prehj/libero-compress-labels

LIBERO binary compression labels (no contact overrides)

266,693 labels, 1,693 episodes, 40 tasks. Derived from the existing LIBERO confidence release. No new VLM inference or policy evaluation was performed.

• conf is the original stored score, unchanged: weighted expected-grade support minus risk, mapped to [0,1]. It is NOT a success probability.
• compress_2p5 = int(conf > 0.50); compress_3 = int(conf > 0.55).
• These are initial heuristic thresholds, NOT calibrated from naive success rates. They do not impose per-task quotas or task ceilings. 3x implies 2.5x by construction.
• Contact policy: none; all contact/proxy overrides disabled. If active, both outputs become zero.
• compress_valid=1 on labeled rows. 1 means compress, 0 means keep original speed. The columns represent separate binary targets; they are not a multiclass speed selector.
• source_ratio and source_fixed are historical provenance, NOT new supervision. No-contact retains contact metadata for auditing but never uses it in the targets.
• A..E, eg_A..E, gp_A..E are preserved. Probabilities are conditional over grades 1..5.

Match to the dataset

Use the unchanged kimtaey/libero_gr00t_delta base. Join exactly on (episode_index, frame_index), not task name or parquet row order alone. Last four frames per episode are unlabeled; the apply script preserves all original rows and sets their validity to zero. Never train these tails as negative examples. The script validates all episode lengths/instructions, exact keys and label checksum.

pip install numpy pyarrow
python apply_to_dataset.py --dataset /data/libero_gr00t_delta --dry-run
python apply_to_dataset.py --dataset /data/libero_gr00t_delta --out /data/libero_binary_sidecar
# Alternatively copy original parquets, preserving existing columns:
python apply_to_dataset.py --dataset /data/libero_gr00t_delta --mode dataset --out /data/libero_binary

The default sidecar contains one parquet per original episode. Dataset mode copies original embedded-image parquets. Two optional [binary_class, valid] float32 carriers are compress_2p5_label and compress_3_label. A model/data configuration must explicitly consume the selected binary target and validity; columns alone do not configure a training loss. Carrier fields are not robot action dimensions. Full join: 273,465 frames, 266,693 valid labels, 6,772 masked tail frames.

Change thresholds later without VLM inference

import pandas as pd
d = pd.read_parquet("labels/libero_training.parquet")
d["compress_2p5"] = ((d.conf > 0.50) & (d.contact_override == 0)).astype("int8")
d["compress_3"] = ((d.conf > 0.55) & (d.contact_override == 0)).astype("int8")

For a new package, update thresholds and the parquet SHA256 in label_policy.json; the apply script intentionally rejects a file changed without updating its manifest. Current positive counts: {"compress_2p5": 118630, "compress_3": 92644}.