Datasetpublic

d3d3shan/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.

sourceHugging Facemitupdated 4mo agoView on Hugging Face

0likes72downloads

Dataset Card

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: wzxuan under /data/wangjinqiu/tta-memory/libero_plus_eval

Summary

—Successes: 1841 / 1915
—Success rate: 96.14%
—Failures: 74
—Failure grid videos: 5
—Failures missing videos in the grid report: 0

Files

—videos/: 16-panel failure grid videos.
—metadata/summary.json: benchmark counts, success rate, and difficulty-level breakdown.
—metadata/metric_records.json: per-episode metric records used for the success rate.
—metadata/eval_info_paths.json: eval files included in the payload.
—metadata/source_configs/: configs/logs copied from the wzxuan run payload when available.

The grid videos are for qualitative failure-mode inspection. The official success rate above is computed only from metric eval outputs and excludes failed-task video reruns.