Datasetpublic

erl-hub/behaviour1k-Qwen3-features

BEHAVIOR-1K Qwen3 skill features Per-frame conditioned features e_t = Phi(f_t, L_sub^(j), L), mean-pooled primitive skill latents S_j, aligned proprioception q_t, actions a_t, and subtask progress p_t. These are the inputs and targets for a Primitive Skill Composer VLA Skill Predictor. Ground-truth primitives come from BEHAVIOR-1K's hand-authored primitive_annotation, so the segmentation is human-labelled rather than predicted, and nothing here depends on a keyframe detector.… See the full description on the dataset page: https://huggingface.co/datasets/erl-hub/behaviour1k-Qwen3-features.

sourceHugging Facemitupdated 2mo agoView on Hugging Face

0likes15kdownloads

Dataset Card

BEHAVIOR-1K Qwen3 skill features

Per-frame conditioned features e_t = Phi(f_t, L_sub^(j), L), mean-pooled primitive skill latents S_j, aligned proprioception q_t, actions a_t, and subtask progress p_t. These are the inputs and targets for a Primitive Skill Composer VLA Skill Predictor.

Ground-truth primitives come from BEHAVIOR-1K's hand-authored primitive_annotation, so the segmentation is human-labelled rather than predicted, and nothing here depends on a keyframe detector.

This version

—version: v1-hosted-task0022
—feature_source: hosted_embedding
—subtask_source: behavior1k_primitive
—sampling: fixed K=8 per primitive, both endpoints included
—d_pre: 4096 d_vis: None
—samples: 3200, segments: 400, episodes: 200
—code revision: unknown created: 2026-08-03T17:15:24.562363+00:00

Reading these features honestly

K is part of every result. S_j is the mean of K sampled frames, so the zero-parameter predictor S_hat_t := e_t has error sigma^2 (K-1)/K against a floor of sigma^2/K. An identity gain quoted without its K is meaningless.

Feature sources are not comparable. hosted_embedding is a retrieval-trained embedding model reached over an API; vlm_prelogit is a VLM's last hidden state. They differ in width and in what they encode. Never pool them.

Layout

<version>/
  config.json          provenance: models, K, feature/subtask source, git rev
  episodes.parquet     episode_id, task, instruction, split
  segments.parquet     episode_id, segment_id, start/end_frame, L_sub
  manifest.parquet     one row per sample; frame_idx, progress, indices
  shards/<episode>/    pre.npy, vis.npy, segment_targets.npy, meta.json

Use

python

from pls_vla.hub import load_published_tensors

tensors = load_published_tensors("v1-hosted-task0022", split="train")
# -> SkillTensors(vis, pre, skill, progress)

Versions in this repo

—v1-hosted-task0000
—v1-hosted-task0001
—v1-hosted-task0002
—v1-hosted-task0003
—v1-hosted-task0004
—v1-hosted-task0005
—v1-hosted-task0006
—v1-hosted-task0007
—v1-hosted-task0008
—v1-hosted-task0009
—v1-hosted-task0010
—v1-hosted-task0011
—v1-hosted-task0012
—v1-hosted-task0015
—v1-hosted-task0016
—v1-hosted-task0017
—v1-hosted-task0018
—v1-hosted-task0022
—v1-hosted-task0022