failures.libero-rldx-demospeedup-slow2-fast4
LIBERO DemoSpeedup: RLDX-1 entropy, slow2 / fast4
Derived from kimtaey/libero_gr00t_delta using prehj/RLDX-1-IMG-LIBERO-60k.
Read FORMAT.md before training. Targets are explicit per-observation action chunks; the original action column is retained for provenance and is not the speedup target.
1693 episodes; 273465 source observations; 271772 usable anchors.
No model was trained for this release. See meta/demospeedup.json and per-episode entropy files.
demospeedup-libero-robocasa-entropy
DemoSpeedup 엔트로피: LIBERO / RoboCasa
기존 GR00T-N1.5 DemoSpeedup 재현에서 실제로 사용한 원본 프레임별 엔트로피입니다.
모델 재추론 없이 구간 분류 역치를 변경하고, 원본 데모를 다른 배속으로 재구성할 수 있습니다.
모델 아키텍처를 변경하는 방법이 아니며, 변환한 데모로 기존 GR00T-N1.5를 학습합니다.
벤치마크
에피소드
프레임
누락/NaN/Inf
LIBERO
1,693
273,465
0
RoboCasa
7,200
2,073,457
0
재현 범위 정정 (2026-09-19)
이 저장소의 엔트로피와 라벨은 우리 GR00T 이식 구현의 결과이며, 논문을 그대로 재현한
결과라고 해석하면 안 됩니다. 논문 §3.2는
클러스터의 평균 엔트로피가 0 미만이면 precision, 그 외(노이즈 포함)는 casual이라고 설명합니다.
공식 공개 코드의… See the full description on the dataset page: https://huggingface.co/datasets/prehj/demospeedup-libero-robocasa-entropy.pi05-libero-plus-light-conditions-failures
pi0.5 LIBERO-plus Light Conditions Failures
Failure summary artifacts for evaluating TensorAuto/tPi0.5-libero on the LIBERO-plus Light Conditions perturbation subset through OpenTau.
Evaluation Setup
Benchmark: LIBERO-plus
Suite: libero_10
Perturbation category: Light Conditions
Policy: TensorAuto/tPi0.5-libero
Tasks: 274
Episodes per task: 5
Total evaluated episodes: 1415
Seed schedule: episode seeds 1000 to 1004
Episode length: 520 steps
Hardware used locally:… See the full description on the dataset page: https://huggingface.co/datasets/d3d3shan/pi05-libero-plus-light-conditions-failures.Swift_libero_longpi05-libero-plus-perturbation-summary
pi0.5 LIBERO-plus Perturbation Summary
This dataset stores the summary report for six completed LIBERO-plus perturbation evaluations of TensorAuto/tPi0.5-libero.
Files:
pi05_libero_plus_six_perturb_hf_report.md: human-readable report with Hugging Face links, success rates, and short analysis.
pi05_libero_plus_six_perturb_hf_report.json: machine-readable summary.
The failure-grid videos and per-category metadata are stored in the linked per-perturbation datasets.
pi05-libero-plus-objects-layout-failures
pi0.5 LIBERO-plus Objects Layout Failures
Failure summary artifacts for evaluating TensorAuto/tPi0.5-libero on the LIBERO-plus Objects Layout perturbation subset through OpenTau.
Evaluation Setup
Benchmark: LIBERO-plus
Suite: libero_10
Perturbation category: Objects Layout
Policy: TensorAuto/tPi0.5-libero
Tasks: 312
Episodes per task: 5
Metric episodes: 1560
Seed schedule: episode seeds 1000 to 1004
Episode length: 520 steps
Source machine: wzxuan under… See the full description on the dataset page: https://huggingface.co/datasets/d3d3shan/pi05-libero-plus-objects-layout-failures.pi05-libero-plus-background-textures-failures
pi0.5 LIBERO-plus Background Textures Failures
Failure summary artifacts for evaluating TensorAuto/tPi0.5-libero on the LIBERO-plus Background Textures perturbation subset through OpenTau.
Evaluation Setup
Benchmark: LIBERO-plus
Suite: libero_10
Perturbation category: Background Textures
Policy: TensorAuto/tPi0.5-libero
Tasks: 289
Episodes per task: 5
Metric episodes: 1445
Seed schedule: episode seeds 1000 to 1004
Episode length: 520 steps
Source machine: wzxuan… See the full description on the dataset page: https://huggingface.co/datasets/d3d3shan/pi05-libero-plus-background-textures-failures.liberoX
liberoX
LIBERO structural-supervision sidecars aligned to the exact Fast-WAM
LeRobot v2.1 release. The four archives retain Fast-WAM's original parquet,
metadata, and AV1 videos byte-for-byte and add lossless active-object masks,
projected Panda skeletons, camera matrices, and replay provenance.
The labels are training-time privileged targets. They are not intended as
policy inputs at deployment.
Contents
Archive
Tasks
Episodes
Frames… See the full description on the dataset page: https://huggingface.co/datasets/zad2eze/liberoX.B0_LIBERO_LongLIBERO-Cosmos-Policy-PointFlowSwift_libero_objectLIBERO_Spatial_B0Libero-Without-Action-ChunkLIBERO_Long_B0LIBERO_Spatial_B3libero-branch-rollouts
LIBERO branch-rollout benchmark (EXP-20260722-004, C-line + Tier-1)
Data generated for a study on metric readiness for video world models: specifically,
whether a learned model's outputs actually respond to the action it is conditioned on,
rather than to the presence of a conditioning signal at all.
Everything here was produced by live MuJoCo simulation on an AMD MI325X node.
It is not a repackaging of the public LIBERO release. See "What is NOT here".
What is here… See the full description on the dataset page: https://huggingface.co/datasets/Minhao-VWM-metrics/libero-branch-rollouts.Libero-Action-Chunklibero-4suites-cachelibero_8tasks_val_newlibero-object-context-rldslibero_object_imageSwift-libero-goal-ActionLibero-Goal-subtask